ABSTRACT. Science is fundamentally transforming our lives, with technologies such as artificial intelligence (AI), gene editing, and quantum computing having profound impacts across society. However, the journey of scientific ideas from conception to societal implementation involves complex socio-technical interaction mechanisms that require systematic understanding (Roco, 2020). While AI has been studied since the 1950s, only recently has it witnessed unprecedented public interest and practical implementation, exemplified by the emergence of generative AI systems like ChatGPT that are rapidly reshaping both daily life and industry. This contrast between AI's long academic history and its recent explosive societal impact raises crucial questions about the mechanisms of scientific knowledge diffusion and social acceptance - a phenomenon that exemplifies the complex convergence-divergence patterns typical in the evolution of transformative technologies (Roco, 2020). This study investigates the dynamic interaction between scientific knowledge and social discourse in the field of AI by analyzing conceptual structure of multi-layer network. By quantitatively examining how scientific concepts interact with and evolve alongside social discourse, we aim to uncover the underlying mechanisms that drive the translation of scientific ideas into societal impact. Our study analyzes an extensive dataset spanning from 1980 to 2020, utilizing over 10 million journal articles and conference papers from OpenAlex and over 800,000 news articles from ProQuest. To understand the conceptual structure evolution in both domains, we constructed co-occurrence based semantic networks (Kedrick et al., 2024). For scientific knowledge, we utilized the hierarchical Fields of Study (FoS) relationships in OpenAlex to build two distinct networks: a convergence network that connects Level 1 FoS through shared Level 2 FoS, and a divergence network that links Level 1 FoS through individual papers. The temporal analysis of these networks reveals distinct evolutionary patterns. The social subject network shows three major peaks of activity around 1998, 2008, and 2017, characterized by balanced growth in both nodes and edges, suggesting organic expansion of social discourse. These peaks appear to coincide with significant AI milestones that sparked widespread public interest. The 1998 peak aligns with IBM's Deep Blue defeating chess champion Garry Kasparov (1997). The 2008 peak follows the introduction of the term "Deep Learning" by Geoffrey Hinton's team in 2007, marking AI's expansion beyond purely technical domains into various social sectors including transportation, finance, and industry. The 2017 peak reflects the societal impact of AlphaGo's victory over Lee Sedol (2016), which dramatically demonstrated AI's potential to the general public and triggered increased investment in AI across both private and public sectors. The scientific concept networks display notably different dynamics: while the number of nodes tends to converge over time, the number of edges continues to increase, indicating that AI's scientific evolution is characterized more by deepening connections between established concepts than by the emergence of entirely new ones. Particularly interesting is the divergence network's strong growth in new edges after 2015, suggesting an acceleration in cross-disciplinary applications of AI concepts. This surge coincides with the widespread adoption of deep learning technologies across various domains, catalyzed by landmark achievements like AlphaGo and, more recently, the emergence of generative AI systems like ChatGPT that have begun reshaping daily life and industry practices. To evaluate the structural evolution of these networks, we employ three key metrics: core churn rate (C), relative number of core nodes (R), and number of cores (S) (Kedrick et al., 2024). The social subject network shows increasing core churn and decreasing relative core size over time, indicating dynamic restructuring of central concepts in media discourse. Conversely, both convergence and divergence networks in the scientific domain exhibit decreasing core churn and increasing relative core size after the mid-1980s, suggesting the consolidation of fundamental AI concepts. Notably, while the number of cores increases across all networks, their concentration remains relatively stable around 0.1, indicating the absence of dominant core concepts. This suggests that both scientific knowledge and social discourse in AI maintain a distributed, multi-focal structure rather than converging around a single dominant paradigm. These findings provide novel insights into how scientific knowledge and social discourse co-evolve, particularly in rapidly evolving fields like AI. The contrasting patterns between scientific and social networks—particularly in conceptual structure dynamics—suggest different mechanisms of knowledge evolution in these domains. While scientific knowledge shows a pattern of conceptual consolidation with increasing interconnections, social discourse exhibits more dynamic restructuring of core concepts. Understanding these patterns and mechanisms could inform strategies for more effective science communication and technology transfer, while also providing a methodological framework for studying knowledge evolution in other scientific domains.

Kedrick, K., Levitskaya, E., & Funk, R. J. (2024). Conceptual structure and the growth of scientific knowledge. Nature Human Behaviour, 8, 1915-1923. https://doi.org/10.1038/s41562-024-01957-x Roco, M. C. (2020). Principles of convergence in nature and society and their application: from nanoscale, digits, and logic steps to global progress. Journal of Nanoparticle Research, 22, 321. https://doi.org/10.1007/s11051-020-05032-0

ABSTRACT. 1. Introduction. Emerging technologies, characterized by radical novelty, rapid growth, coherence, significant impact, and inherent uncertainty (Rotolo et al., 2015; Small et al., 2014), challenge traditional STI policy frameworks. Among these, generative artificial intelligence (AI) has become a prime example, driven by its transformative potential and rapid development. The introduction of the Transformer model by Vaswani et al. (2017) marked an important milestone, driving generative AI towards widespread adoption and interdisciplinary applications. Since then, this technology has demonstrated profound societal impacts and accelerated research worldwide, especially in the United States and China. However, other countries, such as Japan, have not experienced the same pace of growth, raising concerns about their absorptive capacity. Absorptive capacity, as defined by Cohen and Levinthal (1990), refers to the ability of a research system to identify, assimilate, and apply new knowledge. In the context of emerging technologies, "national research absorptive capacity" encompasses the ability of research ecosystems to integrate innovative research findings and transform them into impactful advances. This study extends the concept to examine how resource allocation, interdisciplinary collaboration, and institutional structures influence a nation's ability to advance generative AI research, using Japan as a case study.

2. Methodology To achieve its objectives, this study combines bibliometric analysis and survey research. The bibliometric analysis mapped the landscape of generative AI research in Japan, addressing the challenge of identifying relevant works across different fields. The seminal paper "Attention is All You Need" (Vaswani et al., 2017) served as the core reference, and citation chaining was used to define three data sets. The core set included 18,746 papers that directly cited the core reference, the marginal set included 66,802 papers that cited papers in the core set, and the boundary set included 126,120 papers that cited papers in the marginal set. Together, these datasets represented 211,668 papers, including key works such as BERT, Vision Transformer, and GPT-3. The survey targeted 575 Japanese researchers identified from the bibliometric dataset, yielding 63 valid responses (16.9% response rate). Respondents were asked about their research timeline, including when they started generative AI work, published results, and recognized its societal potential. They also rated whether their research activities expanded rapidly or gradually, and assessed their ability to adapt to the global acceleration of generative AI research.

3. Results and Findings The bibliometric analysis revealed several trends in Japan's generative AI research ecosystem. In the early stages (2017-2018), Japan demonstrated moderate absorptive capacity, ranking seventh in the world in terms of publication output. However, this was followed by stagnation in publication volume and impact metrics. Leading institutions included the University of Tokyo (600 papers), Kyoto University (239), and Tokyo Institute of Technology (233). Early adopters, such as the National Institute of Information and Communications Technology, contributed disproportionately in the early stages, with 11 papers in the first year. This disparity suggests a mismatch between early research activities and subsequent institutional growth, as larger universities entered the field later but failed to capitalize on the momentum created by early contributors. The survey results provide further insight into the dynamics of generative AI research in Japan. Approximately 20 percent of respondents reported starting their research before June 2017, when the Transformer model was introduced. After this milestone, new researchers entered the field at a steady, modest pace. By the end of 2020, approximately 30% of respondents recognized the transformative potential of generative AI, with the largest spike in awareness occurring between January 2022 and May 2023. In contrast to the global trend, where the number of new researchers grew rapidly, Japan's participation expanded gradually, driven primarily by existing researchers rather than new entrants. Regarding the scalability of research activities, 17% of respondents reported rapid expansion, 53% described gradual growth, and 30% noted no significant changes. Respondents assessed their adaptation to the global acceleration of generative AI research, with 4% believing they have adapted very well, 40% believing they have adapted adequately, and 47% responding neutrally. These results are consistent with the gradual growth trends observed. While some researchers adapted successfully, a significant proportion did not see significant change or remained uncertain about their ability to adapt. Key enablers for successful adaptation included prior computational and AI expertise and access to domain-specific datasets, which significantly influenced scalability. However, significant barriers included limited computational resources, talent retention challenges, and inadequate administrative and infrastructure support. Many skilled researchers migrated to the private sector, attracted by better compensation and stability, reducing the capacity of academia to sustain generative AI research. A ridge regression analysis was conducted to explore the factors that enabled adaptation. Missing data were imputed using multiple imputation, resulting in five data sets for analysis. The final model achieved R2 values ranging from 0.400 to 0.502. Key predictors included early engagement in generative AI research, access to key datasets, specialized expertise in AI, secured research support systems, and interdisciplinary research teams. These findings underscore the importance of foundational resources, expertise, and institutional support in enabling early-engaged researchers to effectively scale their activities and adapt to global trends.

5. Discussion This study provides insights into the trajectory of generative AI research in Japan, highlighting factors that enabled or constrained researchers' adaptability to global trends. Researchers who engaged in generative AI activities early on, supported by adequate expertise, access to essential data, robust research support systems, and interdisciplinary collaboration, were better positioned to rapidly expand their activities. However, systemic challenges such as limited computational resources, talent shortages, and insufficient institutional support remain significant barriers. Generative AI highlights broader challenges in fostering emerging technologies within national STI systems. The findings underscore the importance of maintaining a core group of researchers with absorptive capacity in nascent fields and dynamically supporting them when growth signals emerge. Potential strategies include consolidating talent in dedicated research centers, providing rapid access to computational resources, and shifting from individual to team-based funding mechanisms. By addressing these challenges, Japan can improve its ability to adapt to rapid technological advances and ensure its competitiveness in the global research landscape.

ABSTRACT. In this paper, we identify the emerging digital automation technologies and areas of scientific advance (or scientific fields) since 2010 from patented innovations and scientific publications. We cover a large spectrum of digital automation technologies, beyond robots and AI, including for example data acquisition and data management technologies, computing, networking, additive manufacturing, and user interfaces (Savona et al., 2022).

We extract a corpus of digital automation patents since 2001using a combination of Derwent codes, IPCs and keywords in the Derwent Index. Using those patents’ titles and abstracts we build queries to search the OpenAlex publication repository for scientific publications on digital automating technologies and extract a corpus of publications from 2001. Next, from each of the two corpora we identify novel and established patents and publications using an anomaly detection algorithm (Jeon et al., 2022). Novel, are the documents in the top 90th percentile of the novelty measure. These are patents/publications that create substantially different inventions, and which are most likely to introduce new technological directions or new applications within existing technologies. Established, are the documents in the bottom 10th percentile of the novelty measure. These are patents/publications that build on previous inventions and create more inventions of the same type. We also identify the documents that are most similar to those novel and established patents and publications, i.e. which develop the technology in similar directions in the future years — their ‘offshoots’. This allows to exclude patents that do not generate follow up (similar) inventions.

Next, for each set of established and novel documents we cluster them in 500 clusters of similar inventions and research areas. We obtain a large number of granular technologies, ranging from neural networks and self-driving vehicles, to block- chain and medical monitoring devices. We classify each granular technology by their pattern of growth to distinguish the speed at which they grow, the year in which they start growing, when they become established and when most of the novel patents emerge. We finally cluster all 1000 granular technologies in a smaller number of 80 digital automation technologies.

Each of these 80 technologies includes granular technologies that can be either novel or established, and which may have been following different patterns of growth. To identify these patterns, we compute the average year of the bottom 10% of the patent applications by year. We next study the temporal trends between the emergence of the different technologies. We estimate whether novel groups generally precede or follow established groups even within a broad technology. We find that novel groups, on average precede the established technologies. This suggests that a set of radically novel inventions can lead to the creation of a new technology followed by a series of innovations and improvements, resulting in the consolidation of that technology.

We distinguish three types of technologies. • First type are technologies that were novel in 2011-2015, grew exponentially in the number of inventions, forming established technologies by 2021. AI, Blockchain, IoT, Cloud computing and AM all pertain to the first group. • In the second type are technologies that were established in 2011-15 (build on technologies already common since 2000), and which gave way to novel applications (rather than technologies) through 2016-2021. Old and new ICTs related to payment, workflow management, information processing are the main technologies in this group. • In the third type are technologies that fluctuate between established and novel granular technologies more than once between 2011-2021. These are mainly robots. These technologies are relatively similar to technologies developed since 2000, but also create several novel technologies since 2011.

We finally examine the trajectories for novel and established technologies. We find that established technologies typically build upon the novel ones before 2012. These older technologies likely follow a path-dependent pattern, where initial developments create a foundation upon which newer, complementary innovations can emerge. These include more mature technologies of type 2 such as mobile and point-of-sale payment systems, messaging, digital advertising, gaming, logistics, and management systems. In contrast, technologies post-2015 (type 1) seem less path-dependent, beginning as radical innovations before moving toward incremental refinements. These include technologies cantered around autonomous and electrical vehicles, renewable energy forecasting, augmented reality, facial recognition, deep learning, cloud computing, blockchain, internet of things, additive manufacturing, and robotics. For this set of technologies, radical novelty potentially leads to establishment of the technology after a few years. In addition, some of the novel technologies such as quantum computing or artificial intelligence for health management might not have yet stabilized into an established group of technologies. Therefore, these exist only among the novel groups in our data.

Our paper makes several contributions. First, we combines the literature on emerging technologies (Rotolo et al., 2015), and on digital automation technologies (Van Roy et al., 2020; Baruffaldi et al., 2020; Martinelli et al., 2021; Singh et al., 2021; Mateos-Garcia & Klinger, 2023). We contribute by integrating and operationalising attributes of emerging technologies to a corpus of documents that identify a broader set of technologies that included in previous literature. This allows us to not only to identify what are the main technologies and their main trajectories, but also to analyse which ones are most likely to generate more radical innovations in the near future. We also contribute to the literature that uses NLP techniques for mapping and forecasting technologies (Hain et al., 2022; Arts et al., 2023; Shibayama et al. 2021; Yin et al. 2023; Kelly et al. 2021). Our approach is closest to Jeon et al. (2022). With respect to them, we use a pre-trained sentence transformer that has shown the best performance in understanding the context of written text, so far: MPNet (Song et al., 2020). Finally, we contribute to the literature that identifies technologies related to occupations (e.g. Webb, 2020; Kogan et al., 2021). We contribute to this literature by identifying a comprehensive set of 500 narrow and 80 broad digital automation technologies, applications and research areas which can be used to identify the relation between specific digital automation technologies and work.

ABSTRACT. This submission can be related to Thematic Panel: "AI in Science, Technology, and Innovation Policy Studies – Paradigm Shift or Necessary Compromise?”

----- Abstract

This work investigates the use of Large Language Models (LLMs) to identify innovations from web-scraped content, focusing on AI adaptation in Finland. The primary aim is to explore how advanced AI methods can support innovation measurement through unstructured data analysis. To achieve this, the study uses GPT-4o, a long context LLM, to extract relevant artifacts from web content, with a focus on entity identification and relationship extraction to generate knowledge graph (KG) structures. This research aims to understand how the combination of LLMs and KGs can provide a more comprehensive view of innovation landscapes.

Preliminary findings indicate that LLMs effectively capture complex innovation-related information that traditional methods may overlook. However, LLM bias toward over-identifying artifacts poses challenges, which are addressed through additional filtration steps using LLM-as-a-judge evaluations and expert review. The results underscore the potential of LLMs to enhance innovation detection and measurement at scale, while also highlighting the need for human oversight in the process. This study contributes to the growing field of LLM integration in research processes, offering insights into how such data may be evaluated and adopted for use by innovation policymakers and strategic managers.

Introduction

Innovation is critical for economic growth and competitiveness. However, traditional methods for measuring innovation, such as surveys and patent analyses, often lack comprehensiveness and fail to capture emerging trends promptly. These traditional indicators frequently lack sectoral, technological, and geographical detail, leading to a narrow, retrospective view of innovation activities. The rapid advancement of digital technologies and the complexity of innovation ecosystems create a need for more adaptable and scalable approaches to measure innovation efficiently.

This work addresses these limitations by leveraging LLMs to extract innovation indicators from web-scraped content, providing a new approach for identifying and classifying innovations. LLMs, such as GPT-4, can perform high-level analysis of unstructured text-based data, making them well-suited for sources like company websites, news articles, and industry reports. Our research aims to answer the following questions: (1) How can LLMs effectively extract innovation indicators from web-scraped data? (2) What advantages do LLMs offer over traditional methods? (3) How can integrating LLMs with KGs enhance the analysis of unstructured data?

Methodology

The study uses web-scraped data from company websites, industry reports, and other publicly available sources, processed using LLMs to extract relevant entities and relationships. A variety of web pages were scraped, including corporate websites, press releases, and research publications, to ensure a diverse dataset. Data preparation includes cleaning, deduplication, and structuring the extracted information into KGs. LLMs facilitate the extraction of key information, such as types of innovation, thematic focus, and organizational involvement, even from unstructured and noisy sources.

Once the entities and relationships are extracted, the information is structured into KGs. KGs are used to visualize and analyze the relationships among entities, such as companies, products, technologies, and collaborations, providing insights into the innovation ecosystem. By linking these entities, KGs offer a deeper understanding of how different components of the innovation system interact and evolve.

The methodology also includes validation by comparing extracted information with existing innovation databases to assess the accuracy and reliability of the LLM-based approach, ensuring that the insights complement traditional innovation metrics.

Preliminary Findings

Initial results indicate that LLMs are promising in detecting innovation-related information from large datasets, revealing insights that traditional methods may miss. However, even with detailed instructions, LLMs tend to over-identify instances of innovation, leading to inflated results. To mitigate this, we incorporated additional filtering mechanisms, including a secondary evaluation step using LLM-as-a-judge techniques and expert reviews. These steps help reduce biases and improve reliability, but highlight that fully automating the process still requires human oversight and intervention.

We also found that the integration of KGs enhances the representation of complex relationships, providing deeper insights into the innovation landscape. KGs reveal connections between innovations, such as linkages between a company's R&D activities and partnerships, offering a richer understanding of the broader innovation ecosystem.

Another key finding is the ability of LLMs to identify patterns and trends in innovation across sectors. For example, the analysis revealed a growing focus on sustainability-related innovations, with many companies investing in green technologies. This insight is crucial for policymakers and industry leaders to understand the direction of technological development and allocate resources accordingly.

Discussion and Conclusion

The findings suggest that LLMs, when combined with KGs, hold significant potential to improve the measurement and analysis of innovation. LLMs address several limitations of traditional methods, such as their inability to capture real-time developments and reliance on structured data. By analyzing unstructured web data, this study demonstrates how AI integration can offer a more dynamic and up-to-date view of innovation activities. Moreover, KGs provide a structured representation of the innovation ecosystem, helping to identify key players, emerging technologies, and potential areas for collaboration, which is particularly valuable for informed decision-making by policymakers and industry leaders.

Future research should focus on refining LLM models to reduce issues such as over-identification and hallucinations and expand the dataset to include more diverse information sources. Integrating other AI techniques, such as machine learning-based clustering and classification, could further enhance innovation data analysis. Exploring hybrid models that combine LLMs with other AI approaches could also lead to more robust and reliable insights.

ABSTRACT. Abstract

This work explores the integration of Large Language Models (LLMs) into an international innovation policy database (EC-OECD STIP Compass), a system that aggregates data on global science, technology, and innovation (STI) policies. As the volume of data grows, traditional methods for collecting and validating survey responses have proven labor-intensive. To address this, we conducted an experimental pilot using GPT-4o, a long-context LLM, to evaluate whether LLMs can assist in prefilling survey data by extracting relevant information from web-scraped content. The findings suggest that LLM-assisted data collection can significantly enhance efficiency, but challenges remain regarding accuracy and information overlap. This research lays the groundwork for future refinements in the use of LLMs for wide policy data collection and validation.

Introduction

Limited availability of comprehensive and comparable data across countries and regions is a major challenge in understanding the design and effects of innovation policies. This limitation arises due to the costs of collecting such data and challenges related to the unit of analysis—as policies do not easily lend themselves to large-scale data collection (Flanagan et al. 2011). Large Language Models are suited to process complex, unstructured data and are highly scalable, potentially addressing these challenges (Feldman et al. 2015). However, LLMs are not a silver bullet for innovation policy studies. They require input data reflective of the real-world distribution of innovation policies, which cannot be inferred from publicly available data without human expertise. Additionally, LLM-driven data-generation processes introduce noise into the data (Shumailov et al. 2024).

In this paper, we propose a methodology that leverages the strengths of LLMs for data-generating processes while safeguarding against their weaknesses. Our method integrates LLM usage with an expert survey, reducing data collection costs by minimizing experts' time while ensuring accuracy in the LLM-generated data.

The use of LLMs, which have shown promise in other domains of information extraction and text analysis, offers a potential solution to streamline data collection processes. LLMs can process large amounts of text quickly, enabling more efficient data extraction from web-based sources, which could augment traditional human-led surveys. This study focuses on a pilot experiment designed to assess the effectiveness of LLMs in assisting with the EC-OECD STIP Compass survey.

LLMs, particularly models such as GPT-4o, have been lauded for their ability to handle long-form text and perform complex natural language processing (NLP) tasks, including information extraction, summarization, and question answering. However, their application in policy data collection—particularly in STI policy—remains relatively underexplored. The primary goal of this study is to evaluate how effectively an LLM can fill in survey responses based on content scraped from URLs provided by human respondents.

Methodology

The experiment was conducted using a pilot sample of six countries from the OECD—Canada, Finland, Germany, Korea, Spain, and Turkey. The methodology focused on testing two primary approaches to LLM integration: (1) Retrieval Augmented Generation (RAG) and (2) LLMs with Long Context Windows, specifically GPT-4o.

The experiment involved the following steps:

Data Preparation: Survey data was drawn from responses provided by human participants, which included relevant URLs for each policy initiative. Web scraping was performed to collect the full content of the URLs. Initiatives with fewer than 200 tokens or more than 120,000 tokens were excluded from analysis.

LLM Integration: Key survey questions were used to design prompts for GPT-4o. For web content exceeding the token limit, chunking methods were applied. The LLM was prompted to summarize and extract relevant information on policy instruments, objectives, target groups, and policy themes.

Comparison and Evaluation: The LLM outputs were compared to human responses. We evaluated the overlap between human-provided and LLM-generated responses across dimensions such as policy instruments and objectives. The methodological process includes a validation layer that assesses LLM performance (e.g., flagging hallucinations).

Findings

The experiment produced mixed results, highlighting both the potential and limitations of LLMs in this context. In terms of overlap, the LLM captured at least one relevant policy code (e.g., target group, instrument, or theme) in 95% of cases. This result indicates that the LLM was able to identify key themes and instruments that aligned with the policy content.

However, at more detailed levels of comparison, such as full descriptions and objectives of the policy initiatives, the overlap was much lower. Only 1.19% of descriptions had full overlap between LLM and human responses, with 74.05% showing high overlap and 9.52% showing no overlap at all.

A deeper analysis revealed that low overlap was often due to differences in how the LLM and humans prioritized information. Human responses emphasized practical implications and societal impacts, while the LLM focused more on technical details or procedural aspects.

Discussion and Conclusion

This pilot study demonstrates that LLMs can be valuable for prefilling policy survey responses, offering significant advantages in terms of time efficiency and flexibility. However, important limitations must be addressed before LLMs can be fully integrated into large-scale data collection efforts like the EC-OECD STIP Compass.

One key challenge is developing effective validation mechanisms to catch hallucinations and incorrect responses. This could involve using multiple LLMs in tandem, where one model evaluates the output of another, or employing human oversight where LLM-generated responses are unclear or potentially inaccurate. Improving the quality and relevance of input data is also crucial. As shown, LLM performance depends on the quality of the web-scraped content. Ensuring only high-quality URLs are used, or incorporating additional preprocessing steps to filter out irrelevant content, would likely improve LLM performance.

The application of LLMs in STI policy data collection presents a promising avenue for improving survey efficiency and effectiveness. While the pilot results are encouraging, significant work remains to address LLM limitations and optimize their performance. Future research should focus on refining prompt design, enhancing validation methods, and improving input data quality. With these improvements, LLMs could play a transformative role in policy data collection, validation, and analysis, providing policymakers with more timely and comprehensive insights into the global innovation landscape.

ABSTRACT. Thematic Panel: AI in Science, Technology, and Innovation Policy Studies – Paradigm Shift or Necessary Compromise?

Panel Overview Artificial Intelligence (AI) has rapidly transformed the landscape of science, technology, and innovation (STI) studies. Its profound capabilities are reconfiguring the way we collect, analyze, and interpret data, leading to breakthroughs in STI policy and research methodologies. As we gather at the 10th Biennial Atlanta Conference on Science and Innovation Policy, our thematic panel will dive into how AI is reshaping the contours of STI studies, explore the opportunities it presents, and examine the challenges that come with its integration into research frameworks. This discussion is anchored by guiding questions on AI literacy, transparency, and its effects on both the quality and trajectory of STI research.

The panel seeks to address pressing questions about AI’s role in the study of innovation and technology policy. Our guiding inquiries are: How does AI change the nature of STI studies? Is the infamous “AI black box” a necessary trade-off for innovation, or is it merely the beginning of a new, transparent paradigm in research? Furthermore, what kind of AI literacy is essential to ensure that STI studies remain scientifically sound, ethically grounded, and impactful? The integration of AI into STI research promises an era of immense potential, yet it also raises important issues of trust, accountability, and epistemic reliability. These questions are fundamental as we explore the transformative nature of AI-driven methods in policy research—methods that are increasingly being adopted globally in industry, academia, and government alike.

Presentation Summaries Firms’ Knowledge Disclosure: Website, Publication, and Patent Data This presentation will delve into the changing landscape of corporate knowledge disclosure through various mediums, including websites, academic publications, and patents. By leveraging AI and large datasets, the presenter will illustrate how firms use these avenues to communicate their innovation activities and how AI can enhance the extraction and comparison of disclosed information. The talk will focus on how AI contributes to building a comprehensive picture of innovation by examining patterns of disclosure that are otherwise complex and dispersed. AI-Driven Innovation Measurement: Testing the Limits of Large Language Models (and Knowledge Graphs) for Scaling the Mapping of Business Innovations Large Language Models (LLMs) are at the heart of the evolving AI toolkit for STI analysis, but how far can they really go in automating innovation measurement? This presentation will explore the integration of LLMs with knowledge graphs to enhance our understanding of business innovation activities. It will critically assess the ability of LLMs to map firm-level innovations at scale, presenting both the opportunities and challenges encountered in testing these technologies for real-world STI policy applications. What are the limits, and how do they inform the future trajectory of innovation measurement? Large Language Models in Data Generating Processes for Innovation Policy Studies: Method Development for an International Database This presentation will discuss the development of methodologies for generating data that informs innovation policy, focusing on international databases. By using AI, particularly LLMs, researchers can create consistent, cross-country datasets that provide richer insights into science and technology trends. However, challenges around transparency, bias, and reliability of data generation will be examined to highlight the complexities of using AI in building foundational policy datasets. 4. Analyzing Technologies Instead of Counting Patents! Using AI to Cluster Patents by Technologies of Enterprises Traditional methods of assessing technological innovation often involve counting patents, which may miss the nuances of actual technological advancements. This presentation proposes an AI-based alternative: clustering patents according to the underlying technologies they represent. By using AI to understand patent content, researchers can generate more meaningful insights into the technological strategies of enterprises, allowing policy-makers to better assess innovation dynamics and make informed strategic decisions. The presentation will reflect on the potential benefits, but also pitfalls of AI-based approaches.

Panel Discussion: The Path Ahead for AI in STI The panel will explore AI's transformative impact on STI research, focusing on challenges such as the opacity of advanced AI models and the balance between leveraging their power and maintaining transparency for evidence-based policymaking. Audience participation is encouraged to enrich the discussion on the "AI black box" and its implications for research credibility.

As AI capabilities expand, the need for AI literacy among researchers and policymakers becomes vital to ensure scientific integrity and maintain public trust in policy outcomes. The panel will discuss essential AI knowledge required for effective STI research and innovation policy.

We will also examine whether AI is ushering in a new paradigm for STI studies, pushing beyond traditional methodologies towards novel approaches for understanding innovation systems. Presentations will demonstrate how AI models, particularly LLMs, are being used for STI analysis, highlighting both their transformative potential and the complexities involved.

As we celebrate the 10th Biennial Atlanta Conference on Science and Innovation Policy, this thematic panel aims to reflect on AI's role in navigating the complexities of technology and innovation. By addressing AI literacy, transparency, and innovation, we hope to foster conversations that will shape the future of STI policy and research. Participants are invited to contribute, rethink, and advance the potential of AI for informed and effective innovation policies.

ABSTRACT. Traditional methods of assessing technological innovation often involve counting patents, which may miss the nuances of actual technological advancements. Usually, a bundle of patents protects one technology or the technological portfolio of a company. One the one hand, technology development is an evolutionary process within a company that might have a radical or at least “new-to-the-firm” innovation at one point and further improvements or technological advances of these at later stages. On the other hand, the IP protection should split the risk of a technology to be either attacked, circumvented or re-engineered by filing more than one patent document per unit of invention – be it the initial radical or just an incremental step. Technologies might have different propensities to patent, but also different propensities to split. So far, innovation statistics and innovation analytics has mostly used a patent-counting approach (one patent = one technology/invention). Conceptually, however, this can be challenged, given the reality described above. To tackle this problem, the idea is to identify bundles of patents per company (patent clusters), not only to improve the patent/innovation statistics, but to be able to compare technology portfolios in particular technological domains across companies/enterprises, in addition. Hence, this presentation proposes an AI-based alternative to simply counting patents by clustering patents per company/enterprise according to the underlying technologies they represent. As we do not have annotated documents at large scale that allow us to train a model (machine-learning), we developed an alternative approach using the LLaMA-Model to identify technologies in our corpus of worldwide patent documents (titles and abstracts). We follow a two-step approach to accomplish our task of clustering patents and to compare them across companies/enterprises. In a first step, we identify technologies in the patent corpus, which are represented as clusters of synonym words. We use PATSTAT as a source for our patent data and restrict the data to patent applications filed since 2010 at the transnational level. The whole step 1 could be seen as a topic model; the task is to identify technologies (topics) in the texts. But it's not a "standard" topic model, because we do not aim to assign a patent (text) to a technology (topic). We therefore prefer to call it "keyword extraction" and "synonym grouping" to avoid misunderstanding. This serves as a preparation for step 2. We allow that some outlier technology candidates stay unclustered. If we force to assign all technology candidates to (at least) one cluster even when the probability is low, the clustering quality gets much worse. Anyway, it is unnecessary, because our purpose in step 1 is to detect the existing technologies. It does not matter even if one header word cannot cover all possible wordings! In a second step we perform the patent clustering based on a similarity search, combined with a keyword search. For each patent of a company, we assign the binary (yes/no) information, if the patent belongs to a specific technology cluster that we identified in step 1. We use additional information like IPC class or citation links to further narrow down the potential similarity of patent documents per company. It needs to be stressed that we should not conclude "this patent belongs to the technology" from step 1. The experiment is not designed for that purpose. All we can say from this step is "in the texts, technologies xxx, yyy, and zzz are recognized". This does not guarantee the completeness (there might exist other unidentified/missed technologies), and exclusiveness between identified technologies. Our experiments do not support the causation link ("therefore"-reasoning), for example from the keyword "wind turbine" to the technology "wind energy"! In consequence, validation and plausibility checks are needed and are even crucial for approaches like this. Standard methods or procedures to conduct such checks have not yet been established. While our methodology follows a structured approach, based on conceptual thoughts (theorizing/hypothesizing) and produces reasonable results, we still cannot conclude that the results are valid. LLMs have been shown to “hallucinate” by making up things. In this case, the challenge is an assessment scheme for the validity of the results and the scrutiny of the established scientific perspectives in STI policy analyses. By using AI to understand patent content, researchers can generate more meaningful insights into the technological strategies of enterprises, allowing policy-makers to better assess innovation dynamics and make informed strategic decisions. But how reliable and valid are these insights? However, established scientific approaches might be challenged by the new methods and the new possibilities based on LLMs. As the LLMs are “black boxes” themselves, new scientific perspectives might become necessary. The presentation will use the suggested method to reflect on the potential benefits, but also on pitfalls of AI-based approaches in this context, thereby trying to refer to the three overarching questions of the thematic panel "AI in Science, Technology, and Innovation Policy Studies – Paradigm Shift or Necessary Compromise?".

ABSTRACT. Life scientists in universities are subject to an increasing number of rules and regulations stemmed from within and outside. Internally, university policies and procedures are established not only to meet external regulatory demands on transparency, openness and research ethics (Archibald & Feldman, 2008; Johnson, 2020), but also to ensure their own survival and growth. Externally, universities are subject to rules from regulatory, funding and accrediting entities at multiple levels. Research investigating the effects of regulations on scientists’ research outputs has produced mixed results. Some work argues that administrative burden produced by regulations generates increased burden for scientists, which may create barriers to resource and knowledge exchange, delay or sabotage research activities, and interfere with the traditional norms of collaborative science (Bubela et al., 2015; Eisenberg, 2001; Yeh et al., 2017). These effects are generally linked to lower science production. On the other hand, some argue that the negative impact of regulations on research process is overestimated or, at a minimum, are poorly understood and deserve greater attention (Derrick & Bryant, 2013; Mishra & Bubela, 2014). Moreover, many of the rules are designed to serve important rationales such as to conserve of biodiversity, protect of property rights, human and environmental safety and health, and facilitate open (Fusi et al., 2019; Kamau et al., 2010; Rodriguez, 2005). The administrative burden theory is deemed a viable framework to understand rule burden in the context of university research (Bozeman et al., 2020), although it is originally used to examine the effect of rule burden for citizen access to public programs in citizen-state interactions (Fox et al., 2020; Heinrich, 2016; Moynihan et al., 2015). In the context of academic research, administrative burden includes the time and efforts scientists invest in learning about various requirements and specifications, completing applications and reports as well as any psychological strain they may experience in the process. The increased job demands not only add to employees’ workload, enhance role ambiguity, but also occupational stress and phycological risks, which could impair the overall quality of academic work among academic faculty (Boyd & Wylie, 1994; Pace et al., 2021; Tight, 2010). An adjacent concept to administrative burden is organizational red tape. While both administrative burden and red tape can be burdensome, they are distinct in their functionality (Campbell et al., 2023; van Loon et al., 2016). Only rules that entail wasteful consumption of organizational resources without advancing organizational goals are considered red tape (Bozeman & Feeney, 2014; Campbell et al., 2023). At the individual level, a recent synthesis of red tape studies indicates that the effect of red tape on job involvement and employee performance is negligible (Blom et al., 2021). In fact, high-quality rules and the ways in which rules are implemented can even foster rule compliance and improve job satisfaction (Borry et al., 2018; DeHart-Davis et al., 2015). In the context of academic research, some rules and regulations are designed for important public policy reasons, such as to protect human and environmental safety, as well as intellectual property (Bozeman & Youtie, 2020; Woelert, 2023). These rules, albeit burdensome, may not necessarily raise barriers to material sharing nor hinder knowledge diffusion and research progress (Mowery & Ziedonis, 2007; Stern, 2004; Walsh et al., 2003). Accompanying increased administrative rules and enhanced formalization of institutional processes, universities worldwide are experiencing the expansion of bureaucracy (sometimes “administrative bloat”) characterized by increased administrative positions and expenditure, as well as unbundling of functions, services and faculty roles (Macfarlane, 2011; McCowan, 2017; Ramirez & Christensen, 2013). In terms of research management, professional staff are found to have a positive role in securing and managing funds through informing project designs, providing trainings and grounding rules on proposal writing for researchers (Beime et al., 2021; Ito & Watanabe, 2020). To understand whether the ways in which regulations are expressed and realized influence academic scientists’ research performance, this study examined the two aspects of rule: rule burden and rule legitimacy drawing on insights from both the administrative burden and red tape literature. We hypothesize that rule burden is negatively related to research performance and rule legitimacy is positively associated with research performance. We also hypothesize that organizational support in regulation management has a moderating role between rule legitimacy and future research performance. To test our hypotheses, we use a unique database combining survey data and bibliometric data retrieved from the Scopus database in 2024. The survey was administered in 2016 and 2017 to a national sample of 3,933 tenured and tenure-track academic scientists in the United States in three disciplines—marine biology, entomology, and ecology—from research universities classified as having “very high” and “high” research activity as defined by the 2010 Basic Carnegie Classification. We measure rule burden, rule legitimacy and organizational support using survey data.The cumulative number of publications and citations in the four years after the survey are used as dependent variables measuring research performance. Using zero-inflated negative binomial model, we found that neither perceived rule burden nor rule legitimacy is not predictive of academic scientists’ research performance. Additionally, support from university professional staff is positively associated with future research performance. This article contributes to the understanding of the relationship between rule and research performance of academic scientists by connecting the public administrative literature on rules and red tape to science policy literature. Our findings suggest that although rules and regulations may increase workload for academic scientists, they do not necessarily hinder research performance. The compliance burden can be compensated by the legitimate causes rules may serve. Moreover, challenging job demands derived from increased rule compliance burden might provide opportunities to enhance proficiency in job-related tasks, and to demonstrate competences that can get rewarded (Crawford et al., 2010). The findings also suggest that research universities should invest in specialized personnel to assist scientists in navigating the complex regulatory environment.

ABSTRACT. Motivation The global circulation of the scientific workforce is a well-established phenomenon that plays a critical role in the generation and dissemination of knowledge. It also serves as a catalyst for innovation. As globalization intensifies, scientists are becoming more mobile than ever before. The international relocation of scientists, who carry with them unique expertise and knowledge, strengthens research systems by fostering knowledge exchange across universities and regions. While debates persist about the impact of mobility on individual career advancement, there is broad consensus that mobility benefits the research ecosystem as a whole. Numerous policies at various levels actively promote the mobility of scientists to enhance the efficiency and productivity of their respective research systems. However, most research on scientist mobility has focused on what might be considered "normal" periods of scientific development. The effects of global exogenous shocks, such as a pandemic, on the international mobility of scientists remain an open and underexplored question. This study seeks to address this gap by examining how the COVID-19 pandemic has influenced the international mobility of scientists, focusing specifically on the field of Computer Science. We aim to assess whether and in what ways the pandemic disrupted traditional patterns of international mobility. Additionally, this study investigates the applicability of the traditional Push-Pull model in explaining the movement of scientists during this unprecedented global crisis. Specifically, we explore the immediate impact of COVID-19 at both the country and individual levels, seeking to understand how country-level, policy-level, and individual-level factors interacted with the adverse effects of the pandemic. Data and research method The primary dataset for this study is the DBLP-Citation-network V14 (referred to as DBLP), a comprehensive bibliometric dataset released in February 2024. This dataset, built on the original DBLP, encompasses 5.26 million papers and over 36 million citation relationships in the field of Computer Science. DBLP employs a hybrid disambiguation approach—combining algorithmic techniques with manual error correction—to achieve highly accurate author disambiguation. To extract country-level information from authors' address strings, we use a combination of string-matching algorithms and large language models. To validate the accuracy of this process, we manually reviewed 500 random address strings, achieving weighted precision, recall, and F1 scores of 0.948, 0.942, and 0.943, respectively. These results confirm a satisfactory level of accuracy in identifying country information. We use shifts in authors' affiliations across countries as a proxy for international mobility events. Bibliometric data offers an extensive, data-driven perspective on researcher mobility. By tracking changes in researchers' affiliations as reflected in their publications, we provide a comprehensive analysis of global scientific mobility. To examine the impact of the pandemic, we apply a novel causal inference approach: a difference-in-differences method with a continuous treatment variable. This method enables us to measure the disruption caused by COVID-19 on international mobility in Computer Science between 2013 and 2022. To address potential endogeneity concerns, we use the latitude of each country's capital city as an instrumental variable. Latitude serves as a valid instrument because it is correlated with temperature—which affects the transmission of COVID-19—but is exogenous to the pandemic itself. Major findings This study examines the relationship between a country's COVID-19 infection rate and the international mobility of scientists in Computer Science. Our findings indicate that an increase in a country's infection rate by one standard deviation leads to a significant reduction in population flow. Specifically, we observe a 5.35% decline in international mobility, which can be further broken down into a 3.57% decrease in the inflow of scientists and a 6.47% decrease in the outflow. These results are robust across alternative metrics, such as COVID-19 infection and death rates. To address endogeneity concerns, we employ the latitude of a country's capital city as an instrumental variable. This approach allows us to control for confounding factors that may influence both COVID-19 infection rates and scientific mobility. Our analysis reveals that the impact of COVID-19 infections is most pronounced in the second year following the outbreak. Additionally, we find that the decline in inflow is more significant in developed countries, while the reduction in outflow is more pronounced in nations with stringent anti-COVID-19 policies. These findings highlight the complex interplay between public health crises, policy responses, and global scientific mobility. Conclusion This study highlights the significant impact of the COVID-19 pandemic on the international mobility of scientists, particularly in the field of Computer Science. By leveraging a comprehensive bibliometric dataset and employing robust causal inference methods, we provide strong evidence that the pandemic disrupted established patterns of global scientific mobility, with notable declines in both the inflow and outflow of researchers. Our findings suggest that a country's infection rate is a key determinant of these mobility trends, with higher infection rates leading to reduced international movement of scientists. The differential effects observed—where developed countries experienced greater declines in inflow and nations with stringent anti-COVID-19 policies saw more substantial decreases in outflow—offer important insights into the unequal impact of the pandemic across different contexts. These results also suggest that while the traditional Push-Pull model remains broadly applicable, it must be adapted to incorporate additional factors, such as public health crises and corresponding policy responses, to fully explain mobility patterns during periods of global disruption. In conclusion, this research offers a data-driven perspective on the pandemic's disruption of the global scientific workforce and underscores the need for adaptive strategies to sustain international mobility in the face of future crises. Policymakers and research institutions should draw on these findings to develop resilient systems that support global knowledge exchange, even in times of unprecedented challenges. Future research could build upon this work by exploring the long-term recovery of scientific mobility and its implications for innovation and collaboration in a post-pandemic world.

ABSTRACT. Economists have long viewed innovation as the key driver of long-term economic growth. However, recent research has shown that discrimination and systemic racism have throttled the country’s growth path. This project seeks to evaluate how inequitable access to education limits full participation in the innovation pipeline.

Recent literature across the social sciences has consistently found that systemic factors limit who has the opportunity to innovate and that these disparities have quantitatively large detriments to society. Bell et al. (2017/2019) were the first to show, using linked taxpayer-inventor data, that children from lower-earning families, women, and racially disadvantaged groups had meaningfully lower innovation rates.

Given careers in innovation are often viewed as having risky payoffs that come later in life, we seek to uncover whether bigger loan burdens play a role in pushing talented students away from career paths as inventors. On average, Black students hold $7,000 more student loans than White students when they graduate, and this gap grows over time as repayment schedules diverge. Women hold 66 percent of the total outstanding student loan debt.

To better understand sources of disparities in innovation, we undertake a large-scale record linkage effort of national individual-level data on inventors from administrative records of the US Patent Office to student loan data from credit bureau records. The linkage allowed us to view 20 years of complete credit histories for almost 250,000 US patent inventors that have ever lived in California (about ⅓ of all US inventors). Applying imputation methods for race and gender, we find that 88% of our linked inventor sample is White or Asian (relative to 60% of the general CA population) and 17.5% are women.

Research is proceeding in two phases. In the first phase, which will be the primary focus of this talk, we document a number of stylized facts about credit histories and student loan usage among inventors. For instance, people with credit scores in the top 10% of the distribution are ten times as likely to invent as those with below-median credit scores. Inventors take out more loans of all types than the average person of a similar age. During their early 20s, inventors accumulate substantially more student loan debt than most people, but also pay off the debt more quickly; by age 40, inventors actually hold substantially less student loan debt than their peers. Contrary to what one might expect, event studies show that credit scores increase until the time of patenting, and then somewhat level off after patenting.

These descriptive results will lay the foundation for the second, more causal-oriented stage of research. We will test whether there is a causal relationship between education access and later-life innovation, and whether this effect can explain later-life disparities in innovation by race. We will use cutting-edge quasi-experimental techniques leveraging a series of policy changes to causally identify the effect of student loans burdens on later life innovation. Estimates will inform our understanding of the sources of disparities in innovation.

ABSTRACT. Research Questions In this paper, we study the question of whether government support of business enterprise R&D (aka ‘BERD’) affects innovation output qualitatively. The theoretical background is the extensive empirical literature that has evaluated the effectiveness of fiscal R&D incentives. The findings of this longstanding policy-oriented work indicate that public BERD support, broadly speaking, has a positive effect. Bloom et al. (2019:180) conclude that RD&I tax credits and direct public funding ‘seem the most effective measures in innovation policy toolkit’. Much of the work has focused on the return to the public investment to spur private sector R&D investment. This ‘input additionality’ literature indicates that the effect is positive and sizeable, depending on specific settings (eg Almus and Czarnitzki, 2003). This paper contributes to the work that has tried to evaluate the effect on the firm’s innovation output itself (eg Cappelen et al., 2012). The question of ‘output additionality’ matters if the policy goal is to affect not only the level of innovation in the economy, but also the direction of that innovation, for example to promote more climate-friendly solutions. In this context, the policy background is important too. Government spending on BERD support is large, accounting for roughly 460 billion dollars or 0.21 of GDP in OECD countries according to the latest available numbers (2021, see OECD 2023). Spending has also increased substantially (%50 percent since 2000) and there’s been actively rebalanced between different types of measures. As the level of support is actively adapted and rebalanced between direct and indirect policies, it is important to continue to improve our understanding of the effects that the full range of R&D grants and tax credits is having on innovation in the private sector (OECD, 2021). The proposed paper contributes policy-relevant evidence about the effect that BERD support has on innovation output. In general, this is an area where further research is needed. Broadly, the relatively small number of studies have found that BERD support has positive effects on innovation (eg Dechezleprêtre et al. 2016). However, results vary and tend to be sensitive to a range of study-specific factors (Dimos & Pugh, 2016). The immediate background for this paper, however, is recent work that took a ‘completest’ approach to the question. In two recent studies (Nilsen et al, 2020; Iversen et al, 2024) we find that the effect of BERD support tends to diverge: for firms that have not been R&D active, we find a strong and positive effect along the extensive margin; for incumbent R&D active firms, the effect on the intensive margin is marginal. This divergent effect helps to explain why results from the output additionality work are not clearer. At the same time, it begs the question of whether there is another effect among the important incumbent R&D active firms that we are missing by tallying not only patents but patents together with other IPRs. This is what we study in this paper. Data and Methodology This paper builds on a string of empirical work that has followed the well-documented Norwegian case. Following Iversen et al (2024), we follow the ‘new economics of industrial policy’ by emphasizing improved measurement, careful application of causal inference, and a more ‘nuanced and contextual understanding of the effects of industrial policy (Juhasz et al, 2023). Data integration: Three main data sources yield a comprehensive panel dataset: (i) The Business registry consisting of firm-level data for the full population of active Norwegian limited liability business enterprises. (ii) The database on public BERD support consisting of detailed information about R&D subsidies from all major programs whether direct subsidies (grants) or tax credits.(iii) IPR applications (patents, trademarks, industrial design). These micro-data are linked at the enterprise-level to allow us to study the relation between BERD support and IPR outcomes over a period extending at least from 2002 (when the Norwegian R&D tax credit program started) until 2021. Key dimensions of the underlying industrial organization are traced at the micro-level over two decades. The panel observes firms reporting formal R&D and/or innovation activity, and which firms receive what form of R&D subsidy over time. Measurement: The target literature has proxied innovation outcomes using measures from R&D/Innovation Surveys, from firm-linked patenting (e.g. Bronzini & Piselli,2016) or from both (Cappelen et al. (2012). Our innovation metric is based on a composite dataset for the three types of IPR. This goes beyond a simple patent count approach. But an approach based on IPR counts more generally still gloss over the link between the IPR and the specific R&D policy instruments (eg green patents for direct BERD support on climate mitigation). The paper introduces qualitative aspects of firm level innovation outcomes to address such issues. How then to proxy the quality of innovation output? We derive measures about patent quality including application channel, patent family size, renewal, as well as bundles with trademarks and industrial designs. In this sense, patent quality is a function of a potentially multifaceted trajectory of events following the initial patent filing, leading to methodological questions. Method: Patent quality is framed in terms of an event study design. Efforts are made to mitigate challenges this poses to identification of causal effects in the presence of time-varying confounding factors. To address such challenges, we will draw upon ongoing research in this area (see Iversen at al., forthcoming 2024) that apply causal machine learning methods (Belloni, 2016; Chernozhukoc et al, 2018) to identify and estimate treatment effects in non-linear, dynamic models. Findings Our preliminary findings build on Nilsen et al (2020) and Iversen et al, (2024) that fiscal stimulus tends to have greatest impact on previously non-innovative firms. The impact of support measures, alone or in combination, is on the extensive rather than intensive margin. Findings from the extension into the quality of IPR outcomes is ongoing and have not yet finalized.

ABSTRACT. The proposed paper addresses one of the questions intended to animate presentations and discussions in the Conference: to what extent have science and innovation policy interventions generated economic and social impact? Or from a slightly different perspective: what prevents science and innovation policy interventions to generate economic and social impact?

Drawing from the Brazilian case, we start showing compelling evidence that policies promoting innovation have not been effective. Since the late 1990s science and technology policy in Brazil has established innovation as a priority. Starting with the creation of the Sectoral Funds, in 1999, followed by a number of subvention schemes and fiscal incentives to promote business R&D, such as Green and Yellow Fund (2001), Innovation Law (2004), Lei do Bem (2005) and so on, there was a significant increase in the number of instruments aiming to further innovation. Despite all of these initiatives, the results have been disappointing, as there is no indication that Brazilian companies are improving their innovation performance.

We argue that so far policies have been too much concentrated on the supply side, on measures to induce firms to invest in R&D. The demand side has been neglected; the measures that make firms want to spend on R&D and seek to innovate.

We show that globalization changed the conditions under which the Brazilian economy functioned therefore limiting the effectiveness of growth strategies followed by the country until then. A key feature of globalization, which is the functional integration of geographically dispersed activities of multinational companies, has implications for the economic dimension of the domestic markets of the countries in which these companies operate. As a result, development strategies based on the size of the domestic market became highly compromised. This take is valid not only for Brazil but for Latin America overall.

Failing to adapt to this new scenario brought up by globalization implies that firms have limited stimulus to engage in innovation activities. To be effective, policies must focus on both the supply and demand side. That is, it is necessary to provide physical and human capital necessary to carry out innovation activities, institutions that facilitate innovation, subvention schemes, fiscal incentives, governmental programs that promote innovation, and so on. Nevertheless, it is equally necessary to create an economic environment, particularly a competitive regime, which compels firms to innovate.

ABSTRACT. The “wickedest” problems facing society today will require a revolutionary, not evolutionary, breakthroughs in physics, engineering, neuroscience and other scientific fields, but will remain ineffective if not widely adopted. It is widely recognized that over and over again human behavior, especially contingent, interactive and aggregated human behavior proves to be the point of failure in models in cybersecurity, pandemic relief and mitigation of global warming trends, among others. Social scientists are trained to study precisely what is overlooked in these models, and yet there is little engagement of social scientists on teams to address these issues and relatively little investment in social science research, which accounts for less than 5% of funded research in the US. This is largely because few people outside of social science are aware of the substance or methodologies available to social scientists and do not appreciate the value of social theory and research training for developing research programs or evaluation strategies required to interest people in adopting the technology solutions that are the focus of scientific research programs. Consequently, social scientists, and their expertise, are overlooked in the development of strategic plans and crises mitigation doomed to fail if the people part of the model is not properly scoped or specified. Here we provide examples of the consequences of underinvestment in social research, discuss some of the reasons social science is marginalized and undervalued, and recommend new ways to communicate the value of validated models of human behavior to a broader scientific audience.       

ABSTRACT. Research Questions

The paper addresses critical gaps in the evaluation of innovation policy, particularly concerning the tracking of invention commercialization.

Specifically, the research aims to answer the following key questions: 1. Can a web-based method of linking patents to products serve as a viable means of tracking commercialization outcomes? 2. What are the comparative commercialization probabilities between the Department of Defense's (DoD) SBIR-funded and privately-funded patents? 3. How do factors such as the phase of the SBIR project (Phase I vs. Phase II) and the stage of research (basic, applied, or developmental) correlate with the commercialization of these inventions?

Method

The paper proposes a web-based approach to innovation policy evaluation that overcomes the traditional limitations of tracking the commercialization of inventions. It applied the approach to DID SBIR-funded inventions. The method is structured around three primary steps: 1. Identification of SBIR-funded Patents: The research identifies all U.S. patents that acknowledge support from the DoD’s SBIR program by leveraging the legal requirement that federal funding be disclosed in patent documents. We exploit the Bayh–Dole Act and Federal Acquisition Regulation (FAR) rules to extract this information from the patent text. 2. Tracking Commercialization through Web Data: To assess whether an SBIR-funded invention has been commercialized, we search for online traces that link patented inventions to commercial products. This is achieved by identifying virtual patent marking pages and product information on the websites of the patent assignees. The existence of a web page that associates a patent with a product is treated as evidence of commercialization. 3. Econometric Analysis: We construct a control group of patents with similar characteristics but without SBIR funding. We then perform econometric analyses to compare the commercialization rates of SBIR-funded inventions with those of privately-funded patents. The commercialization probability is measured by tracking whether a patent is directly or indirectly (i.e., through a citation) linked to a commercial product.

The study’s dataset comprises 2,896 DoD-SBIR-funded patents and a control group of 4,622 privately-funded patents, spanning from 1984 to 2018. These patents are matched based on priority year and technological class to ensure a robust comparison. We use regression models to examine the factors influencing commercialization, controlling for various patent-level characteristics such as the number of claims, citations, and geographical patent family size.

Key Findings

The study delivers several important findings that provide insights into the effectiveness of the SBIR program: 1. SBIR-Funded Inventions Are More Likely to Be Commercialized: The results indicate that patents arising from SBIR-funded projects are 17% more likely to be commercialized compared to patents in the control group. This finding suggests that the SBIR program plays a significant role in transitioning federally funded inventions into commercial products. 2. Influence of Research and Development (R&D) Stage on Commercialization: The stage of the R&D project also plays a crucial role in commercialization outcomes. Inventions connected to developmental R&D contracts are much more likely to be commercialized than those resulting from basic or applied research contracts. Specifically, SBIR-funded patents connected to developmental R&D have a commercialization rate approximately 4.9 percentage points higher. 3. Impact of SBIR Phases: The results confirm the importance of the SBIR’s multi-phase structure. Inventions funded during Phase II are significantly more likely to be commercialized than those that only received Phase I funding. (Phase I projects primarily serve as feasibility studies, while Phase II projects receive substantial funding to push the invention closer to commercialization.) We find that Phase II projects have a commercialization probability much higher than comparable privately-funded inventions. 4. Effect of the Phase II Plus Policy: The paper highlights the success of the 2000 policy reform, which introduced Phase II Plus, providing additional funding to incentivize commercialization. SBIR-backed patents that received Phase II funding after the policy change were more likely to be commercialized, indicating that the policy effectively encouraged firms to transition their inventions to market. 5. Commercialization Pathways: The study explores two pathways for commercialization—direct and indirect. Direct commercialization occurs when a patent funded by the SBIR program is directly linked to a product. Indirect commercialization happens when an SBIR patent is cited by another patent that protects a product. Interestingly, about 40% of the indirect commercialization cases are driven by self-citations, implying that the firms receiving SBIR support continue to invest in the development of their own inventions.

Conclusion

The web-based approach proposed in the paper provides a scalable method to track commercialization outcomes. The findings suggest that the DoD’s SBIR program is an effective tool for fostering the commercialization of government-funded inventions. Importantly, the study highlights the long-term benefits of public funding, as many SBIR-funded inventions contribute to commercialization indirectly through subsequent innovations. The research also underscores the critical role of Phase II funding and policy changes, such as the Phase II Plus program, in ensuring the commercial success of federally funded projects.

Future work could expand this method to evaluate other public innovation programs or explore the commercialization of non-patented inventions. The study also opens the door to using web-based data for broader policy evaluations in areas such as technology transfer and intellectual property.

ABSTRACT. Background The commercialization of university research has become increasingly crucial for technological innovation and economic development. Following the success of the Bayh-Dole Act in the United States, many countries have adopted similar IP rights reforms to enhance university technology transfer. In China, despite implementing comparable IP rights policies since 2000 that grant universities patent rights and establish revenue-sharing mechanisms, the commercialization rate of university patents remains disappointingly low at below 5%. This limited success presents an intriguing puzzle: why do similar IP rights reforms yield dramatically different outcomes across institutions? The challenge is particularly complex in China's institutional context, where university-generated intellectual property has traditionally been classified as state-owned assets, subjecting technology transfer decisions to rigorous government oversight and auditing requirements. This institutional arrangement creates significant risks for universities in asset valuation and transfer pricing, potentially undermining the intended effects of IP rights reform. The variation in technology transfer performance across universities, despite facing similar institutional constraints, suggests that the effectiveness of IP rights reform may depend on its interaction with other organizational and environmental factors. Understanding these complex relationships is crucial for improving the design and implementation of technology transfer policies. Research Questions Given this context, our study addresses three interrelated research questions: First, why do similar IP rights reforms yield dramatically different technology transfer outcomes across universities operating under the same national policy framework? Second, what configurations of institutional arrangements, organizational capabilities, and environmental conditions enable successful university technology transfer? Third, how does the effectiveness of IP rights reform implementation depend on its interaction with universities' internal capabilities (scientific discovery, technological application, human capital, and reputation) and external conditions (regional economic vitality)? These questions move beyond the traditional focus on policy adoption to examine the complex conditions under which IP rights reforms actually deliver their intended outcomes. Research Methods To unravel this puzzle, we employ fuzzy-set Qualitative Comparative Analysis (fsQCA) to examine 28 cases from China's prestigious 985 Project universities. Our research design offers two key methodological innovations. First, we measure technology transfer performance through actual contract values rather than conventional patent-based metrics, thereby capturing the true extent of commercialization success while avoiding the limitations associated with patent counting and valuation ambiguities. Second, we adopt a configurational approach to analyze how IP rights reform interacts with multiple conditions to produce successful outcomes. These conditions include scientific discovery capability (measured by research output and quality), technological application capacity (indicated by applied research and development activities), human capital (represented by research staff qualifications and industry experience), university reputation (based on national and international rankings), and regional economic vitality (reflected in industrial development and market demand for innovation). This comprehensive analytical framework allows us to identify complex combinations of conditions that enable or constrain successful technology transfer. Findings Our analysis reveals several important patterns. First, the necessity analysis shows that no single condition, including IP rights reform implementation, is sufficient alone to guarantee high technology transfer performance. However, the persistence of traditional state-owned asset management approaches (without effective IP rights reform implementation) emerges as a necessary condition for poor performance, highlighting the fundamental importance of institutional arrangements. Second, our sufficiency analysis identifies three distinct pathways to successful technology transfer: (1) effective implementation of IP rights reform combined with strong accumulated knowledge capital, enabling universities to leverage their existing research capabilities; (2) reform implementation paired with robust research capacity and complemented by either strong human capital, established university reputation, or favorable economic conditions; and (3) reform implementation coupled with strong regional economic vitality, which can compensate for relatively modest university capabilities. These findings suggest that while effective IP rights reform implementation is crucial, its success depends on specific organizational and environmental configurations. Contribution This study makes several significant contributions to our understanding of university technology transfer. First, we advance the methodological approach by directly analyzing technology transfer contract values, addressing fundamental limitations in previous studies that rely solely on patent metrics. This innovation provides a more accurate picture of actual commercialization success and captures various forms of technology transfer that might not be reflected in patent statistics. Second, we enrich theoretical understanding of research management and science & technology policy by demonstrating that the effectiveness of IP rights reform depends on complex configurations of institutional, organizational, and environmental factors. This finding challenges the simple policy transfer approach that assumes similar reforms will yield similar results across different contexts. Third, our identification of multiple pathways to successful technology transfer provides valuable insights for policymakers and university administrators. It suggests that successful implementation of IP rights reform requires careful consideration of institutional contexts and complementary capabilities. Universities may need to develop different strategies based on their specific strengths and environmental conditions to achieve effective technology transfer. Finally, our research contributes to the broader literature on institutional reform by showing how the same policy instrument can yield varying outcomes depending on its interaction with other organizational and environmental factors. These insights are particularly valuable for emerging economies seeking to enhance their innovation capabilities through institutional reforms.

ABSTRACT. 1. Introduction

Universities drive innovation through intellectual property, technology transfer, and research collaboration, essential to national and regional innovation systems (Nugent & Chan, 2023; Perkmann et al., 2013). The shift to a knowledge economy has expanded universities’ missions to include fostering economic and social development (Compagnucci & Spigarelli, 2020).

This study examines how internal characteristics of Brazilian federal universities—such as size, governance, and scientific output—affect their innovation activities. Using data from all 69 federal universities, we analyze factors influencing patenting and research collaborations. Our findings highlight key institutional traits linked to innovation, providing insights into how universities contribute to Brazil’s knowledge economy. The article covers literature, methodology, results, and conclusions.

2. Literature Review

Universities are increasingly integral to innovation systems, translating research into commercial applications through patenting, technology transfer, and industry partnerships (Aghion & Howitt, 2008; Mansfield & Lee, 1996). This shift is aligned with the Third Mission (TM), which emphasizes universities' contributions to economic and social development, and with the growing proximity of research outputs to outcomes and innovation (Compagnucci & Spigarelli, 2020).

Factors influencing university patenting include its size, the institutional support, and fields of research, with larger and older institutions often patenting more, particularly in engineering and chemistry (Fisch et al., 2015; Yamaguchi et al., 2019). Additionally, patent quality, as seen in Brazilian universities, is significant, as impactful patents contribute more effectively to innovation ecosystems (Méndez-Morales et al., 2022).

The literature also suggests a positive correlation between publishing and patenting, challenging the notion that commercialization detracts from research output (Stephan et al., 2007). Instead, patenting often complements scientific productivity, particularly in knowledge-intensive fields (Breschi et al., 2008). However, successful academic patenting requires supportive institutional frameworks, as highlighted in studies on emerging economies (Murat AR et al., 2023).

3. Method

This exploratory and descriptive study uses quantitative data from Brazil's Federal Court of Auditors (CGU) audit project "Economics of Innovation in Federal Public Universities", and bibliometric data extracted from large data sources. The audit examined how strategies within Brazil's innovation ecosystem contribute to innovation-driven research, considering the Triple Helix model. Complementary bibliometric data allowed to correlate primary data collected by CGU with secondary data collected by the authors.

3.1 Research Question and Variables

The study aims to determine:

Research Question: a) To what extent are the constitutive characteristics of Brazilian federal universities correlated with their performance variables? b) What are the factors that most influence technological production and transfer?

Variables Used:

Constitutive Variables: University size (number of faculty), age, and geographic location. Performance Variables: Scientific publications, citations weighted per area of knowledge and timeframe, intellectual property (IP), patent citating scholarly outputs, articles co-authored with companies, technology transfer rates, and internal regulatory guidelines towards IP and tech-transfer.

3.2 Data Collection

Data were gathered through:

Census Surveys: Questionnaires covering intellectual property, technology transfers, and institutional strategies related to innovation. Secondary Data: Information from the Ministry of Education, Sucupira Platform (from the Brazilian Federal Agency for Support and Evaluation of Graduate Education - CAPES), Scopus database, and Scival platform.

3.4 Data Analysis

Bibliometric analysis of publications (2011–2022) was performed using Scopus and Scival. Key indicators related to technology transfer, intellectual property and institutional regulations were selected. Pearson and Spearman correlation analyses and Qualitative Comparative Analysis (QCA) were conducted.

4. Empirical Context

Brazil’s higher education system includes over 2,500 institutions, with about 87% private and 13% public. Despite their smaller share, public universities, particularly federal ones, are central to scientific research and PhD training, employing more than half of Brazil’s PhD faculty and contributing significantly to scientific output (INEP, 2022). In 2023, Brazil ranked 13th globally in scientific production, largely due to federal universities, which excel in health sciences, agrarian sciences, and engineering (Clarivate, 2024).

Federal universities, under the Ministry of Education (MEC), have standardized governance, with deans appointed by the President of Brazil following academic consultations. These institutions vary widely in size, age, and regional distribution, reflecting Brazil’s economic and infrastructural diversity. For example, some are over a century old, while others were established within the last 20 years, highlighting the range in institutional characteristics.

This study is the first comprehensive analysis covering a large set of indicators of scientific, technological and institutional regulation of all Brazilian Federal universities.

5. Results

Here are some findings yet to be discussed and developed: • There is no correlation between constitutive characteristics of universities (age and tradition, location and size) and their performance on variables of outputs. • The are strong correlations between scientific indicators (as for scholarly outputs, citations per field and timeframe and co-authorship with companies) and IP and tech-transfer. • The presence of internal regulations about IP and tech-transfer act as a factor that positively influences the rates of IP and tech-transfer. • There is evidence that tradition (in terms of age, size and location in more developed states) does not serve as a differentiating factor for universities being more oriented toward IP and tech-transfer.

ABSTRACT. Proof of Concept Centers (PoCCs) are an increasingly prevalent strategy for university technology translation (Bradley et al., 2013). PoCCs support technology transfer and commercialization by providing academic innovators with training, mentorship, and non-dilutive funding to de-risk technologies and reduce information asymmetries that could create barriers to commercialization (Gulbranson & Audretsch, 2008). While PoCCs are associated with better translational outcomes (Hayter and Link, 2015), there is limited evidence or understanding of how PoCCs impact commercialization trajectories across a wide range of technologies and for PIs with different backgrounds (Munari and Toschi, 2021; Battaglia et al., 2021; Abudu et al. 2022).

This paper asks how projects supported by PoCCs with different principal investigators (PIs) and technology characteristics vary in reaching three intermediate commercialization outcomes: patents filed, licensing agreements, and Small Business Innovation Research (SBIR) or Small Business Technology Transfer (STTR) award applications by startups founded around PoCC-supported technologies. These three outcomes are important to the technology transfer process and are of interest to the research and policy communities (Link, 2024). Our analysis shows that initial paths to outcomes within PoCCs vary based on technology type and the PI’s academic title, gender, and commercialization experience. These findings make novel contributions to the literature on PoCCs and technology transfer.

We use a dataset of 275 biomedical technologies that have received programmatic funding from 6 distinct PoCCs supported by the National Institute of Health (NIH): three centers funded under the NIH Centers for Accelerated Innovations (NCAI) beginning in 2014, and 3 hubs from the first iteration of the Research Evaluation and Commercialization Hubs (REACH) program beginning in 2015. For each PoCC-supported technology, our dataset tracks the trajectory of the technology over a period of 4 to 9 years, observing outcomes associated with the core technology supported.

The dataset is unique in the timespan of observations following funding, the scope across multiple PoCCs, and the ongoing efforts to update and ensure validity of outcomes. These data are summarized by Anderson et al. (2022) in a programmatic review but have not previously been used in an empirical analysis. While Munari & Toschi (2021) tracks outcomes for multiple European PoCCs, data of this type have not been previously used in the United States.

To estimate the impacts of project-level characteristics on patents filed, licensing agreements, and SBIR/STTR awards, we employ a Tobit estimation to predict the number of outcome events. We control for NIH PoCC program funding, technology type, and the PI’s title, gender, and commercialization experience. Controls for the six PoCCs in the data are included and standard errors are clustered at the center level. For robustness, models are also run with ordinary least squares (OLS) and Probit estimations, which support initial fundings.

Results show important variation in outcomes achieved by PoCC-supported projects based on PI characteristics and technology type. PIs with prior commercialization experience had higher numbers of SBIR/STTR applications and were much more likely to pursue licensing agreements but had no significant differences in the number of patents filed. Full professors more likely to license technologies, while assistant professors were more likely to file patents or apply for SBIR/STTR awards. Associate professors had low rates across all outcomes. Female PIs are estimated to be more likely to patent and apply for SBIR/STTRs but had fewer licensing agreements. However, we found no statistically significant differences between male and female PIs. Among technology categories, drug projects had fewer outcomes than other technology types, likely due to lengthy timelines required for development and regulatory approval. Medical devices and diagnostics had the highest patenting rates. We also observe a strong positive correlation between the level of NIH funding and the frequency of licensing agreements, with a weaker yet positive association with patents filed and SBIR/STTR applications.

The experiences of projects in NIH PoCC programs support show the importance of considering different translational pathways and trajectories which may be relevant to different projects. They also support the efficacy of PoCC funding in spurring academic research with commercialization potential towards translation activities. Policymakers can use results to better support a range of innovators and nascent technologies.

Works Cited

Abudu, R., Oliver, K. and Boaz, A. (2022) ‘What funders are doing to assess the impact of their investments in health and biomedical research’, Health Research Policy and Systems, 20(1). https://doi.org/10.1186/s12961-022-00888-1.

Anderson, B. J., Leonchuk, O., O’Connor, A. C., Shaw, B. K., & Walsh, A. C. (2022). Insights from the evaluations of the NIH Centers for Accelerated Innovation and Research Evaluation and Commercialization Hubs programs. Journal of Clinical and Translational Science, 6(1), p. e7. https://doi.org/10.1017/cts.2021.878.

Battaglia, D., Paolucci, E., & Ughetto, E. (2021). The role of Proof-of-Concept programs in facilitating the commercialization of research-based inventions. Research Policy, 50(6), 104268. https://doi.org/10.1016/j.respol.2021.104268

Bradley, S.R., Hayter, C.S. and Link, A.N. (2013). Proof of Concept Centers in the United States: an exploratory look. The Journal of Technology Transfer, 38(4), p. 349–381. https://doi.org/10.1007/s10961-013-9309-8.

Gulbranson, C.A. and Audretsch, D.B. (2008). Proof of concept centers: accelerating the commercialization of university innovation. The Journal of Technology Transfer, 33(3), p. 249–258. https://doi.org/10.1007/s10961-008-9086-y.

Hayter, C.S. and Link, A.N. (2015) On the economic impact of university proof of concept centers. The Journal of Technology Transfer, 40(1), p. 178–183. https://doi.org/10.1007/s10961-014-9369-4.

Link, A.N. (2024). Public Sector Technology Transfer. Edward Elgar Publishing. Available at: https://www.e-elgar.com/shop/usd/public-sector-technology-transfer-9781035310524.html

Munari, F. and Toschi, L. (2021). The impact of public funding on science valorisation: an analysis of the ERC Proof-of-Concept Programme. Research Policy, 50(6), p. 104211. https://doi.org/10.1016/j.respol.2021.104211.

ABSTRACT. Corporate research laboratories of the 20th century advanced science and technology enormously. The iconic Bell Labs alone won eleven Nobel prizes and four Turing awards and launched the fields of microelectronics and information science. Some such laboratories still exist today, but with a much-diminished role in the national research ecosystem [1]. In this talk, building on our recent modern synthesis of the nature and nurture of research [2], we distill the reasons such laboratories were so productive into a set of funding and organizing principles for how they might be re-established in a 21st century form. Our aim is for these principles to catalyze debate on policy opportunities, which we believe could be transformative for national research productivity, and on policy challenges of practical implementation.

A first principle is “inspire, but don’t constrain, research by particular use.” As is well known by corporate scientists and engineers, intimate familiarity of use—human-desired interactions with the real world—stimulates a rich space of problems associated with such use. As is also well known by corporate scientists and engineers, however, the immediacy of solving exactly those problems, of constraining research to particular use, is typically paramount. Occasionally (often, even, if one is intentionally looking), however, surprise intervenes: an unexpected new idea for solving the problem more generally but less immediately; or an unexpected new observation that contradicts expectation [3]. Pivoting to that surprise means moving into adjacent problem spaces—a detour that likely doesn’t solve, at least not very efficiently, the original problem. But pivoting to that surprise is the essence of research, of disruption to existing knowledge [4], and a signature of Bell Labs 1.0 research. In other words, corporations, with their problem-rich use environments, are powerful starting points for research, but, because of competitive market pressures, are inherently unable to support the pivoting to surprise. Our proposed Bell Labs 2.0 policy solution for this is a public-private partnership in which corporations host and cost-share, but funders with public-goods benefit in mind provide majority funding, for such research. Such a public-private partnership would be different from academic research in being embedded in real-world problem-rich environments, and from both traditional corporate “development” and the newer focused research organizations (FROs) in seeking surprise.

A second principle is “fund and execute research at the institutional, not individual researcher, level.” As is well known, virtually all goods and services in modern economies are produced by firms, not by individual contractors who manage their own individual transactions with a complex web of customers and suppliers. The production of goods and services is much more optimally orchestrated by empowered leaders with tacit contextual knowledge of complex resources, including that most complex of resources, people [5]. Most importantly for our purpose, new knowledge, particularly the surprising and disruptive new knowledge associated with research, is no different—its production is likewise more optimally orchestrated, as it was at Bell Labs 1.0, by empowered leaders. Our proposed Bell Labs 2.0 policy solution is thus a “firm” organizational model—one in which a research institution is block funded and whose leaders actively orchestrate and nurture researchers as employees, not contractors. Such a “firm” organizational model for research is opposite to the matrix-funded models that dominate the current research landscape—models in which researchers are employed by other institutions but contracted by funding organizations to perform research.

A third principle is “evolve research institutions by retrospective, competitive reselection.” As is well known, selection and reselection of leadership is an unsolved problem in organizational management. Great leaders are not guaranteed and, because in a “firm” organizational model leaders are of extreme importance, great research institutions are also not guaranteed. Moreover, there is much we do not know about the optimal organization of a research institute and how that optimal organization might vary with the advance of technologies for learning and collaboration [6]. Our proposed Bell Labs 2.0 policy solution is thus a network of Bell Labs 2.0’s: each hosted and cost-shared by a different corporate host; each majority funded by an organization that has the broader public good in mind; and, most importantly for this purpose, all competing with each other for re-selection. Such an internally competitive network of Bell Labs 2.0 is different from most existing research institutes which are captive to their research funder. For example, Janelia Research Campus is captive to the Howard Hughes Medical Institute and is virtually guaranteed funding without competition.

In summary, we believe the policy opportunity for these “Bell Labs 2.0” research constructs is enormous—the 2024 Chemistry Nobel Prize is a modern testament to research that can be done at corporate research labs, if properly configured. There are also policy challenges, of course, including: how can “inspire, but don’t constrain, research by particular use” be best operationalized; how might differences between the production of goods and services vs of knowledge alter trade-offs between firm vs contractor research models; and how can competition between research institutes in potentially vastly different knowledge domains be operationalized? We hope this talk catalyzes debate on these policy challenges, as well as on the policy opportunities.

[1] A. Arora, S. Belenzon, L. Sheer (2021). “Knowledge spillovers and corporate investment in scientific research.” American Economic Review 111, 871-898. [2] V. Narayanamurti, J.Y. Tsao (2024). How technoscientific knowledge advances: A Bell-Labs-inspired architecture. Research Policy 53, 104983. [3] A Bell Labs 1.0 example of the first kind of surprise was Claude Shannon’s information theory, a general way of thinking about efficient information communication. A Bell Labs 1.0 example of the second kind of surprise was Karl Jansky’s discovery of radio signals from the Milky Way, leading to his founding of the field of radio astronomy. [4] M. Park, E. Leahey, R.J. Funk (2023). “Papers and patents are becoming less disruptive over time.” Nature 613, 138-144. [5] O.E. Williamson (2010). “Transaction cost economics: The natural progression.” American Economic Review 100, 673-690. [6] M.A. Nielsen, K.J. Qiu (2022). “A Vision of Metascience: An Engine of Improvement for the Social Processes of Science.”

ABSTRACT. Generative artificial intelligence (GenAI) has demonstrated remarkable capabilities in generating text, images, and other forms of content by leveraging large language models. Its potential applications span diverse domains, including automated decision-making, content creation, and personalized services. The release of OpenAI’s ChatGPT-3 in 2022 marked a pivotal moment, accelerating the adoption of generative models in commercial settings and highlighting their potential to revolutionize industries.

Despite its immense potential, GenAI also introduces risks that stem from its technical architecture and deployment environment. GenAI models operate as “black-box” systems, where their high-dimensional and complex neural networks make it challenging to interpret, explain, or correct outputs. This opacity is particularly problematic when GenAI is integrated into decision-making processes in critical sectors like finance and healthcare, where errors or biases could have serious consequences. In addition, the growing interdependence between developers and deployers, often through third-party collaborations, compounds the risks of data breaches, regulatory non-compliance, and cybersecurity vulnerabilities. It is crucial to understand what risks are involved in developing and deploying GenAI in different sectors and what approaches can be taken to properly govern the risks in facilitating innovation.

This study examines the implementation of GenAI in the financial services and healthcare sectors, focusing on adoption patterns, emerging risks, and risk management approaches. Using a sectorial innovation system framework, we analyze how different industries integrate GenAI into their decision-making processes and manage associated risks. A comprehensive approach is taken to data collection by conducting 37 interviews between February and May 2024 with stakeholders across financial services, healthcare, technology firms, and regulatory bodies. A combination of stratified, purposive, and snowball sampling is adopted to ensure representative data collection across sectors.

The financial services and healthcare sectors provide a compelling basis for comparative analysis due to their distinct regulatory environments, priorities, and systemic risks. Financial institutions prioritize maintaining market stability, consumer trust, and regulatory compliance, while healthcare institutions focus on patient safety, clinical accuracy, and data privacy. These differences influence how each sector adopts and manages GenAI technologies.

Our findings reveal distinct patterns in GenAI ecosystem development across sectors. In financial services, GenAI adoption has been driven by the need to enhance operational efficiency and decision-making. Use cases such as automated credit underwriting, personalized investment recommendations, and trading algorithms illustrate the sector’s reliance on generative models. However, explainability and performance remain central concerns. Financial institutions must ensure that decisions made by GenAI systems can be substantiated to regulators, clients, and other stakeholders, particularly to avoid legal liabilities under consumer protection laws. Moreover, the performance of GenAI models in trading and investment decisions has been scrutinized, as these systems are trained on historical data and may struggle to adapt to dynamic market conditions.

In healthcare, GenAI holds promise for applications such as medical diagnosis, research, and clinical support. However, the sector’s emphasis on patient-centricity introduces unique challenges. Healthcare professionals prioritize accuracy and robustness in GenAI outputs, as even minor inaccuracies can have profound consequences for patient outcomes. In addition, IT infrastructure limitations, including data privacy concerns and compatibility issues with electronic health record systems, hinder the scalability of GenAI solutions. The sector’s reliance on sensitive patient data also necessitates rigorous cybersecurity measures, further complicating the deployment of generative models.

This study identified two broad categories of risks associated with GenAI adoption. Technical risks include concerns about model accuracy, explainability, and robustness. For example, in financial services, explainability is critical for ensuring that generative models’ outputs align with legal and ethical standards. In healthcare, accuracy and robustness are paramount, as clinicians rely on GenAI systems to support decisions that directly impact patient care.

Implementation risks, on the other hand, stem from regulatory uncertainty, data protection requirements, and infrastructure constraints. In financial services, regulatory uncertainty is a significant barrier to scaling GenAI solutions. Institutions face difficulties in navigating evolving regulatory landscapes, which could render their investments in GenAI non-compliant. Outsourcing and third-party collaborations, while offering cost-effective solutions, exacerbate these challenges by introducing additional layers of risk related to data security and model transparency. In healthcare, infrastructure limitations and data-sharing restrictions hinder the broader deployment of GenAI. Hospitals often lack the necessary IT infrastructure and machine learning expertise to develop and deploy generative models in-house, leading to a reliance on partnerships with academic institutions.

This study explored two primary approaches to managing the risks associated with GenAI. Rule-based governance relies on formal regulations and standards to enforce compliance and accountability. For instance, financial institutions are subject to consumer protection laws and regulations that govern data privacy and cybersecurity. However, rule-based approaches can be rigid and may not adequately address emerging risks introduced by new technologies. In healthcare, strict data protection regulations govern how patient data is managed, adding layers of complexity to the adoption of GenAI solutions.

In contrast, principle-based governance emphasizes ethical guidelines and frameworks to promote responsible AI development. Approaches such as the OECD AI Principles advocate for transparency, trustworthiness, and human-centered design. While principle-based governance allows for flexibility and adaptability, it often lacks accountability mechanisms, making it challenging to enforce compliance. A mixed approach, combining rule-based and principle-based mechanisms, is essential for managing the risks of GenAI while fostering innovation.

The findings of our study reveal both similarities and differences in how the financial services and healthcare sectors approach GenAI adoption and risk management. Both sectors face challenges related to technical uncertainties and regulatory compliance, but their priorities and strategies differ. Financial institutions emphasize explainability and market stability, while healthcare organizations prioritize accuracy, robustness, and patient safety. These differences underscore the importance of sector-specific policy interventions.

Policymakers can draw valuable insights from this study to enhance GenAI governance. For financial services, regulatory frameworks should be adapted to address the unique risks posed by generative models, such as biases in credit underwriting and algorithmic decision-making. For healthcare, investments in IT infrastructure and data-sharing mechanisms can support the scalability and robustness of GenAI solutions. In both sectors, fostering collaboration between regulators, industry stakeholders, and academic institutions can facilitate the development of targeted governance mechanisms that balance innovation with risk management.

ABSTRACT. The advent of the internet has ushered in a new era of economic development and social transformation. At a national level, the Internet, particularly the emergence of the social media, has proven to increase GDP with such effects varying between economies (Czernich et al., 2011; Gruber et al., 2014), and to shape formal and informal institutions (La Ferrara, 2016), etc. At the firm level, the internet helps to improve business performance and employee productivity (Mack & Faggian, 2013). At an individual/household level, broadband has been found to improve both individual and family income (Ariansyah, 2018; Atasoy, 2013). However, another strand of literature has spotted the potential welfare losses it might bring to society (Goldfarb & Tucker, 2019), including political polarization (Levy, 2021) and the exacerbated inequalities.

Existing research highlights the “skill bias" created by the broadband internet adoption, where it complements skilled workers while substituting for low-skilled workers, thereby widening the wage gap (Akerman et al., 2015). Nonetheless, technological changes do not necessarily always favor skilled workers. Acemoglu (2002) posits that the impact of technology on the labor market depends on its relative profitability in replacing different skill groups. If replacing skilled workers with new technology is more profitable for the market, new technologies will attempt to replace skilled workers as well. A limited number of studies have investigated the effects of expanding the internet on inequality of income due to potential endogeneity issues. For example, the relationship between the Internet and social inequality is often bi-directional. On the one hand, national development and socioeconomic status greatly influence internet access (DiMaggio et al., 2001). On the other hand, there are mixed findings on the role of the Internet in reducing inequality (Chen & Wellman, 2005). Furthermore, there are numerous factors that are affecting both internet development and income inequality (the existence of confounders).

Despite the growing body of literature on the internet's impacts from various perspectives, evidence regarding its influence on intra-city inequality remains limited. One close study in recent years shows that higher internet penetration rates increase consumption inequality in counties across China in 2010 - 2016, but such effects are reduced when the internet penetration rate exceeds a certain threshold. Additionally, higher education levels help to offset some of these effects (Zhang et al., 2020). Current discussions on the relationship between the internet expansion and labor income underscore the role of technological advancements in exacerbating wage differentials among workers (Acemoglu & Autor, 2011; Acemoglu & Restrepo, 2019; Kogan et al., 2023; Kogan et al., 2020). Consequently, it is likely that the internet also plays a significant role in shaping the evolving patterns of income inequality, as possibly, wealth inequality in China.

This study seeks to address the existing research gap by offering the first empirical evidence on the impact of enhanced internet infrastructure on income distribution inequality in a developing context. We use data from China Family Panel Studies (CFPS) and China City Statistical Yearbook amid the staggered rollout of the Broadband China Strategic Program from 2014 to 2016, which is a natural experiment characterized by robust government support for the upgrade of regional internet infrastructure. Our findings reveal that the enhancement of internet infrastructure led to a statistically significant reduction in intra-city income inequality, as measured by the Gini index, by 0.035 points (control mean: 0.459, standard deviation: 0.061). This reduction is primarily driven by the program's impact on cities in central China. Furthermore, we observe that improved internet infrastructure increases household net income and labor income. Notably, the reason for the simultaneous increase in household income and reduction in city-level income inequality is that the income gains are concentrated among low- and middle-income households: The increase in income for the high-income families is significantly and substantially smaller. Therefore, the income gap is effectively narrowed. We also find that the expansion of broadband internet access reduced income inequality through its impact on the labor market. Specifically, the improvement of internet infrastructure facilitates the development of the platform economy, which creates new opportunities for low-skilled workers in sectors such as logistics and (online) retail.

This study contributes to the existing literature in several ways. First, it addresses a critical gap in understanding the relationship between internet access and inequality by examining the impact of enhanced internet infrastructure on intra-city income inequality. This paper uses Broadband China Strategic Program as a natural experiment to overcome endogeneity issues and provides novel evidence of its potential to reduce income disparities within cities. It further highlights the importance of considering the distributional effects of internet access within a developing context. Second, by focusing on the Broadband China Strategic Program, this study offers valuable insights for policymakers seeking to leverage technological investments to promote inclusive growth. Our findings suggest that cities’ efforts, such as targeted investments in broadband infrastructure, and the extra support from upper level governments can effectively reduce income inequality. This evidence supports the idea that digital inclusion policies can play a crucial role in ensuring that the benefits of technological progress are shared more equitably across society. Last, our analysis sheds light on the potential mechanisms through which better internet infrastructure can influence income distribution. By facilitating the growth of the digital economy and creating new employment opportunities for low-skilled workers, the upgraded internet infrastructure can empower historically disadvantaged groups and contribute to a more inclusive labor market.

ABSTRACT. Sociotechnical transitions are often framed as consequences of technological advancement, with pervasive and broadly applicable technologies serving as the catalysts for these transformations. These technologies, due to their configuration, undergo an evolutionary process where they agglomerate capabilities, gradually acquiring diverse functionalities before becoming the established technology in a sociotechnical regime. Concurrently, various barriers – technical, social, cultural, or economic– impede realization of envisioned sociotechnical system transitions. This study posits that the sequence of barrier resolution and the process of technological capability agglomeration are interconnected, presenting an opportunity to address transition barriers proactively during technology development.

Using a narrative review of extant literature from diverse sociotechnical transition cases, this research explores the alignment between these sequences. The research yields three insights: i.) identification of the sequential nature of barriers and technological capabilities; ii.) delineation of barriers according to the types of capabilities (technical, social, economic, cultural) required to address them; and iii.) development of a pathway for proactively addressing these barriers through strategic acquisition of capabilities during technology development.

The significance of this study lies in its effort to bridge the gap between resolution of transition barriers and technological evolution. By providing stakeholders with a framework to proactively facilitate transitions through informed decision-making during the technology development process, we offer a nuanced perspective to proactively manage sociotechnical transition. This research contributes to the broader understanding of sociotechnical systems by highlighting the interplay between technological evolution and barrier mitigation. It provides valuable insights for policymakers, technologists, and other stakeholders involved in shaping future sociotechnical systems, enabling more holistic and proactive approaches to facilitate desired transitions.

ABSTRACT. The National Center for Science and Engineering Statistics (NCSES) within the U.S. National Science Foundation (NSF) is the principal source of analytical and statistical reports, data, and related publications that describe and provide insight into the nation’s science and engineering resources. NCSES collects and provides data on the science and engineering workforce; research and development (R&D); competitiveness in science, engineering, technology, and R&D; and the condition and progress of STEM education in the U.S. These data are collected as sample surveys or census.

The CHIPS and Science Act, PL 117-162, was signed into law in August 2022. Section 10375 establishes a National Secure Data Service Demonstration Project (NSDS-D) to "develop, refine, and test models to inform the full implementation of the Commission on Evidence-Based Policymaking recommendation for a governmentwide data linkage and access infrastructure for statistical activities conducted for statistical purposes, as defined in chapter 35 of title 44, United States Code." The Advisory Committee on Data for Evidence Building: Year 2 Report recommends several approaches for the NSDS that focus on supporting a high-quality user experience. Recommendation 3.4 states "The NSDS should employ data concierges to help users refine their research projects, discover relevant data, and acquire access to that data.”

Statistical agencies produce a large amount of data and analysis every year. These data can answer questions such as “How do most people die?” and “How many employed engineers with doctorates live in Arkansas?”. However, simply typing these questions into an internet-wide search engine or search bar on a statistical agency website will not always return the direct information requested. In addition, it may be difficult to know with an internet-wide search if the result is based on a statistic from federally published data or other non-federal data. The expansion of machine learning, including natural language processing and AI, may provide an updated approach to directly answering user queries related to public statistical information. The team is working to build a pilot tool to respond to users’ queries using publicly available data on statistical websites.

In the first part of the pilot, the team is building a Retrieval Augmented Generation (RAG) based system (process that controls the output of the Large Language Model before generating a response by leveraging database access) that is compatible with and builds on the open-source framework behind Google’s data commons. The chatbot will focus on three data sets that represent unique ways statistical agencies publish their data: (1) public use file, (2) data tables, and (3) analytical reports.

An important feature of this effort is how users interact with the system. The pilot tool, grounded in the Data Commons framework, offers: • Built-in visualizations, including timeline charts, scatter plots, and maps. • A natural-language query interface • Interactive and programmatic query capabilities. • Exploration tools for interacting with extensive datasets. • A Statistical Variable Explorer for exploring data variables. • A Knowledge Graph browser for examining data relationships. • Customization capabilities for organization-specific data implementations. • An enhanced Explore interface that uses large language models (LLMs) to map user queries to public datasets, generating relevant visualizations.

These features are designed to make public data more accessible, useful, and relevant for a broad range of users, including those in science, policy, journalism, and more.

This research presentation will introduce the pilot tool to the audience and detail salient technical details and lessons learned. Training of the model will be presented along with quality metrics such as precision and recall and benchmarking tests. In addition, we will present lessons learned about the size of input data tables, making statistical data “AI ready”, and other engineering issues encountered while building the pilot tool.

ABSTRACT. We have previously conducted a series of research profiles (basic and clinical) on COVID-19 (Porter et al., 2020) and a follow-on study of the pattern of Long COVID (LC) research (Porter et al., 2023). Both relied on the standard search protocols provided by the National Library of Medicine (NLM) to identify relevant PubMed papers. Recent research on LC prompts us to strive to enhance that search and apply a new analytical approach to help understand the challenging LC research domain and post-COVID-19 mechanisms.

The NLM LC definition is not fully accepted in the US; in contrast, the UK early settled on LC operational definitions. Fineberg et al. (2024; Figure 1) convey the messiness of specifying what to include as LC-related. They identify 11+ symptoms and 6+ physiologcal systems affected – including cardiovascular, cognitive, kidney, lung, dysautonomia, and autoimmune. We aim to bootstrap from the UK definitions to investigate global research set characteristics. The Long COVID research domain is challenging to specify and to bound. While MeSH terms lag significantly in Pubmed assignment, there are currently nearly 300 articles tagged with the MeSH term “Biomarker.” To illustrate, three biomarkers of note are elevated levels of cytokines such as Interleukin 6; biochemical markers like D-dimer indicating clotting abnormalities; and neurological markers such as elevated neurofilament light chain. We will devise Boolean search strings, with additional delimiters, to identify LC-indicative markers. We then intend to generate a learning model based on these, and query against the entire PubMed corpus. This should generate unique “non-LC” articles that do address LC-relevant research warranting inclusion.

In terms of methodology, this research advances “intelligent bibliometrics” by combining Machine Learning-enhanced analytical capabilities with established policy analytics. The paper then analyzes the expanded LC research data set, seeking to identify research opportunities (gaps) and policy opportunities. We aim to identify potential initiatives to connect topical research across research players -- countries, institutions, and researchers –that are not presently well-linked.

LC research both seeks biomarkers and draws on clinical solutions. However, those are hard to specify for a disease lacking in clearcut biomarkers, distinctive clinical features, or identifiable temporal patterns. Because LC symptoms are diverse and there is no lab test or simple diagnostic approach for isolating LC, the research space will most likely continue to operate inefficiently. Our approach should identify a more inclusive research literature.

Elements of interest in this study include use of Machine Learning trained routines to cluster text terms and phrases, so as to identify vital LC topics. Porter et al., 2023 (Table 7) show topical clusters from the 5,000 LC records compiled through 2022 in the PubMed and Web of Science databases. At the time, 13 topic clusters showed, including, neuropsychiatric, neurological sequelae, fatigue, cognitive deficits, and long-term effects. A simple search of PubMed, as of late 2024, finds over 30,000 papers. We plan to expand that search early in 2025 to more fully cover research publications through 2024. We will reach beyond the bounds of Boolean search on LC by applying the LC knowledge models to be built. The topic clusters will be different and richer than for the earlier results. Denoting changes over time should aid in projecting emerging topics.

Construction of knowledge models entails a several step process – clustering the core of the LC research space using established statistical methods; training a Machine Learning routine in our text analysis software, VantagePoint to form topical clusters; then engaging biomedical experts to check those results. This process aims to produce knowledge models for the major research domains within LC.

These knowledge models would be used to formulate queries within a specially encoded version of the full PubMed database to expand our LC dataspace. The results would be two-fold: expanded search results for knowledge spaces that are difficult to isolate with traditional techniques, plus a collection of topic models that can be used to understand and characterize the LC research space.

Fuller and insightful characterization of LC research is an aim of this study. Because LC symptoms are diverse and there is no definitive lab test or simple diagnostic approach for isolating LC, the research space will most likely continue to be characterized by “lots of loose ends.” Our approach should suggest productive pathways toward integrating this research to enable LC prevention and treatment. By providing a “map” of LC research and by tracking it over time, one could improve research connectedness to advance efforts to mitigate LC. Given that CDC (Centers for Disease Control and Prevention) researchers estimate that, as of August 2024, 17.9% of adults in the United States have experienced Long Covid, LC issues are not going away (Ducharme, 2024), progress is critical.

References

Ducharme, J. (2024), What’s the risk of getting Long COVID in 2024?, Time; [time.com/6999274/long-covid-risk-2024]

Fineberg, H.V., Brown, L., Worku, T., and Goldowitz, L. (Editors) (2024), Committee on Examining the Working Definition for Long COVID, A Long COVID Definition: A Chronic, Systemic Disease State with Profound Consequences, The National Academies Press.

Porter, A.L., Markley, M., and Newman, N.C. (2023), The Long COVID Research Literature, Frontiers in Research Metrics & Analytics.

Porter, A.L., Zhang, Y., Huang, Y., and Wu, M. (2020), Tracking and mining the COVID-19 research literature, Frontiers in Research Metrics and Analytics 5: 594060, 1-18.

ABSTRACT. Background As AI technologies continue to shape critical sectors, the risks associated with AI deployment—ranging from bias and discrimination to unintended system failures—are becoming more apparent (Acemoglu, 2021; Weidinger et al., 2022). Countries are taking different approaches to address AI potential harms through policies and regulations (Elgesem, 2023; Lupo, 2023; OECD.AI, 2021). One area where regulations and policies are being developed is incident reporting (Lupo, 2023), which is seen as a potential tool to prevent repeated AI failures (McGregor, 2021).

Prior work has analyzed incidents database to understand the ethical issues covered (McGregor, 2021; Nasim et al., 2022; Stanley & Dorton, 2023; Wei & Zhou, 2022) and how they could influence policy (Lupo, 2023). However, analysis is still in its infancy, and incident databases have failed to inform AI policies for different reasons such as underdeveloped reporting requirements, lack of standardized taxonomies, etc. (Lupo, 2023; McGregor, 2021; Stanley & Dorton, 2023). Without agreed categories and thresholds, regulators struggle to incorporate AI incident reporting into actual regulation, and companies lack incentives to report and use existing incident databases.

One particular challenge in using incident databases is processing a growing amount of textual information. This is a similar challenge that policymakers face when implementing participatory processes and analyzing civic participatory data. Policymakers are often overwhelmed by large amounts of textual data, reaching the point of cognitive overload (Chen & Aitamurto, 2019). In this domain, it has been explored how to use NLP tools to support policymakers in making sense of civic data (e.g., Arana-Catania et al., 2021; Guridi, Pertuze, et al., 2024; Romberg & Escher, 2022). However, many challenges remain unsolved, ranging from organizational complexities (Guridi, Cheyre, et al., 2024) to technical issues, including the lack of friendly interfaces (Romberg & Escher, 2023). Recent advancements in Large Language Models with chat-like interfaces, such as ChatGPT and NotebookLM, could address existing challenges and support policymakers in analyzing large amounts of text data.

We explore how policymakers could leverage existing LLMs to analyze incident databases and inform AI Policy. Using AI in the public sector could support policymakers’ analysis, but their use has significant risks (Dwivedi et al., 2023; Salah et al., 2023; Weidinger et al., 2022). Since most prior work using NLP analyses summarizing and organizing information, we focus on how LLMs can help policymakers large text databases to existing policy documents. Moreover, we use existing applications that do not require coding, which are more likely to be quickly adopted by policymakers.

Methods Our research question is: How can existing LLMs interphase support policymakers in relating AI incident databases to existing regulations to inform policymaking? To answer this, we conduct a case study using NotebookLM to analyze how incidents in the Artificial Intelligence Incidents Database (AIID) and the AIAAIC Repository relate to the European Union Artificial Intelligence Act (EU AI Act).

We analyze 818 incidents in the AIID, which have the following fields: title, description, date, alleged deployer, alleged developed, alleged harms, and a classification. We also analyze the AIAAIC repository, which has the following fields: title, type, date, country(ies), sector(s), deployer(s), developer(s), system name(s), technology(ies), purpose(s), media trigger(s), issue(s), and a URL to the full web description. For the AIID, we use the title and descriptions, while for the AIAAIC, we use the link to the description.

We use NotebookLM to analyze its ability to analyze incident report databases and relate them to existing policies. We chose this system because it can be trained with specific documents (e.g., PDFs, Google Docs, Websites, and Google Slides) and then be prompted to work only on the uploaded data. The system performed better than ChatGPT in our pilot testing for the specific tasks we aimed to do.

We used a series of prompts so that the system could relate incidents to specific sections in the AI Act and classify them on different criteria such as severity, likelihood of violation, etc. We manually revised if the resulting analysis corresponded with the data in the incidents’ description and the AI Act. We documented the prompts and results alongside notes about the system's performance.

Findings and Discussion Our findings demonstrate that out-of-the-shelf LLMs, such as NotebookLM, can assist in classifying AI incidents under the framework of the EU AI Act by identifying links between reported incidents and relevant provisions. For example, incidents involving deepfake scams and the use of AI for generating illegal content align clearly with Prohibited Practices under Article 5 of the EU AI Act. Similarly, using AI-generated voice clones without consent raises concerns related to Transparency Obligations outlined in Article 52, while anti-shoplifting systems potentially fall under High-Risk Systems due to their use of biometric identification and potential biases, as detailed in Annex III. These findings underscore the potential of LLMs to classify incidents based on risk levels and regulatory relevance, even with limited contextual information.

However, the analysis also reveals critical gaps in current incident reporting practices. More detail in incident descriptions limits their utility for direct application to regulatory frameworks, highlighting the need for standardized reporting criteria and comprehensive taxonomies. Despite these challenges, LLMs show promise in reducing the cognitive burden on policymakers by augmenting their ability to link incidents to regulatory provisions. This capability enhances the efficiency of regulatory analysis and ensures a more systematic approach to evaluating AI risks and their policy implications.

We faced several limitations that must be addressed before the widespread use of LLMs in this kind of analysis. For example, NotebookLM could not analyze tables or spreadsheets, forcing URLs or PDFs to be used, making the analysis less accurate. Moreover, more analysis is needed into how the LLMs use different levels of certainty in their response to inform people and how policymakers could systematically analyze that.

ABSTRACT. The transition to green technology is critical for addressing climate change, yet our understanding of how this transition occurs across different stages of production processes remains limited. While existing literature has extensively analyzed green technology adoption at the industry level (e.g., Acemoglu et al., 2012; Feng et al., 2022) or focused on specific technologies (e.g., Liu et al., 2023), these approaches mask important variations in transition patterns across different production process stages. Traditional industry-level analyses face three key limitations in understanding green technology transitions. First, they overlook how different stages within the same industry face varying barriers to innovation adoption - a critical oversight as technological challenges often vary significantly between basic materials development and final system integration. Second, they fail to identify potential technological synergies across similar processes in different industries, missing opportunities for cross-sector learning and technology transfer. Third, they provide limited guidance for designing targeted policies that address specific technological bottlenecks in the production process, potentially leading to inefficient resource allocation in technology development programs. This study introduces a novel process-level framework for analyzing green technology transitions, focusing particularly on the battery sector as an exemplar case that clearly demonstrates the value of this approach. By examining transitions at the process level - from upstream (raw materials and resources) through midstream (core component manufacturing) to downstream (system integration and application) - we can better understand where technological bottlenecks occur and how they might be addressed. Our analysis leverages advanced natural language processing techniques applied to 3.2 million patents, providing unprecedented granularity in mapping technological evolution across process stages. The importance of this research lies in its potential to inform more effective policy design. Understanding how green technology transitions vary across process stages can help policymakers develop more targeted interventions and optimize resource allocation. The battery sector, representing 28% of our sample and showing clear process-stage differentiation, provides valuable insights for developing stage-specific policy approaches. While our detailed analysis focuses on the battery sector, our methodological framework offers a new lens for examining technological transitions more broadly.

Our analysis framework consists of three distinct process stages: upstream processes involving raw materials and resource conversion, midstream processes focused on core component manufacturing, and downstream processes covering system integration and application. This three-stage framework provides systematic structure while maintaining sufficient granularity to identify specific technological bottlenecks. The framework is particularly well-suited for analyzing the battery sector, where we can clearly observe these stages from material development through cell manufacturing to system integration. We construct our dataset from USPTO patents granted between 1990 and 2023, identifying green technology patents through the intersection of three major classification systems: OECD's ENV-TECH, WIPO's Green Inventory, and EPO's Y02/Y04S tagging schema. This comprehensive approach results in a dataset of 3.2 million patents, with battery technology representing 28% (approximately 896,000 patents). For each patent, we collect and combine title, abstract, and citation information to create a rich text corpus for analysis. Our methodological approach leverages BERT for Patents, a specialized language model pre-trained on USPTO patent documents, which offers several advantages for our analysis. First, it better captures patent-specific terminology and technical context through its training on patent documents. Second, it understands hierarchical relationships between technical concepts, enabling more nuanced classification of process stages. Third, it can identify subtle technological relationships that might be missed by traditional keyword-based approaches. The analysis proceeds in three steps. First, we generate patent embeddings using BERT for Patents, capturing the semantic content of each patent's title and abstract. Second, we apply k-means clustering (k=30, determined through silhouette analysis) to these embeddings, as this method efficiently handles our large-scale dataset while providing clear, interpretable clusters. Finally, we map these clusters onto our three-stage process framework based on their technical characteristics and position in the value chain, validated through manual review of representative patents from each cluster. Initial clustering results reveal distinct patterns in the battery sector. We identify clear clusters for material development (clusters 29 and 8, comprising 1,366 patents), cell manufacturing (clusters 15, 22, and 4, totaling 3,104 patents), and system integration (cluster 17, containing 2,231 patents). The varying sizes and distinct characteristics of these clusters provide a robust foundation for analyzing process-stage differences in green technology innovation.

Our analysis of the battery sector reveals how different stages of the production process face unique technological challenges: Upstream processes (materials development, ~1,400 patents) show highly specialized innovation focused on fundamental research, particularly in metal materials and electrolyte development. These clusters reveal the deep technological changes required at the material level for green transition. Midstream manufacturing (electrode and cell manufacturing, ~3,100 patents) represents our largest cluster group, highlighting the critical challenge of scalability. These patents focus on translating basic materials into manufacturable components, addressing diverse technical challenges in electrode design, conductive structures, and assembly processes. Downstream integration (battery systems, ~2,200 patents) demonstrates a more application-oriented innovation pattern, focusing on practical implementation challenges including battery management systems and pack design. This stage reveals how new green technologies must be effectively integrated into existing systems. The varying cluster sizes and technical focuses across stages suggest fundamentally different innovation dynamics requiring distinct policy approaches. This evidence demonstrates why examining green transitions through a process-stage lens is crucial for effective policy design.

This research makes three principal contributions to understanding green technology transitions. First, methodologically, we demonstrate how combining a process-level framework with advanced NLP techniques can reveal the multi-layered nature of technological transitions. Second, empirically, we provide detailed evidence of how green innovations manifest differently across process stages, from fundamental material changes to system-level adaptations. Third, we show why considering process stages is crucial for designing targeted policies that can effectively address specific technological bottlenecks in the green transition. Our findings suggest that effective green transition policies must be tailored to address the distinct challenges at each process stage - from supporting fundamental research in materials development, to facilitating manufacturing scale-up, to enabling system integration. This nuanced understanding can help policymakers develop more targeted interventions across the production process.

ABSTRACT. Service robots are a key part of emerging technologies and drive innovation across industries. Their rapid growth and interdisciplinary nature make measuring and classifying them difficult. This study develops a framework to measure SR technologies. It focuses on two questions: (1) How can we create a reliable measurement framework for SR technologies? (2) How does combining human judgment with AI, such as ChatGPT, improve the identification and classification process? The study begins with an initial dataset of 174 firms identified using the International Federation of Robotics (IFR) categorization. ChatGPT was asked to generate lists of 50 SR-related firms in multiple iterations, resulting in 429 unique firms. These results were compared to a refined list of 48 firms, finalized through manual expert review. ChatGPT showed scalability but often included duplicates or irrelevant firms. In contrast, human review ensured contextual accuracy, particularly for firms without patents or explicit SR references. To link product descriptions to patents, four embedding models were tested: (1) Word BERT with individual word embeddings, (2) Word BERT with averaged word vectors, (3) Sentence BERT, and (4) Patent BERT. A reliability test on 20 random sample patents showed that Patent BERT produced the best results regarding precision and contextual relevance. Using Patent BERT, IFR product descriptions were matched to relevant patents through cosine similarity. Patents with scores above 0.65 were assigned to IFR categories. Patent data was sourced from PatentsView, which focused on 100,000 patents filed between 2021 and 2023. The resulting treatment group of SR patents was used to train a supervised model. The model classified patents into SR and non-SR categories. Stratified 5-fold cross-validation ensured balanced data and robust performance evaluation. This framework provides a reliable way to measure SR technologies. It addresses gaps in traditional methods by combining human expertise with AI tools like ChatGPT and Patent BERT. The approach ensures accurate and scalable measurement, making it applicable to other emerging technologies.

ABSTRACT. Since the enactment of the Bayh-Dole act, universities have increasingly become a source of innovation that leads to commercial products. In the U.S. alone, university researchers publish nearly one million articles annually. Only a small percentage of the technologies described by these articles are formally disclosed by universities through their technology transfer offices; even fewer are patented and/or licensed for commercial application. For example, while researchers at the University of Michigan publish well over 10,000 papers annually, less than 5% of these publications result in technical disclosures, with even fewer advancing to patent applications. In addition, many tools and other byproducts of research have commercial utility but are not reported to university technology transfer offices.

Our preliminary research, based on analysis of more than 80,000 research articles collected over an 18-month period, suggests that the number of university publications with high commercial potential significantly exceeds current disclosure rates. We developed and validated two complementary AI-based approaches to identify commercially promising research outputs: one focused on identifying disclosure-ready material, and another designed to replicate the evaluation criteria used by technology transfer professionals.

In validation experiments with licensing experts from the University of Michigan and the University of Chicago, our system demonstrated remarkable effectiveness. When presented with 100 previously unexamined research articles (50 classified by our model as commercially promising), the blinded expert reviewers identified 39% and 51% as promising candidates for commercialization, respectively. More significantly, when they were then provided with the AI-generated reasoning for each classification, the experts independently modified some of their responses, increasing their alignment with the AI and each other. As a result, the inter-rater reliability between experts improved dramatically from 66% to 82%, suggesting that our system not only identifies promising research but also provides useful context for human evaluators.

Our preliminary work already has important implications for science policy and university technology transfer: 1. There appears to be a substantial gap between potentially commercializable research and actual disclosures, suggesting that current manual screening processes may be missing valuable opportunities for economic impact.

2. AI-assisted screening could help TTOs more efficiently allocate their limited resources, allowing them to focus on the most promising opportunities while expanding their overall screening capacity.

3. More comprehensive identification of commercializable research could significantly increase universities' contributions to economic development, better fulfilling the intentions of the Bayh-Dole Act.

4. The high inter-rater reliability achieved with AI assistance suggests potential for more standardized evaluation processes across institutions.

This research suggests that by leveraging AI-assisted screening, universities could significantly increase their innovation pipeline while maintaining high-quality standards for commercial development. The next stage of our work focuses on incorporating actual licensing outcomes to refine our predictive capabilities, moving beyond disclosure worthiness (predicting expert judgments) to directly assess commercial viability (predicting executed licensing outcomes). The dramatic improvement in inter-rater reliability with AI assistance (66% to 82%) suggests that automated screening tools could not only increase the quantity of identified commercial opportunities but also improve the consistency and quality of evaluation processes. This could lead to more equitable access to commercialization opportunities across different research domains and institutions.

ABSTRACT. The structure of the scientific communication system, influenced by research policies prioritizing "high impact" venues, plays a crucial role in shaping research agendas worldwide. These policies, aimed at promoting the publication of "excellent" research in mainstream indexing systems, often align with the interests of international readers rather than addressing local scientific and societal needs (López-Piñeiro & Hicks, 2015). This misalignment disproportionately affects the Global South, exacerbating national inequalities and undermining the critical roles played by local journals, such as training junior researchers, disseminating region-specific findings, and translating global scientific knowledge for local communities (Chavarro et al., 2017). While recent studies have explored the effects of research policies on locally oriented research (e.g., Mongeon et al., 2022; Orduña-Malea et al., 2024), a consensus on defining and operationalizing "local research" remains elusive.

Building on our earlier work (Di Césare & Robinson-Garcia, 2024), where we conceptualized local research through six complementary approaches, this study examines how local research dimensions are represented within two leading global indexing systems: Web of Science and Scopus. We focus on three dimensions of local research:

1. Locally Informed Research: Measured by the geographic concentration of references. 2. Locally Situated Research: Measured through the use of toponyms in journal content. 3. Locally Relevant Research: Measured by the geographic concentration of citations and authors.

To address these dimensions, we analyze a dataset of over 17,000 journals extracted from Dimensions, enriched with bibliometric indicators capturing these dimensions. Using fuzzy matching techniques, we merge this dataset with metadata from the Master Journal Lists and Journal Citation Reports from Web of Science and the Scimago Journal Rank list from Scopus, evaluating the inclusion and representation of local journals across regions, particularly in the Global South.

Our study investigates two interrelated research questions:

- How is local research from the Global South represented in Web of Science and Scopus? - Which scientific areas and disciplines are most affected by potential underrepresentation?

By highlighting the underrepresentation of local journals, this study contributes to a better understanding on the effects of research evaluation policies on national research priorities and topic delineation of scholars, as well as to the differences of cost in terms of citation impact and visibility for publishing local research between researchers from the Global North and the Global South.

REFERENCES Chavarro, D., Tang, P., & Ràfols, I. (2017). Why researchers publish in non-mainstream journals: Training, knowledge bridging, and gap filling. Research Policy, 46(9), 1666–1680. https://doi.org/10.1016/j.respol.2017.08.002

Di Césare, V., & Robinson-Garcia, N. (2024). What is local research? Towards a multidimensional framework linking theory and methods. Zenodo. https://doi.org/10.5281/zenodo.14033473

López Piñeiro, C., & Hicks, D. (2015). Reception of Spanish sociology by domestic and foreign audiences differs and has consequences for evaluation. Research Evaluation, 24(1), 78–89. https://doi.org/10.1093/reseval/rvu030

Mongeon, P., Paul-Hus, A., Henkel, M., & Larivière, V. (2022, September 7). On the impact of geo-contextualized and local research in the global North and South. Zenodo. https://doi.org/10.5281/zenodo.6956978

Orduña-Malea, E., Miguel, S., González, C. M., Arias, R. R., & Ortiz-Jaureguizar, E. (2024). An analytical framework for the study of geographical places in the scientific literature. Revista Española de Documentación Científica, 47(3), Article 3. https://doi.org/10.3989/redc.2024.3.1571

Rafols, I., Ciarli, T., & Chavarro, D. (2019). Under-reporting research relevant to local needs in the South: Database biases in rice research. In R. Arvanitis & D. O’Brien (Eds.), The Transformation of Research in the South Policies and outcomes. Éditions des archives contemporaines. https://digital.csic.es/handle/10261/226953

Ràfols, I., Molas-Gallart, J., Chavarro, D. A., & Robinson-Garcia, N. (2016). On the Dominance of Quantitative Evaluation in ‘Peripheral’ Countries: Auditing Research with Technologies of Distance (SSRN Scholarly Paper No. ID 2818335). Social Science Research Network. https://papers.ssrn.com/abstract=2818335

ABSTRACT. Scientific peer review fundamentally shapes scholarly discourse and researchers' career trajectories, particularly for early career scholars.  Growing concerns around a "peer review crisis," including concerns around biases in scientific publishing and difficulties finding suitable reviewers, raise important questions about equity and efficiency in the peer review process. Leveraging large language models (LLMs) within an augmented resume audit design, we first experimentally test how explicit author identity cues influence LLM-generated editorial evaluations of identical scientific papers. Initial findings reveal some systematic biases: papers attributed to female authors, those affiliated with lower-prestige institutions, and—to a lesser extent—marginalized racial groups receive lower quality scores and a greater number of critical, yet less constructive, reviewer comments.  We further evaluate the quality of LLM-generated comments and assess feasibility in serving as a substitute or compliment to human-generated reviews.  LLM generated reviews are promising, albeit inconsistent. LLM biases may increase barriers to scientific productivity and career advancement, highlighting broader implications of LLM-based bias detection methodologies applicable to other areas of research.

ABSTRACT. Introduction

Large language models (LLMs) are increasingly being integrated into scientific research to automate and enhance various tasks, from code generation and data analysis to literature review, content creation, and even scientific writing (Lin., 2023). However, studies consistently show that LLMs carry biases in gender, race, and language, among others (Gallegos et al., 2024). Within the scientific context, known disparities in prestige, gender, and race (Wapman et al., 2022; Kozlowski et al., 2022) could contribute to these issues. However, the presence and impact of these biases in the scientific domain remain underexplored. In this study, we aim to detect and examine potential disparities in LLMs for and about science, focusing on gender, race, location, field of study, socioeconomic status, and academic prestige. We analyze their impact on the representation of scientific knowledge and researchers. We apply a well-known method for estimating biases, the Word Embeddings Association Test (WEAT), to measure issues in popular text models, BERT and SciBERT. We also analyze biases directly in an advanced LLM model, Anthropic’s Claude 3.5 Sonnet. Our results reveal significant biases across all dimensions examined, with a consistent tendency to favor dominant groups, such as people and institutions in the Global North.

Background Previous studies show that LLMs are increasingly integrated into scientific research, enhancing research productivity, diversity, and innovation by leveraging their three key attributes: intelligence, versatility, and collaboration (Lin., 2023). Although there is limited research on biases in LLMs within scientific contexts, prior studies have documented significant disparities in the research workflow related to prestige, gender, and race. Prestigious institutions dominate faculty placements, reinforcing hierarchical structures (Way et al., 2019). Gender disparities and career sacrifices disproportionately affect women, while racial disparities hinder the career progression of faculty of color (White et al., 2021).

Data and Methods Our analysis focuses on two main investigations: 1) whether disparities exist in underlying models used by LLMs, and 2) whether LLMs’ responses about science favor certain geographical or demographic groups. The first investigation uses the embedding level of LLMs. We apply the WEAT method, which calculates semantic distances between target and attribute word pairs using cosine similarity, to detect associations that reveal biases. This analysis is performed in BERT, which is general, and SciBERT, which is trained on scientific literature. In our experiments, we studied attribute and target words. Attribute words represent two sets of positive and negative words that we want to contrast with target words such as populations. We use positive and negative words, respectively, generated by GPT-4 to describe five key areas of research and academia including performance, productivity, teaching effectiveness, funding acquisition, and awards/ recognition. A set of targets was obtained from Caliskan et al. (2017), representing European vs. African-American names, male vs. female names, and science vs. art terms. We used GPT-4 to generate Global North vs. South, high vs low socioeconomic status, and prestigious vs. non-prestigious institutions. The second investigation involves LLM prompting tasks. The first task simulated email inquiries to professors about Ph.D. positions, varying the gender and location of the sender. Responses were rated on a Likert-like scale from 1 (Highly unlikely to reply) to 5 (Highly likely to reply). In another task, we simulate manuscript cover letter submissions, with the editor’s decision to desk reject or send for peer review.

Results The first investigation revealed consistent biases in both BERT and SciBERT through the WEAT task. SciBERT shows stronger biases in associating the overall dominant groups (eg. male and Global North names) with research and award successes than BERT (0.641 vs 0.569 and 0.824 vs 0.424 respectively on a scale from +2, indicating a strong bias toward dominant groups, to -2, toward underrepresented groups), and specifically, the Global North institutions with teaching success and research productivity (0.849 vs 0.670 and 1.199 vs 0.574 respectively), while BERT more strongly links Northern and European traits with academic success in teaching (1.232) and research productivity (1.226). The second investigation also showed consistent biases in both email inquiry and cover letter tasks. The email response task revealed disparities across gender and regional origin. Female candidates received a higher likelihood of response (M = 2.62, SEM = 0.12) compared to male candidates (M = 2.50, SEM = 0.11), while Global North candidates received significantly higher response rates (M = 2.90, SEM = 0.15) than Global South candidates (M = 2.22, SEM = 0.13). In the editorial review task, female authors from the Global North had the highest likelihood of their manuscript being reviewed (M = 3.34, SEM = 0.04), followed by female authors from the Global South (M = 3.07, SEM = 0.05).

Conclusion Our results emphasize the potential for LLMs to amplify existing biases, even as they become more integrated into scientific workflows (Wagner et al., 2022). Future research will expand prompt categories and test models like Llama 3 and ChatGPT, examining bias persistence across platforms. This will build on existing research on LLM biases, informing more equitable use of AI within scientific and educational contexts.

References Gallegos et al. (2024). Bias and fairness in large language models: A survey. Computational Linguistics, 1-79. Caliskan et al. (2017). Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334), 183-186. Clauset et al. (2015). Systematic inequality and hierarchy in faculty hiring networks. Science Advances, 1(1), e1400005. Wapman et al. (2022). Quantifying hierarchy and dynamics in US faculty hiring and retention. Nature, 610(7930), 120-127. Kozlowski et al. (2022). Intersectional inequalities in science. Proceedings of the National Academy of Sciences, 119(2), e2113067119. White et al. (2021). Race, gender, and scholarly impact: Disparities for women and faculty of color in clinical psychology. Journal of Clinical Psychology, 77(1), 78-89. Way et al. (2019). Productivity, prominence, and the effects of academic environment. Proceedings of the National Academy of Sciences, 116(22), 10729-10733. Lin, Z. (2023). Why and how to embrace AI such as ChatGPT in your academic life. Royal Society open science, 10(8), 230658.

ABSTRACT. Scientific breakthroughs are often associated with interdisciplinarity. Consequently, to increase the likelihood of breakthroughs, institutions that invest in science often invest in initiatives to incentivize interdisciplinarity. Here we investigate the effectiveness of one approach to encouraging interdisciplinarity: creating “problem centered” research unit with scientists and scholars from many disciplines with similar substantive interests in favor of traditional disciplinary academic units.  For this project we use archives of submissions for external funding by researchers at a southwestern university in the United States to study scholars and their proposal submissions over time. Collaborative research proposal writing provides opportunities for scholars to formulate and articulate their hopes for future research.  Investigators bring to proposal preparation their disciplinary background, research interests, skills, and priorities. They also bring their aspirations for future work. Participation in a collaborative proposal, then, indicates a scholar’s willingness to contribute to a collaborative endeavor, if not a willingness to evolve and grow from that experience. Viewed in this way, submission archives of proposals for funding contain valuable information for understanding research networks. For this project, we define as connected all investigators affiliated with a common proposal and all proposals that list any common researchers with specific focus on project teams with six or fewer investigators. We leverage data from multiple other sources such as human relations data, public web pages, and publisher databases, to populate the attribute file for scholars including a list of key words of self-identified interests.  We hypothesize that over time researchers who are members of an interdisciplinary problem focused unit will collaborate with more colleagues from other disciplines on proposal teams than their non-affiliated colleagues.  We also investigate if those embedded in this interdisciplinary environment are more central in the research enterprise than non-affiliated peers. 

ABSTRACT. Social capital incorporates the social process and status procured from the intricate structures of an individual's social network (Bordeau,1986). Social connections are a resource, and, as with many resource-based advantages, with whom we connect and when can make it easier to procure further advantages. This is true for various individuals (e.g., in business and other organisations), but also very true for academic researchers. Those who undertake an academic career will understand that 'social capital' is a significant precursor to initial and continued academic success (Abassi et al., 2011). However, success can mean many things, for example, acquiring research funds, scientific puzzle solving, and/or advancing a university’s mission (Ioro et al., 2017). At the very least it could be as simple as “becoming familiar with alternative ways of thinking and behaving” and being “at a higher risk of having good ideas” (Burt, 2004, p. 349).

The aim of our research is to examine how scientists’ collaborative preferences and strategies vary by groups, and how they are realized in actual collaborative behaviors. We focus specifically on academic faculty researchers who are embedded in institutions that provide both resources and rule structures important for career outcomes and advancement. How individual scientists prioritize the type of social capital that they seek in collaborative relationships may have an effect on the outcomes and visibility of their work and their careers overall.

This research advances an initial study, whereby collaboration habits and preferences of academics were investigated via a pilot questionnaire (Zuccala et al., 2024), using Nahapiet and Ghoshal’s (1998) theory of social capital based on three related dimensions: cognitive, relational, and structural. Shared narratives and shared codes and languages are features of the cognitive dimension. A shared understanding of norms, feelings of interpersonal trust, feelings of mutual obligation, and a shared form of identification are all features of the relational dimension. Structural social capital, which is the third dimension, reflects ties to an appropriate organization (e.g. a university department/research field), ties (bonds/bridges) to others in a network, and the overall network configuration.

We use Social Cognitive Career Theory (Lent & Brown, 2019) as a theoretical foundation for considering how scientists prioritize collaborative relationships. At various points in a faculty member’s career, different pressures, expectations, and opportunities may shape how an individual strategizes their collaborative work. For example, pre-tenured faculty may focus more on preferences related to relational and cognitive networks, whilst senior faculty will seek to broaden their network, and thus alter its structure with the inclusion of more international, cognitive diverse relationships. Given this, we hypothesize that:

H1: Collaboration preferences and resulting structures change as academics advance through their careers.

A known challenge is that access to social capital varies in important ways specific to collaboration and career advancement. Elitism in science has created insurmountable boundaries for some individuals, as evidenced by hiring and other professional networks (Feeney & Bernal, 2010; ). Demographic differences in professional social networks and related social capital point to exclusion from resource-rich networks. For example, collaborative and other professional networks of underrepresented groups (women, people of color, LGBTQ) may start off with similarities at a cognitive and relational level, but a majority member network will change structurally more rapidly with the inclusion of international nodes, and nodes representing more cognitive diversity (Siciliano et al., 2018). Informed by studies of varied access to social capital, we hypothesize the following:

H2: Collaboration preferences of faculty across demographic groups will be similar, but underrepresented faculty members will experience greater difficulties with structural manifestation of these preferences.

Data and Methods Drawing from bibliometric and/or survey data, studies of social capital in academic science have used both global and ego-based network approaches (Kadushin, 2012) that capture various forms of professional networks. Rather than a dyadic co-author and collaborative approach, this study uses an approach similar to studies of scientific cosmopolitanism (Wang et al, 2019) or collaboration (Fox, 2010) where collaborators are identified by type or location. Because we want to examine the relationship between collaborative preferences and mobilized social capital, we use both survey and bibliometric data.

Our survey is based on the approach developed by Zuccala et al. (2024) where social capital was measured in terms of preferences. In this regard, social capital is not recognized as an outcome, but a catalyst or lubricant for cooperative work. In addition, we will match Scopus bibliographic data to survey respondents in order to examine which types of social capital preferences are linked with strategic outcomes.

These data constitute a significant part of a larger national survey (State of the Professoriate), implemented in October 2024 with faculty working in doctoral-serving universities across the United States. It was distributed to over 30,000 tenured and tenure-track faculty from 9 STEM disciplines across more than U.S. research-extensive institutions. The survey will close in January 2025, and given survey response trends we anticipate a 15% response rate.

Our analyses will begin with a descriptive analysis of how research collaboration preferences vary by demographic characteristics and career stage. Using bibliographic Scopus data we will further analyse respondents’ co-authorship network, where links represent ties to a shared article, and nodes represent a co-author by affiliation and gender (relational social capital) and area of expertise (cognitive social capital).

Preliminary Findings Although the survey is still in the field, we reviewed preliminary findings (1,000 respondents) in order to examine what we expect to see in the data. A difference of means analysis shows some differences in collaborative preferences by gender, rank and citizenship. Given the low proportion of non-whites in the U.S. professoriate, We will examine differences by race once all data are collected. These preliminary data are important in establishing patterns of collaboration preferences. The next step is to match them (and the full dataset) with publication data (already acquired and matched to individuals) to complete our analysis.

References: omitted due to length restriction.

ABSTRACT. In 2018, we delivered a survey to scholarly parents to understand the gendered relationship between parenting and productivity (Derrick, et al., 2022). Our responses, received from 10,445 active scientists around the globe showed a strong parenting penalty based on levels of parenting engagement. That is, while gender differences were observed, they were largely driven by the higher engagement of women in parental duties. In 2020, the onset of the social restrictions wrought by the pandemic had a demonstrable immediate effect of production of science by younger women globally (Vincent-Lamarre et al., 2020). Universities responded by creating policies around COVID: allowing for extensions to evaluative periods (e.g., tenure-clock extensions in the United States) and introducing “COVID impact statements” into evaluations (Malisch, 2020). Surveys conducted nine-months after the pandemic suggested that there may be longer term production consequences, particularly for women with children (Gao et al., 2021) and many of these accommodations are still in place, indicating that the effects of the pandemic may be experienced long after social restrictions have ceased. This study seeks to examine the long-term consequences caused by COVID on academic productivity for parents considering not only the disruption of COVID by the potential recovery period to return to pre-COVID levels of productivity.

Using authors in Web of Science as a population, we constructed a sampling frame of one million disambiguated authors associated with a unique email address who had published at least one work as a corresponding author within the last five years and at least three papers in their publication history. The survey replicated the questions from the 2018 survey, but added additional questions focused explicitly on the experience of parents during the pandemic. This resulted in data that examined three time periods: pre-pandemic (2018); pandemic (2020-2021), and post-pandemic (2024). We received 7291 completed surveys at the time of this analysis. These responses were then matched back to productivity data within the Web of Science to compare parenting engagement with productivity.

We analyzed the average number of publications per year by the cohort from 2000 to 2022 and found a relatively stable increase, peaking, for both men and in 2019 and experiencing decline since 2022, suggesting a productivity loss that has not recovered (Figure 1). We also see that this sample contains several women with higher rates of production in the earlier years and a decline in the later years. This is a function of age: when we limit the population to only those who published their first paper after 1992, we find a much stronger rate of increased productivity and sharper rate of decline (Figure 2). The filter question for the survey was whether the respondent had children. Several individuals started the survey but then indicated that they did not have children. These data are mapped as “no kids” on Figures 1 and 2. As shown, there is no clear relation for the younger cohort between those with and without children, in terms of productivity.

We then relate these data to parenting engagement, to understand the effect of parenting on rates of production. We examine the difference between the number of papers published in 2020-2022 and those in 2017-2019. As shown in Figure 3, those individuals who had children prior to the pandemic experience a dramatic loss in productivity as compared to those who had their children in 2020 and beyond. We also see that men had a greater loss of productivity during the pandemic, compared to women.

Given these observed differences, we examine whether degrees of parenting engagement also varied during this period. As shown (Figure 5), women shifted strongly into lead parenting positions during the pandemic, sharply decreasing in satellite parenting and moderately declining in dual parenting. Men had a large increase in dual parenting, mirrored by a decline in satellite parenting. Overall, levels of parenting engagement increased for all parents during the pandemic and particularly for men. Post-pandemic, however, women shifted back to dual parenting, while men rebounded away from lead and dual to more satellite parenting. The return, however, was not back to 2018 levels for either—showing higher levels of post-pandemic parenting engagement.

Differences, however, varied dramatically by country. For example, while the United States (which comprises about a third of all respondents) mirrors the global level, variations are observed for other countries. Women in the UK, e.g., had much stronger lead roles during the pandemic and men maintained a slightly higher rate of dual parenting post-pandemic. Nordic countries observed an increase in women in lead positions during the pandemic, with women taking strong dual roles in the post-pandemic period.

Building on these initial results, we will explore the survey to gain a deeper understanding of the disruption of the pandemic on academic parents, whether this disruption was gendered, how it varies across country and discipline, and dynamics of the recovery period. The goal of this study is to inform policy makers to the extent of disruption and how policies can account not only for the immediate event, but the lingering consequences of this event on expected productivity, particularly across countries and by gender. This work also informs the broader body of literature on the post-pandemic labor market and sociological changes to work-life balance in the aftermath of COVID.

REFERENCES Gao, J., Ying, Y. Myers, K.R., Lakhani, K.R., & Wang, D. (2021). Potentially long-lasting effects of the pandemic on scientists. Nature Communications, 12, 6188. Malisch, J.L., et al (2020). In the wake of COVID-19, academia needs new solutions to ensure gender equity. PNAS, 117(27), 15378-15381. Vincent-Lamarre, Sugimoto, C.R., & Lariviere, V. (2020). The decline of women’s research production during the coronavirus pandemic. Derrick, G.E., Chen, P-Y., van Leeuwen, T. Lariviere, V, & Sugimoto, C.R. (2022). The relationship between parenting engagement and academic performance. Scientific Reports, 12, 22300.

ABSTRACT. OpenAI's demonstration of multimodal artificial intelligence (AI) services, featuring a blind person utilizing a mobile application for taxi hailing, illustrates the technology's capacity to simulate human-like assistance through contextual intelligence. Nevertheless, the development and deployment of such capabilities are contingent upon significant computational resources, including extensive datasets and advanced computing resources like GPUs, which may amplify existing socio-economic inequalities. This talk examines the intricate relationship between the rapid advancement of AI and social inclusion, with a particular focus on marginalized communities, such as older adults and persons with disabilities. I will provide the implications of concentrated investment and accelerated development within the AI sector, delineating both the potential for enhanced accessibility and the risk of heightened digital exclusion. Specifically, this talk will scrutinize recent legislative initiatives designed to promote inclusive AI, with a primary emphasis on the Digital Inclusion Act, passed by the Korean National Assembly in December 2024. Furthermore, I will explore the Act's legislative process, including its multi-stakeholder consultations and background, as well as the unfolding details of its subordinate legislation and implementation strategies. Finally, I will address the ongoing ethical discourse surrounding the integration of AI within social infrastructure, considering the long-term sustainability of inclusive AI initiatives and the imperative for continuous evaluation of their societal impact.

ABSTRACT. Artificial intelligence (AI) holds significant promise in enhancing the lives of neurodiverse individuals—those with neurological variations such as autism, ADHD, and dyslexia—by providing personalized support in communication, daily routines, and social participation. Recent studies demonstrate how AI-driven services can empower these users; for instance, large language model-based conversational agents offer autistic individuals a non-judgmental source of advice and coaching for everyday tasks, fostering greater independence (Cha et al. 2021, Choi et al. 2024). Similarly, AI-powered tools designed explicitly to accommodate neurodiverse users, such as gamified scheduling apps and adaptive social interaction platforms, have shown improvements in user engagement, task management, and self-efficacy (Kim et al., 2023, Kim et al., 2024). These findings underscore critical ethical considerations in AI development for marginalized communities. Embracing the social model of disability—which frames disability as arising from societal barriers rather than inherent deficits—requires designers to recognize cognitive variations as natural diversity to support rather than deficits to correct. For example, autistic individuals navigating online dating seek clarity on social conventions yet prefer preserving their unique characteristics, even if these diverge from neurotypical norms (Choi et al., 2023). However, mainstream matchmaking algorithms aligned with majority preferences inadvertently exclude such users, prompting implicit compromises in their authentic self-expression. Similarly, neurodivergent YouTube creators strategically portray their disability identities to educate the public but frequently find algorithms insufficiently supportive, diminishing their visibility and outreach (Choi et al., 2022). Considering that current AI systems often fail to adequately reflect neurodivergent users' self-defined priorities and inadvertently marginalize their authentic voices and perspectives, it is essential to explore new ethical design approaches—actively integrating neurodiverse individuals' unique needs, values, and identities, while consciously amplifying rather than suppressing their contributions within algorithmic decision-making processes. To foster responsible and inclusive AI innovation, this research advocates for a paradigm shift in the development of AI systems for neurodiverse populations. Co-design and participatory methodologies involving neurodivergent individuals as a key informant are essential for ensuring AI systems authentically represent their lived experiences, needs, and values. Transparency and fairness in algorithmic decision-making processes also must be critically examined to prevent systemic disadvantage or stigmatization of neurodiverse users. Moving forward, comprehensive frameworks and AI literacy initiatives should be established to: (1) recognize and respect neurodiversity as a cultural identity with unique experiences and perspectives, rather than a deficit or disability; (2) ensure AI systems preserve users’ distinct identities, values, and agency by not implicitly enforcing neurotypical standards; and (3) facilitate reciprocal understanding and bridge societal divides by enabling neurotypical populations to appreciate neurodiverse strengths and characteristics (Kim et al., 2024). This approach aligns with responsible AI principles—prioritizing equity, autonomy, and cultural sensitivity—and charts a path for AI innovation that uplifts marginalized communities. References Kim, S. I., Jang, S. Y., Kim, T., Kim, B., Jeong, D., Noh, T., ... & Kim, J. G. (2024). Promoting self-efficacy of individuals with autism in practicing social skills in the workplace using virtual reality and physiological sensors: Mixed methods study. JMIR Formative Research, 8, e52157. Choi, D., Lee, S., Kim, S. I., Lee, K., Yoo, H. J., Lee, S., & Hong, H. (2024, May). Unlock life with a Chat (GPT): Integrating conversational AI with large language models into everyday lives of autistic individuals. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (pp. 1-17). Kim, B., Jeong, D., Hong, H., & Han, K. (2024, May). Narrating routines through game dynamics: Impact of a gamified routine management app for autistic individuals. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (pp. 1-15). Kim, B., Kim, S. I., Park, S., Yoo, H. J., Hong, H., & Han, K. (2023, April). RoutineAid: externalizing key design elements to support daily routines of individuals with autism. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (pp. 1-18). Choi, D., Kim, S. I., Lee, S., Lim, H., Yoo, H. J., & Hong, H. (2023, April). Love on the Spectrum: Toward Inclusive Online Dating Experience of Autistic Individuals. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (pp. 1-15). Cha, I., Kim, S. I., Hong, H., Yoo, H., & Lim, Y. K. (2021, May). Exploring the use of a voice-based conversational agent to empower adolescents with autism spectrum disorder. In Proceedings of the 2021 CHI conference on human factors in computing systems (pp. 1-15).

ABSTRACT. AI offers innovative approaches to transform social service delivery and coordination, with significant potential to address persistent, long-standing inequities. Applications already in development include predictive risk modeling to identify families and children in need (Ahn et al., 2024; Chouldechova et al., 2018) and simulation modeling to optimize resource allocation across service systems (Mashiat et al., 2024). Despite these advances, there remains a critical gap in the development of comprehensive governance frameworks specifically designed for AI deployment in public sector social services.

Private sector AI governance frameworks prioritize organizational value creation, technical performance, and regulatory compliances (Batool et al., 2025; Birkstedt et al., 2023). These general frameworks organize governance around ensuring AI aligns with corporate strategies by emphasizing data management, algorithm functionality, and risk mitigation primarily from business and legal perspectives. In contrast, public sector frameworks shift focus toward procedural fairness, institutional accountability, and collaborative governance for public interest. For instance, Wirtz et al.'s (2020) model extends beyond private sector concerns by explicitly addressing societal impacts, ethical considerations, and multi-stakeholder policy development, acknowledging unique public sector responsibilities while balancing innovation with public protection.

When social services target inequality reduction among marginalized communities, existing public sector AI governance frameworks require further adaptation. These specialized contexts demand incorporation of structural inequities analysis, trauma-informed principles, and social work ethics into governance processes (Ahn et al., 2025). Effective governance for these applications must establish mechanisms that address power differentials and systemic inequalities beyond what generalized public sector models typically provide (Bricout & Goldkind, 2024). Without this critical extension of public sector governance for social service contexts, AI implementation risks reinforcing existing inequities rather than mitigating them. References Ahn, E., An, R., Jonson-Reid, M., & Palmer, L. (2024). Leveraging machine learning for effective child maltreatment prevention: A case study of home visiting service assessments. Child Abuse & Neglect, 151, 106706. https://doi.org/10.1016/j.chiabu.2024.106706 Ahn, E., Choi, M., Fowler, P., & Song, I. (2025). Artificial Intelligence (AI) Literacy for Social Work: Implications for Core Competencies. Journal of the Society for Social Work and Research, 735187. https://doi.org/10.1086/735187 Batool, A., Zowghi, D., & Bano, M. (2025). AI governance: A systematic literature review. AI and Ethics. https://doi.org/10.1007/s43681-024-00653-w Birkstedt, T., Minkkinen, M., Tandon, A., & Mäntymäki, M. (2023). AI governance: Themes, knowledge gaps and future agendas. Internet Research, 33(7), 133–167. https://doi.org/10.1108/INTR-01-2022-0042 Bricout, J., & Goldkind, L. (2024). Meeting the Grand Challenge to Harness Technology for Social Good: Social Workers Support Vulnerable Community Members’ Rights to Access Safe, Beneficent and Effective AI. American Academy of Social Work & Social Welfare. http://grandchallengesforsocialwork.org/ Chouldechova, A., Putnam-Hornstein, E., Benavides-Prado, D., Fialko, O., & Vaithianathan, R. (2018). A case study of algorithm-assisted decision making in child maltreatment hotline screening decisions. Proceedings of Machine Learning Research, 81, 1–15. Mashiat, T., DiChristofano, A., Fowler, P. J., & Das, S. (2024). Beyond Eviction Prediction: Leveraging Local Spatiotemporal Public Records to Inform Action. The 2024 ACM Conference on Fairness, Accountability, and Transparency, 1383–1394. https://doi.org/10.1145/3630106.3658978 Wirtz, B. W., Weyerer, J. C., & Sturm, B. J. (2020). The Dark Sides of Artificial Intelligence: An Integrated AI Governance Framework for Public Administration. International Journal of Public Administration, 43(9), 818–829. https://doi.org/10.1080/01900692.2020.1749851

ABSTRACT. Public-private partnerships play a crucial role in leveraging AI to address social challenges. While private companies possess advanced technological capabilities, public institutions provide the organizational infrastructure and direct engagement with marginalized populations, such as older adults and individuals with disabilities. This talk will present case studies of such partnerships in South Korea, highlighting collaborative efforts between SK Telecom and various non-profit and governmental organizations. SK Telecom, a global leader in AI and a key player in entrepreneurship with a strong focus on Environmental, Social, and Governance (ESG) principles, has pioneered several AI-driven initiatives to support vulnerable communities. One such initiative is AI Call Care, an AI-powered monitoring service designed to enhance the well-being of older adults. The successful implementation of this service has been made possible through long-term collaborations with local non-profit organizations and government agencies. Additionally, SK Telecom recently introduced Carevia, an AI-driven initiative that utilizes vision-based AI to analyze and manage challenging behaviors in individuals with developmental disabilities. Another notable innovation is Sullivan EYE, a multimodal generative AI technology jointly developed with Tuat, a tech startup, which recognizes images, text, and faces to provide verbal descriptions, assisting individuals with visual impairments. By examining these cases, this talk will explore the impact of AI-driven public-private partnerships in fostering social inclusion and improving the quality of life for marginalized populations.

ABSTRACT. Summary: This paper investigates the impact of collaborating with foreign-educated researchers on Colombian non-mobile researchers’ agendas. Data from a centralized academic CV system, scholarship lists, and a publication repository are combined to assess whether collaborating with foreign-educated drives non-mobile researchers to broaden or narrow the scope of his or her research agenda. By applying a Propensity Score Matching strategy and an Event Study analysis, we find that non-mobile researchers collaborating with foreign-educated researchers become less diversified than their colleagues. The effect is amplified as non-mobile researchers collaborate with a greater number of foreign-educated peers.

Extended abstract:

Background and rationale

Pursuing training abroad has been an effective strategy for researchers looking to advance their careers (Netz et al., 2020). Extant evidence at the individual level has shown that those who move abroad for training (i.e., foreign-educated researchers) have, on average, higher productivity (Baruffaldi et al., 2020; Baruffaldi & Landoni, 2012; Petersen, 2018), visibility (Jonkers & Cruz-Castro, 2013), and social capital (Jonkers & Tijssen, 2008; Scellato et al., 2015) than those who never experience mobility (i.e., non-mobile researchers). This is largely due to the fact that while aborad foreign-educated researchers increase their human and social capital (Bozeman & Corley, 2004; Lee & Bozeman, 2005). Also, according to knowledge recombination theory, mobility allows researchers to access distant sources of knowledge, which is often more beneficial in sparking creativity than is local knowledge (Fleming, 2001; Franzoni et al., 2018).

While much of the focus of the literature has been on the impact of mobility at the individual level, less attention has been given to the extended effects of mobility on non-mobile. Drawing on the argument that knowledge and social capital from foreign-educated researchers can be absorbed by other researchers through formal collaboration, or through being located at the same department, scholars have provided evidence that foreign-educated researchers positively influence nonmobile researchers’ productivity and networks (Borjas & Doran, 2015a; Franzoni et al., 2018; Fry & Furman, 2023; Fry, 2023; Müller et al., 2023; Yang et al., 2022). However, among the questions that remain unaddressed, we ask the following: since foreign-educated researchers access different sources of knowledge, does collaboration with foreign-educated researchers help non-mobile researchers expand their research agendas?

Methodology

In this paper, we assemble a unique dataset on Colombian researchers by combining three data sources: the national academic CV repository, Ph.D. scholarship lists, and OpenAlex, using observations from 1990 to 2021. We reconstruct researchers’ educational profiles and work trajectories, identifying foreign-educated and non-mobile researchers. To assess the impact of collaborating with a foreign-educated researcher on a non-mobile scientist’s research agenda, we consider two diversification indicators: (i) we calculate the share of papers published after having started the collaboration in which 75% of their (OpenAlex defined) topics are not present in their previous publications; (ii) we calculate a “pivot index” to capture the directional change of research. For the latter indicator, we modify the pivot index measure proposed by Hill et al. (forthcoming) by using topics assigned to publications in OpenAlex instead of referenced journals from a researcher’s publications. The methodological approach combines a matching and difference-in-differences estimator, relying on nearest-neighbor matching to find comparable treatment and control groups. We then rely on Callaway & Sant’Anna’s (2021) difference-in-differences estimator to perform an Event Study and obtain the Average Treatment Effect on the Treated (ATT). Results

We find that collaboration with foreign-educated researchers reduces the diversification of non-mobile researchers with respect to the other Colombian researchers. The share of publications of non-mobile researchers who co-author with foreign-educated researchers having 75% new topics is 18 percentage points lower than that of the control group. This indicates that non-mobile researchers collaborating with foreign-educated researchers are more inclined to publish on topics they are already familiar with. Regarding the pivot index, we observe an 11-percentage point decrease in the research direction trajectory of the non-mobile researchers. Overall, these results indicate that, for non-mobile researchers, specializing in fewer topics is the result of collaborating with foreign-educated researchers. Furthermore, by conducting an Event Study, we observe a consistent decrease in the share of publications with new topics and the direction of research trajectory (measured by the pivot index) over the years as non-mobile researchers increase their stock of foreign-trained coauthors.

Significance

We contribute to the literature on international scientific mobility in two ways. First, we analyze the extended effects of international mobility. Along with a few scholars (Fry & Furman, 2023; Fry, 2023; Müller et al., 2023), we depart from the mainstream literature, which emphasizes the impact of mobility on those who experience it, to demonstrate that an international mobility experience can also influence individuals who have never moved abroad. Second, we advance the literature by focusing on the extended effects on non-mobile researchers’ agendas. Previous studies have primarily focused on the effects of productivity and social capital. However, the impact of collaborating with foreign-educated researchers on the topics of non-mobile research remains barely explored. In our case, we show that by collaborating with foreign-educated, non-mobile researchers are stimulated to focus on topics they are familiar with. This seems likely to suggest that foreign-educated researchers play a role in helping locals to specialize, leveraging the absorptive capacity matured in the topics they know to refine their skills and include new methods or applications in their research.

ABSTRACT. This paper’s title paraphrases Robert Axelrod’s best seller “The Evolution of Cooperation”. Instead of focusing on how international cooperation emerges and develops, we aim at understanding why cooperation does not evolve. Departing from International Political Economy perspectives, we question to what extent the global economic position of countries in certain sectors explains the absence or insufficient regulatory measures adopted by states through the design and implementation of international agreements. We focus on the case of bioprospecting in Antarctica, a practice that has emerged as a new form of exploitation of the continent's resources between the late 1980s and early 1990s. Though Antarctica has historically been a source of economic resources and disputes, the Antarctic Treaty, signed in 1959, has not accounted for economic issues. Even if, since then, the trajectory of the Antarctic regime has been characterized by the evolution of cooperation in some economic-related areas, bioprospecting has not been part of Antarctic binding instruments. While the vast majority of the Antarctic Treaty's consultative members, at different degrees, have sustained a cooperative position by offering information and feeding debates about bioprospecting practices during Consultative Meetings, the fact that substantial decisions demand consensus could mean that a minority possesses a veto power and exercises it by remaining silent. In order to understand to what extent economic interests have influenced cooperative and non-cooperative positions adopted by consultative members of the Antarctic Treaty we have departed from two hypotheses: that governmental decisions which contribute to keeping the topic of bioprospecting out of the agenda of Antarctic regulations are influenced by the economic preferences of countries that are well positioned as global technological and commercial players; that cooperation can be promoted by players that aim at improving their global relative economic position. We adopt the geographic distribution of Antarctic simple patent families filed until 2023 as a proxy for global competitiveness. The number of bioprospecting-related working, information, and background papers (which we call 'position papers') submitted by governments to consultative meetings until 2023 is adopted as a proxy for cooperativeness. We recognize that the best indicator of economic competitiveness is trade. However, as there is no comprehensive database on Antarctic biotechnology-related traded products we have relied on a Boolean research in Lens.org that combined the list of OECD's biotechnology-related patent codes with Antarctic-related terms. Through a bidimensional analysis the mean of both variables was calculated in order to distribute the analyzed countries in four quadrants, considering the number of positioning papers and of patents above and below the mean. As the US has been identified as an outlier (with 1240 patents, followed by Japan, with 133 patents) it was removed from the dataset for mean calculation. Our hypothesis on the relationship between economic competitiveness and non cooperation seems to hold mainly in the case of Antarctic patent leaders that did not submit any position papers - cases of the US, Japan, China and South Korea. As hosts of powerful Antarctic innovation actors, those four silent countries might hinder movements that can lead to the regulation of their economic activities (as they prosper in laissez-faire) and to sharing strategic information that can enhance the know-how of potential competitors. We also found that other Antarctic patent leaders, such as the UK, Germany, Norway and Australia, have adopted lower cooperative positions. Higher cooperative positions were found in the cases of the Netherlands (15 position papers) and Belgium (11 position papers). Though both do not host many inventors of Antarctic biotechnology related patents, they are well positioned in the biotechnology-based economy in general and thus could benefit from cooperation initiatives that promote transparency and access to information related to Antarctic activities developed by current leaders. We suggest caution in treating both cases as similar to other countries that also appeared in the same quadrant, such as Argentina and Brazil, which, not being reference countries in biotechnology, may not have enough absorptive capacities to benefit from more access to information that can be promoted by cooperative arrangements. The case of France, the only country that appears in one of the quadrants, is also worth mentioning, as it can be considered an Antarctic patent leader, but, having issued six position papers, also seems to hold a more cooperative position towards bioprospection. As a country that maintains territorial claims in Antarctica - also the case of Argentina, Australia, Chile, New Zealand, Norway, and the UK - France can feed debates on Antarctic bioprospecting in an attempt to induce reciprocity and promote transparency from countries that have been bioprospecting in the Antarctic territory it claims. We conclude by stressing that both cooperation and non cooperation, as political economy phenomena, have distributional consequences. The non-regulation of bioprospecting activities in Antarctica benefits current leaders, as its regulation can benefit potential leaders, but not countries that do not have the indigenous networks needed to become leaders. For future research we recommend that the role of non-consultative members, such as Canada and Denmark, is accounted for, since they seem to be important Antarctic innovators. We recognize that decision-making in international negotiations is a complex process that involves not only economic interests, but also their interaction with other interest groups and institutions, including social movements, scientific organizations and the military. Therefore the next stage of our research is focused on case studies. Qualitative analysis will be crucial to allow us to hypothesize on the interaction between economic interests and other variables that can affect cooperation.

ABSTRACT. Scientific research funding plays a crucial role in fostering talent and advancing innovation across various disciplines. By providing financial support to researchers, these funds aim to promote academic productivity, career development, and collaborative networks. However, the general mechanisms and impacts of such funding on talent development require further exploration. This study examines the broad influence of scientific research funding on the growth and career trajectories of researchers, focusing on its effects on research output, professional advancement, and international collaboration. Research Questions: This study addresses the following questions: (1) How does scientific research funding influence researchers' academic productivity and career progression? (2) What are the short-term and long-term effects of funding across different disciplines and career stages? (3) How do funded researchers compare with non-funded peers regarding professional growth and academic influence? Research Methods: A quasi-experimental design is adopted, utilizing difference-in-differences (DiD) and propensity score matching (PSM) techniques. The analysis draws on data related to research publications, citation metrics, career promotions, and collaboration networks. Funded researchers are compared to a matched control group of non-funded researchers to isolate the causal effects of funding on their development. Research Design: This study uses recipients of the National Natural Science Foundation of China (NSFC) Excellent Young Scientists Fund (EYSF) between 2013 and 2018 as the primary sample. The primary growth metric is whether recipients subsequently obtain the Distinguished Young Scientists Fund (DYSF), and the time taken to transition from EYSF to DYSF is analyzed as an indicator of career advancement. Key academic outcomes are measured using metrics such as the total number of publications, citations, and H-index, providing a quantitative assessment of academic productivity and impact. Furthermore, a co-authorship network is constructed based on the publications of the sampled researchers to evaluate their collaborative relationships and the development of academic networks over time. Propensity score matching (PSM) is applied to select a control group of non-funded researchers with similar academic and professional backgrounds. A difference-in-differences (DiD) approach is used to compare the trajectories of funded and non-funded researchers, with multivariate regression models controlling for discipline, institutional type, and demographic factors. This design aims to provide a comprehensive understanding of the impact of scientific research funding on talent development, focusing on both individual growth and collaborative innovation. Expected Outcomes: The study is expected to reveal that scientific research funding significantly boosts academic productivity, enhances professional advancement, and fosters international collaboration. These effects may vary by discipline and career stage, highlighting the need for tailored funding strategies. The findings will offer evidence-based recommendations to optimize funding mechanisms for supporting talent cultivation and achieving broader research and innovation goals.

ABSTRACT. Critical to academic training is the relationship between students and supervisors. In this relationship, the learning process has been traditionally viewed as the supervisor unidirectionally transferring knowledge to the student (Delamont and Atkinson, 2001). However, unidirectionality is not always the case. Students may acquire specific knowledge faster than their supervisors (Stephan and Levin, 1992) and transfer that knowledge back to them. This is especially true in fast-changing technological landscapes (Fleming and Sorenson, 2003), in which supervisors often learn from their students the use of new technologies. One such technology is Artificial Intelligence (AI). The use of AI in science has grown fast in the last decade, transforming scientific research across research fields (Flanagan et al., 2023). Given its recent upsurge, students might have better and more up-to-date AI knowledge than their supervisors, becoming a source of AI knowledge. The goal of this study is to provide a comprehensive understanding of academic training by illustrating the bidirectionality of knowledge transfer between supervisors and students.

To empirically document the AI knowledge transfer process, we draw on a unique dataset covering the entire population of 51,826 French PhD students in STEM who graduated between 2010 and 2018. We focus on AI as the area of knowledge transferred for two main reasons. First, AI is a contemporary and growing technology (Flanagan et al., 2023), and advanced economies, including France, have made significant investments in research and education concerning the technology. Second, AI is a general-purpose technology (Cockburn et al., 2018) applicable in all fields of science, and thus, the learning of the technology has increasingly become critical.

In our empirical analyses, we test the existence of a supervisor-to-student knowledge transfer and the student-to-supervisor knowledge transfer using the following approach.

Supervisor-to-student knowledge transfer To analyze the supervisor-to-student knowledge transfer, we identify the student-supervisor pairs in which the supervisor has AI knowledge before the student’s enrollment (〖AI supervisor before〗_i=1). To address selection mechanisms that might intervene and bias our estimates, for each pair having a supervisor with AI knowledge, we find a similar pair having a supervisor with no AI knowledge before the student’s enrollment (〖AI supervisor before〗_i=0). To do so, we employ a Propensity Score Matching (PSM) approach based on the student-supervisor pairs’ observable characteristics. We compare the probability of writing an AI thesis (〖AI student〗_i=1) for students in student-supervisor pairs having an AI supervisor to the probability in a control sample of similar pairs having a non-AI supervisor. Our findings reveal a statistically significant difference in outcomes based on the supervisor's AI background. Specifically, pairs in which students are mentored by AI supervisors exhibit a 16% probability of showing a student writing an AI-related thesis, compared to the 4% probability in pairs in which non-AI supervisors mentor students. The statistically significant difference between the two probabilities, i.e., 12 percentage points, indicates an AI knowledge transfer from the supervisor to the student.

Student-to-supervisor knowledge transfer To analyze the student-to-supervisor knowledge transfer, we isolate the student-supervisor pairs in which supervisors have no previous AI knowledge before mentoring the students. We identify the student-supervisor pairs in which students develop AI knowledge during their thesis (〖AI student〗_i=1). Then, for each of these pairs, we find a similar pair characterized by a student with no knowledge in AI (〖AI student〗_i=0) using a PSM approach. Finally, we compare the probability of a supervisor initiating AI research within the three years after their student's defense year (〖AI supervisor after〗_i=1) for supervisors in student-supervisor pairs having an AI student to the probability in a control sample of similar pairs with a non-AI student. Our results reveal a statistically significant difference in outcomes. For pairs in which the student writes an AI thesis, the probability of the supervisor publishing in AI within the three-year window following the student's defense is 33%, compared to a lower probability of 14% for supervisors in pairs where the student does not write an AI thesis. The statistically significant difference between the two probabilities, i.e., 19 percentage points, indicates knowledge transfer from the student to the supervisor.

Our results are relevant for policymakers who want to steer the research agenda of a country toward specific topics. They show two main channels through which AI can permeate the research system of a country. One channel is to incentivize established researchers to develop AI knowledge, which they can transfer to their students. Another channel is the AI education of young generations of undergraduate students. Indeed, our results show that individuals acquiring AI knowledge before enrolling in a PhD program or during the PhD period can transmit this knowledge to their supervisors, steering supervisors’ research efforts toward AI.

ABSTRACT. Aims and Research Questions： This study empirically examined whether fellowship programs are effective in fostering talented young researchers. Specifically, it focuses on the Japan Society for the Promotion of Science (JSPS) Postdoctoral Fellowship (JSPS-PD), a typical program that supports young researchers before and after obtaining their PhDs. The main objective is to clarify the impact of receiving a fellowship on the research productivity of researchers and, in particular, to examine this from the perspective of the number of academic papers published and the number of citations received. Furthermore, this research also explores how factors such as gender, researcher mobility, speed of finding employment after being awarded a fellowship, and overseas experience affect research productivity. The specific research questions are as follows: (1) Does receiving a fellowship improve research productivity? (2) Is research productivity higher for those who find employment quickly after being awarded fellowships? (3) Does overseas experience improve research productivity? (4) Does research productivity increase for researchers who frequently change their affiliations? This study focuses on researchers in the humanities and social sciences, including economics, where the effects of fellowship programs have not been studied as extensively as in the natural sciences. However, the core mechanisms of fellowship support - financial security, resources, and time for intensive research - are relevant across various disciplines. Therefore, the findings and methods of this study are expected to provide universal insights that will help policymakers design effective fellowship programmes.

Method： A comprehensive microdataset was created to address these research questions, including data from 236 fellowship recipients and 454 non-recipients. The propensity score matching method was used to pair researchers with similar educational backgrounds and personal attributes, minimizing bias and enabling fair comparisons. The main variables analyzed included personal characteristics (e.g., gender, number of transfers between institutions, time to first job after fellowship adoption, overseas research experience, type of university/graduate school, etc.) and research performance (e.g., number of English papers published, number of citations, number of Japanese papers published, etc.). Academic paper data was obtained from databases such as “Web of Science” and “CiNii”, and information on doctoral degrees was obtained from databases such as “CiNii Thesis Information” and “ProQuest”, while details of career history were collected from publicly available CVs. A regression model was used for the analysis, and the impact of fellowships on subsequent research productivity was evaluated while controlling for potential confounding factors, such as past research achievements, academic background, and gender. This approach was essential for isolating the effects of fellowships and accurately evaluating their impact on the research results without distortion by external variables.

Results： The results clearly show that receiving a fellowship increases research productivity, particularly in terms of English publications and citations. On the other hand, the impact of fellowships on publications in Japanese is limited. Therefore, it is suggested that the JSPS-PD program is effective in fostering young researchers who can play an active role internationally. Furthermore, we found that researchers’ mobility is positively correlated with improvements in research productivity. Researchers who move between institutions tend to have a wider network and access a greater variety of resources, which is likely to contribute to improvements in research results. We also found that fellows who could find employment quickly after receiving their fellowship were more productive. This indicates the importance of stability during the early stages of a researcher's career. Interestingly, while it is generally thought that research experience abroad is beneficial, this analysis did not find a consistent positive correlation with increased productivity. The results of this survey call into question the common belief that international experience improves research performance and suggest that other factors, such as the quality of the research environment, may be more important. The survey also reveals a significant gender gap. Female researchers tend to publish fewer English-language papers than their male counterparts, possibly because career interruptions related to personal life events such as marriage and childbirth often coincide with the fellowship period. Despite the support provided by fellowships, female researchers may face unique challenges that hinder productivity. To address this issue, more targeted support mechanisms are needed to mitigate the effects of career interruptions and enable female researchers to pursue their academic goals without being disadvantaged.

Conclusion： In conclusion, this study provides robust evidence that fellowship programs such as the JSPS-PD effectively enhance the productivity of young researchers, particularly in an international context. Not only do recipients publish more papers but they also achieve higher citation counts, indicating that the program is successful in developing researchers from a global perspective. The importance of early career stability was also revealed, with the results showing that fellows who found employment quickly after being awarded a fellowship were more productive. However, the results showing that there were no consistent benefits from overseas experience suggest that simply encouraging people to gain international experience is not sufficient. Furthermore, the observed gender gap highlights the need for tailored support systems to address the unique challenges faced by female researchers. Although this study focuses on economics researchers, the insights are broadly applicable to a variety of fields. The main benefits of fellowships - financial security and the time to focus on research - are essential for young researchers across disciplines. These findings provide valuable guidance for policymakers in maintaining fellowship programs as an important tool for cultivating the next generation of academic talent and building effective support systems for early career researchers.

ABSTRACT. Research training programs are essential for developing the next generation of researchers, because they provide the knowledge, skills, and contacts necessary for young scientists to become successful independent investigators. Due to their importance, biomedical research training programs supported by the US National Institutes of Health (NIH) are frequently evaluated, but such evaluations are typically limited in scope for at least two reasons. First, they typically focus on a single program or training type and include fewer than 100 trainees. These evaluations can provide useful information for the specific program being evaluated, but their findings are difficult to generalize or compare with other programs. Second, most of these evaluations focus on subsequent NIH funding as the primary measure of trainee success. Although subsequent NIH funding can be an important step in becoming an independent biomedical investigator in the US, reliance on this measure undervalues other successful trainee outcomes.

In this presentation, we introduce a new method of obtaining trainee outcome data that addresses both limitations. We introduce a fully automated method of linking trainees to their Scopus author profiles, allowing us to obtain information on the current affiliations and research activities of thousands of NIH trainees from all NIH-supported training types. This also allows us to introduce a new measure of trainee success: whether the trainee is still actively publishing in the academic literature, and therefore still engaged in a research-related career. We then run this procedure on all trainees participating in all training programs supported by the Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD) from fiscal year 2000 through 2019. We then compare the traditional funding measure to the new active publishing measure for these trainees overall and by training type.

The proposed method has five steps. First, we obtain training award data and data on subsequent NIH applications and awards from our internal NIH data systems and perform some standardization and processing steps to ensure accurate matching. Second, we search the Scopus Author APIs for each trainee by name and affiliation at the time of training and retrieve up to three possible matches. Third, for trainees not matched in step 2, we search the APIs by name and subject area and retrieve up to five possible matches. Fourth, we use the APIs to obtain all Scopus Author profiles for all possible matches. Finally, we run a custom decision algorithm to select the most probable author profile for each trainee. Accuracy testing of the method against manual searches for 200 randomly selected trainees indicated that the method achieves 91% recall and 97% precision.

A total of 10,613 individuals participated in NICHD training programs in FY2000-2019. Approximately 20% of these trainees received one or more subsequent research project grants from the NIH, with that proportion remaining relatively constant across the years in our analysis. However, these rates varied substantially by training type, with our graduate and immediate postdoctoral training programs having lower funding rates (at around 15% and 30%) than our career development programs (at around 40% and 60%). Training programs awarded to institutions had lower rates than programs awarded directly to individuals across all trainee career stages. Individuals receiving student loan repayment awards had subsequent funding rates (at around 50%) comparable to those of individuals participating in career development awards, even though loan repayment awards do not include any formal training or mentoring.

Approximately 89% of NICHD trainees were successfully matched to a Scopus author profile. Match rates were higher for career development trainees than for graduate and postdoctoral trainees. Approximately 52% of all trainees in our analysis are actively publishing, defined here as authoring or co-authoring at least one peer-reviewed publication in 2023 or 2024. As with subsequent funding, active publication rates varied substantially by training type, with graduate and postdoctoral programs having rates of approximately 44% and 56%, respectively, and career development programs having rates around 80%. Loan repayment recipients had a rate around 72%. Approximately half of the trainees in our analysis who are actively publishing have not applied for subsequent research project funding from the NIH.

These findings have important implications for NIH training programs. First, we demonstrate that the proposed method can be successfully scaled to obtain outcome information for large numbers of trainees, making it a useful tool for subsequent analyses of other training programs at the NIH and beyond. Second, we find a substantial difference between the two outcome measures across all years and all training types in our analysis. Focusing on subsequent NIH funding as the primary measure of success therefore substantially underestimates the success of NIH training programs. Broader outcome measures are needed to adequately evaluate these programs. Third, we find that trainee success rates vary by career stage, with early career trainees having lower success rates than later career trainees. This suggests that factors external to the training are affecting these rates and that training program success should therefore be defined differently at each trainee career stage. Finally, these results suggest that approximately one-third of NICHD trainees are actively publishing but have not directly received any NIH research project funding. More research is needed to understand why these trainees have not received NIH funding and what might be done to support them in doing so.

ABSTRACT. This project is funded by the NSF (https://nsf.elsevierpure.com/en/projects/eager-pbi-paving-the-path-to-market-the-role-of-place-based-innov) and is the result of a public−private collaboration between researchers at Elsevier-Science-Metrix (David Campbell, PI) and the University of Iowa (Jiajie Xu, CoPI)

Background: Open innovation leverages collaboration to fuel the transition of research and development (R&D) outputs to commercial markets (1). Open innovation is most fully realized in place-based innovation (PBI) ecosystems in which diverse stakeholder types leverage geographic proximity to engage in joint R&D. Geography is a crucial factor in these ecosystems, as it allows for easier connections and the ability to leverage the unique attributes and potential of a specific location to enhance its competitiveness and resilience (2).

Leveraging PBI, the National Science Foundation's Regional Innovation Engines (NSF Engines) program aims to foster and accelerate economic growth to address the nation’s widening regional divide in technological advancements (3). However, there is a gap in research on, and comprehensive data sources for (4), understanding the key drivers and barriers to translating R&D outputs into new products, a key outcome sought through NSF Engine awards. Addressing this knowledge gap could support the adaptive management and success of each NSF Engine.

Goal: By linking large, regionalized data sources covering the entire innovation pipeline from R&D investments to their actualized products in the market, our study attempts—for the first time—an at-scale characterization of both the translation rate of R&D outputs into new products, and the time lag from R&D to commercialization. The study also tests a broad range of potential factors, including those that have been and continue to be subjected to PBI interventions, to address the following questions: • Which of these factors influence the translation of R&D outputs into new products and how? • Which of these factors influence the time from R&D to market and how?

These questions are addressed for each of the key technology and challenge areas (hereafter key technology areas, KTAs ) outlined in the CHIPS and Science Act (https://new.nsf.gov/chips#key-technology-areas-4f3).

Methods: Our study is based on a large-scale data set comprising all US R&D outputs produced since 1996, which consists of close to 6.6 million US publications (mainly peer-reviewed journal articles, conference papers and reviews extracted from Scopus) in the Natural and Applied Sciences, and roughly 1.2 million USPTO patents originating from the US in KTAs (extracted from PatentsView and LexisNexis).

By linking publications to patents, and then the patents to their actualized products in the market (extracted from IPRoduct), descriptive statistics are produced on the translation rate of R&D outputs into new products, as well as on the time lag from R&D to commercialization. The data covers all KTAs, overall and by area, for the US and for NSF-supported R&D outputs (close to 900K NSF-supported publications and 10K NSF-supported patents were identified). Data is disaggregated by US state and county for both groups. It is also sliced according to factors, with a focus on those of highest relevance to PBI, that may influence the commercialization of R&D outputs.

Such factors include, as an example, team size, sectoral and international collaboration, interdisciplinary knowledge and collaboration, gender diversity, entrepreneurship, availability and mobility of workforce, patent quality, local capabilities, availability of financial capital, as well as NSF support codified as PBI relevant or not. Data on these factors originates from a range of sources: Scopus, PatentsView, LexisNexis, IPRoduct, NSF Awards database, NSF-PAR, CHORUS, OpenAire, USAspending, VentureXpert, Business Formation Statistics (Census Bureau), and NamSor. Using this data, it is, for example, possible to compare the share of NSF public–private co-publications linked to patent-protected products to that of all other NSF publications.

The resulting data set is then used to model the influence of selected factors on the targeted PBI outcomes: number of patent-protected products linked to a publication or a patent, and the time lag to product release, either from the date of the related publication or patent. We decided to analyze factors relevant to PBI, rather than assessing PBI interventions, because NSF Engines are too recent and have limited data availability, and other PBI interventions are heterogenous and implemented on a smaller scale, restricting the ability to draw conclusions and/or limiting external validity of the results across the US.

While several of the factors considered in this study can be directly measured at the level of individual publications and patents, which can be regionalized based on the addresses of authors and inventors/assignees, some factors are only available on a regional basis. Two approaches are therefore used to address the study questions: • Multivariate multilevel regression models with some factors measured at the level of individual publications or patents combined with other factors measured for regions (US counties) within which the publications or patents are grouped; and • Multivariate panel regression models with all factors aggregated at the regional level (US counties).

Preliminary results: This study is still in its data collection phase but a complete draft of the descriptive analysis and statistical modeling will be ready for presentation at ATLC25. Our team has already published data on the distribution of patents across KTAs at the level of US states (https://ncses.nsf.gov/pubs/nsb20241/assets/data/swbinv-3.xlsx) and successfully regionalized all patent data at the level of US counties (see Figure INV-4, https://ncses.nsf.gov/pubs/nsb20224/invention-indicators-protecting-useful-ideas#). Additionally, preliminary descriptive results from a test sample for four US universities (University of Washington, Cornell University, Arizona State University and Howard University) show promising results. For all four universities combined, 3,667 patents were linked to products of which 2,108 had a product release date (representing 1.9% and 1.1% of all patents, respectively). Interestingly, the share of patents linked to products was 32% higher for NSF-supported patents (2.5% vs. 1.9%) providing early indication that federal support likely is instrumental in generating R&D outputs with higher market potential and/or in promoting the conversion of R&D outputs into economically valuable products.

References: (1) https://doi.org/10.1142/S1363919622400254 (2) https://www.thecgo.org/research/university-research-technological-commercialization-and-location/ (3) https://doi.org/10.3386/W31541 (4) Manuscript in preparation and shared with our team: https://www.gate.cnrs.fr/évènement/gaetan-de-rassenfosse-ecole-polytechnique-federale-de-lausanne-the-commercialization-of-dod-sbir-patents-a-counterfactual-analysis/

ABSTRACT. Place-based policies are widely employed globally to foster regional economic growth, yet the economic development implications of abruptly terminating these initiatives remain poorly understood. Leveraging a large-scale reform in China as a quasi-natural experiment, this paper examines how the sudden dismantling of Economic Development Zones (EDZs)—triggered by the central government's concerns over waste, fraud, and inefficiency—affected local economic competitiveness and welfare. Urban districts exposed to EDZ closures experienced sustained declines in export competitiveness and worker wages, significantly eroding the economic development benefits accrued during the zones' operational period. Supporting analyses suggest these adverse outcomes primarily arise from disrupted agglomeration economies and increased policy uncertainty resulting from the abrupt withdrawal of government support. However, districts with more innovation capacity, integration into regional supply chains, and trade specialization demonstrated greater economic resilience. These findings highlight the substantial risks associated with abrupt reversals in place-based policies, underscoring the critical importance in the design, placement, and management of zones while operational to sustain regional competitiveness and economic welfare even after explicit policy support is withdrawn.

ABSTRACT. Mechanisms of Firm Learning in the Global South: Political Economy of Quality Infrastructure

Rick Doner Gerry McDermott Dept. of Political Science Moore School of Business Emory University University of South Carolina

Abstract Submitted to the 2025 Atlanta Conference on Science and Innovation Policy (Topic / Area: Forms of governance (e.g., regulation, policy, scientific organization)

Through an analysis of quality infrastructure (QI), this paper addresses the institutional and political challenges of increasing productivity in Low- and Middle-Income Countries. QI refers not to physical infrastructure such as roads and bridges, or even IT-related supports, but rather to a particular form of governance -- the institutional ecosystem that helps firms identify and meet high-quality process and product standards. Concretely, QI institutions, include those, whether national, sectoral or multi-sectoral, or government, private or public-private -- responsible for metrology, standards, testing and quality certification (MSTQ). This list is sometimes supplemented to include accreditation, inspection, technical regulations, personnel certification (and thus training).

Appreciation of these functions has grown in response to concern with disappointing productivity in low- and middle-income country firms. Weak productivity is reflected in the “great divergence,” (unanticipated by traditional development theory), in which poor countries have generally failed to catch up as rich countries continue to pull ahead (Pritchett 1997); it is central to the “middle-income trap” problem (Doner and Schneider 2016); and it is relevant to an understanding of the “missing middle” (i.e. firms operating between a very few large, often multinational, firms and a very large group of small, unproductive firms) and related problem of inequality. The obvious solution to this problem is better adoption of new technologies (Comin and Ferrer 2013). An understanding of QI can help to address two gaps in the literature on technological catch-up by LMIC firms.

First is the nature of the technology relevant to the productivity of these firms. Recent scholarship has acknowledged that relevant technologies are typically not those at the world frontier and more those “new” to the firms, a point captured by the concept of firms’ “capabilities escalator” (Cirera and Maloney 2017; Lall 1992). We thus explore the ways in which QI facilitates incremental learning -- attention and access to diverse sources of applied, experiential knowledge in order to transform existing capabilities to meet basic international quality standards.

Second is the actual mechanisms through which firms might identify and make use of technologies “new” to them. Two traditional answers have proven unsatisfactory. Investments in R&D, while necessary, have been shown to be insufficient. And positive spillovers from MNCs and global value chains have been uneven: helpful to some countries but disappointing in many others where firms “lack the complementary capabilities that would allow them to accumulate knowledge” ( Cirera and Malone 2017: xxii). A key problem involves the ease of knowledge transfer even with relatively standardized systems. Much of the work on emerging markets believes that because MNCs tend to bring more mature, modularized production systems, with discrete packages of technologies and interfaces, then the relevant knowledge and practices are highly standardized and can be “bought off the shelf” or easily replicated by suppliers. (Gereffi, Humphrey & Sturgeon, 2005) However, the work on even apparently highly modularized manufacturing sectors within advanced countries such as automotives, aircraft, and energy equipment has shown increasingly how the diffusion of capabilities for lean production, TQM, and continuous process improvements depends on the tacit knowledge of translating the codified practices from one context to another (Camuffo & Cabigiosu, 2012; Sako, 2004). Moreover, recent research notes how even within MNCs, subsidiaries often have difficulties in such “plug and play” scenarios because they lack the original experiential knowledge of implementation to overcome the ambiguities of cause and effect when developed processes are transferred from one context to another. (Szulanski et al., 2004).

In response to these disappointing spillovers from FDI and GVCs, the mainstream multilateral organizations have increased attention to QI. By incorporating a focus on standards, testing and quality certification into National Innovation Systems and Triple Helix frameworks, the QI literature focus draws attention to the broader institutional ecologies in which local firms identify and adopt “new” technologies.

But if the existing QI literature constitutes important progress on the role of institutions in economic development, it is largely technocratic and supply-driven. As a result, existing analyses say little if anything about puzzles of variation: One involves differences in institutional design: Do all QI systems look the same? Another is the puzzle of performance variation: Why are some QI systems more effective than others? Here we move into more explicitly political concerns: In what ways do broader systems of governance affect performance? What incentives – pressures and opportunities – influence political leaders’ commitment to the challenge of building QI agencies? By addressing these puzzles, the paper aims to fill important theoretical and substantive gaps in the QI literature, while also helping policy makers to identify important opportunities and constraints.

Empirically, the paper draws on a set of comparative cases. Using a combination of most-similar and most-different designs, the paper two sectors (autoparts and wine) in one country (Argentina), and one sector (rubber) in two countries (Thailand and Malaysia).

ABSTRACT. In recent years, tax incentives, mainly provided to firms in the form of tax credits, have become increasingly popular among OECD countries. So that in 2023, out of 38 OECD countries, 33 countries offer R&D tax incentives for their firms, with 21 of them having R&D tax credits. R&D tax credits, reduces the risk of R&D expenses and help stimulate innovation by directly deducting from the firm’s tax liability. Over the past few years, more countries have used this tool to boost innovation within firms. The success of these programs depends on the effective design of their features. Although granting tax incentives has a financial burden (tax revenue loss) for governments, if properly designed and implemented, its fiscal impact is less than other forms of support. Determining eligible activities to receive tax credit, target groups, tax credit rate and permanent or temporary measures, are important decisions that have great effects on program's effectiveness and efficiency. Poor design of these features can lead to resource waste, unintended consequences like rent-seeking, and even market failure. Despite the significance of the issue, there is not enough theoretical and empirical guidance in this field. This research, with a comparative method, has examined the differences and similarities in design and implementation of R&D tax credit programs in developed and developing countries in order to provide suitable policy recommendation for developing countries. The comparative data analysis is conducted in two stages, descriptive and comparative, and success and failure factors were identified in order to extract key policy lessons for improving research and development tax credit systems. The design feature of R&D tax credit policy framework has been extracted from policy reports and global literature, leading to identification of seven key features, each contributing to enhancing effectiveness and efficiency of these programs in promoting research and innovative activities. The feature examined include definition of eligible operations for tax deduction, eligible R&D costs (current, labor, and capital expenditures), basis of credit calculation (volume-based, incremental or hybrid) tax credit), thresholds and caps on R&D expenditures and credit, target group (firm size, firm age, sector or industry, and region), permanent or temporary measures, and specific incentives for small and medium-sized firms. To provide a comprehensive and targeted analysis of R&D tax credits based on these key feature, comparative studies have been conducted, and components have been separately analyzed to clarify various aspects of design and implementation of this policy tool. The proposed policy framework for developing countries in designing R&D tax credit system can be more effective by focusing on four key components. First, focus on firms size as target group, noting the differences between large firms and SMEs. Results of using tax credit are different for two groups, and giving tax incentives can be more cost-effective considering the firm size. SMEs will benefit more from such incentives due to financial constraints and higher risks in R&D projects, and due to knowledge spillovers, they have greater economic effectiveness compared to large firms. Therefore, focusing on supporting SMEs can significantly contribute to sustainable economic growth and increased innovation. Second, there are three approaches to allocating R&D tax credits: volume-based, incremental, and hybrid. For SMEs, the volume-based approach is suitable due to its simplicity, greater generosity, and provides easier access. But for large firms that have more financial resources, the combination of incremental and hybrid approaches can be a suitable option, because this approach helps to better manage operational costs. Third, incentive generosity should be structured in a way that encourages investment in R&D but does not impose too much of a financial burden on government. Developed countries have maintained a balance between effectiveness and efficiency in their policies by setting threshold or cap for R&D expenditures and credits. This strategy also helps developing countries to optimally use public resources to stimulate R&D activities and thus create a sustainable path for economic growth and technological development. Fourth, In the proposed policy framework for developing countries, focusing on improving existing products and services and developing indigenous technical knowledge is more important. Unlike developed countries that are more concerned with innovation at the edge of knowledge, development of new technologies, and global competitiveness, developing countries can benefit from focusing on the operational and practical aspects of R&D to strengthen their indigenous infrastructure and existing capabilities. Thus, rather than pushing the boundaries of knowledge and facing the high risks associated with it, emphasizing on existing technologies improvement and production of new products and processes at the domestic industry level is presented as an effective strategy for achieving sustainable economic growth. In this policy framework, it is preferable to adjust definition of R&D activities to prioritize the improvement of existing technologies and the commercialization of enhanced technical knowledge at the national level. This approach not only controls R&D costs but also increases productivity and effectiveness at operational levels, contributing to sustainable economic development and enhancing regional competitiveness. Additionally, in proposed policy framework for developing countries, focusing on high-tech industries due to their high absorptive capacity and spillover capabilities, along with credit allocation based on firms’ performance and carryforward of unused credits to future years, enhances the effectiveness of tax incentives. This study is ongoing, and we have currently presented part of the findings. We continue to analyze the topics related to policy framework for developing countries, and we hope that by completing this research, we will obtain more comprehensive results that can contribute to improving R&D policies in these countries.

ABSTRACT. 1. Introduction In the globalised and competitive academic landscape, universities strive to enhance their research excellence, reputation, and visibility. International highly cited researchers (HCRs), are known to significantly contribute to a university’s scientific advancements, increase institutional prestige, and foster international collaborations. International HCRs not only elevate a university's research output but also play a pivotal role in securing funding, improving global rankings, and enriching the academic environment. Academic networks are known to shape research productivity and mobility. A university's position within international research networks can influence its visibility and accessibility to HCRs. Closeness centrality is a fundamental concept that measures how close a node is to all other nodes in a network. In the organisational knowledge network, the concept captures how efficiently a university can utilise knowledge within the network. High closeness centrality in the international HCRs network implies that a university can quickly reach and be reached by other prominent international research organisations and HCRs globally. Hence, making them more attractive destinations for international HCRs seeking to maximise their research impact and collaborative opportunities. Therefore, we propose: H1. A university’s closeness centrality in the international HCR network is positively correlated with its share of incoming international HCRs. Furthermore, the average shortest path reflects how quickly and efficiently information can flow through the network from one actor to another. A lower average shortest path indicates that the organisation is closely connected to others, facilitating rapid exchange of knowledge amongst organisations and researchers globally. It also reduces coordination costs and barriers to collaboration, as researchers can connect with fewer intermediaries. From the perspective of international HCRs, universities with a lower average shortest path length offer a more efficient platform for global collaboration and knowledge exchange. These organisations provide access to a broad network with minimal effort, making them attractive for HCRs aiming to maximise their impact and productivity. Thus, we hypothesise: H2. A university’s average shortest path in the international HCR network is negatively correlated with its share of incoming international HCRs. Similarly, the local clustering coefficient measures the likelihood that a node's neighbours are also connected to each other, forming tightly knit groups within the network. A high local clustering coefficient indicates strong interconnections among an organisation's collaborators, leading to cohesive subgroups. While this can foster trust and facilitate the sharing of tacit knowledge within the cluster, it may also result in redundancy of information and limit exposure to diverse perspectives. High clustering can create echo chambers where novel ideas are less likely to emerge due to the homogenisation of information. On the other hand, a lower local clustering coefficient indicates that a university's collaborators are less interconnected, exposing the university to a wider array of information from different parts of the network.For international HCRs, universities with lower local clustering coefficients may be more appealing because they offer access to diverse networks and reduce the risk of redundant collaborations. These universities provide opportunities to connect with unconnected or loosely connected collaborators, enhancing the potential for impactful research. Hence, we propose: H3. A university’s local clustering coefficient in the international HCR network is negatively correlated with its share of incoming international HCRs. 2. Method and Results We utilised panel data comprising 86 German universities from 2005 to 2020. Our dependent variable is the number of incoming international HCRs to a given university as a share of its total publishing authors for a given year. We use the author’s change of affiliations to determine their joining to a university. Our key explanatory variables are network measures derived from the universities' HCR collaboration networks. The nodes of our networks are the universities, and the edges are the number of publications among a university and other publishing organisations. Our network measures are closeness centrality, average local shortest paths, and local clustering. To establish whether the effect is significantly different when establishing connections domestically or internationally, each measure is calculated for a university’s international and domestic network. A university’s international network includes the connections amongst the focal domestic university and other international organisation, whilst the domestic network is between domestic organisations. We included controls for time, disciplines, regions, and university characteristics. We ran three separate regressions for each of the network measures using fixed effects. In the first model, the coefficient for external closeness centrality is negative and statistically significant (-0.0124***), while its interaction is positive and significant (0.0138***). This indicates a U-shaped relationship between external closeness centrality and the share of incoming international HCRs. In the second model, the external average shortest path length has a negative coefficient (-0.0013*) with a positive squared term (0.0002*), also suggesting a U-shaped relationship. Whilst, in the third model, the international local clustering coefficient shows a negative effect (-0.0030**) with a positive squared term (0.0037**), confirming the U-shaped relationship. We find domestic network measures are not significant throughout. 3. Discussion The results highlight the non-linearity between the international network measures and the share of incoming international HCRs, which contrasts with our initial hypotheses that posited linear relationships. One possible explanation is that, on the one hand, universities in peripheral positions offer unique research opportunities that attract international HCRs seeking niche areas. On the other hand, well-connected universities become hubs, providing extensive collaboration opportunities, which are attractive to international HCRs. Universities in the middle range might not offer the advantages of peripheral or well-connected organisations, making them less attractive to international HCRs. Hence, high connectivity may appeal to those seeking strong, tight-knit research communities with established collaboration patterns. Conversely, low connectivity may attract HCRs interested in bridging structural holes and accessing diverse, unconnected networks. These non-linear relationships challenge the simplistic assumption that better network positions attract more HCRs. Instead, they highlight the complexity of network dynamics and the varied preferences of HCRs. Moreover, the insignificant effects of domestic network measures indicate that domestic collaborations within Germany do not impact the attraction of international HCRs. This emphasises the importance of international over domestic networking in the global competition for top researchers.

ABSTRACT. Introduction

Across academic fields in the U.S., tenured and tenure-track faculty are approximately 25 times more likely to have a parent with a PhD than the typical American, and this tendency is even higher at elite universities [1]. The impact of this lack of socioeconomic diversity on scientific discoveries and innovations remains largely unknown. While gender, race, and ethnicity are known to influence the researcher's specialty (see [2]), socioeconomic background has not been studied due to the lack of data. This project studies and uncovers large-scale differences in scholarship across socioeconomic status (SES). To do so, we apply latent Dirichlet allocation (LDA), an unsupervised topic model, to publications of U.S. tenured and tenure-track faculty for which we have self-reported SES. Our results show that some research topics exhibit strong correlations with the SES of faculty.

Background and Research Question

Previous research shows that the composition of the scientific workforce shapes the pace and direction of discovery, and that increased diversity can produce more innovative science. Hence, representational disparities in the scientific workforce can represent a kind of “epistemic inefficiency” in the production of knowledge, as too few individuals focus on some questions and too many on others. Prior work has linked research topics with researcher gender and to a lesser extent to race and ethnicity. This systematic association suggests losses of innovation in identity-correlated topics like women’s health, ethnic studies, history, and heritable or infectious diseases. No comparable research exists for SES, largely due to the inability to infer SES algorithmically from bibliographic, name, or employment data, and because SES is complex and multidimensional. Instead, SES must be measured through self-reporting, which is difficult to perform at scale.

Studies have shown that childhood SES influences individuals’ educational attainment, including obtaining a PhD, academic socialization, and building ‘doctoral capital’. This literature illustrates the multiplicity of pathways by which SES can influence a person’s scholarly trajectory. Our study extends this to characterize the SES influence on scientific discovery and innovation. Understanding how researcher SES correlates with different scholarly outputs raises several questions, including: What scientific discoveries and innovations are we missing because the scientific workforce is dominated by higher SES researchers? How does SES’s impact on scholarship intersect with researcher gender, race, and ethnicity?

Data and Methods

Our study links three sources of data with state-of-the-art methods to address these questions: (1) a census of tenured and tenure-track faculty at U.S. PhD-granting universities from the Academic Analytics Research Center (AARC); (2) a complementary survey of faculty that collects self-reported SES; and (3) clean publication records from the open-source bibliometric database OpenAlex. Our census encompasses 239,949 tenure-track or tenured faculty across 111 academic fields at PhD-granting institutions in the U.S. who were active in 2011 to 2020. Here, we present results from combining self-reported SES variables for faculty in Computer Science (CS) and Biology, for whom we algorithmically linked with OpenAlex researcher IDs using their names and then affiliation institutions using string functions. Given the data available from the survey, we use the highest level of parental education reported by the respondent as a proxy for childhood SES, acknowledging the limitations of only using one of three dimensions commonly used to estimate SES. After cleaning the resulting 3-way linked dataset (faculty, SES, and publications) to correct incorrect linkages, we concatenated the publication abstract, title, and venue and applied Latent Dirichlet Allocation (LDA) topic modeling to represent each faculty member’s record of scholarship as a k-dimensional vector.

In the fields of Computer Science (CS) and Biology, the corpus has 40,978 and 162,190 publications respectively, for 363 computer scientists and 1,388 biologists. When fitting the LDA model on each field and stratifying by the researcher SES, we can observe if topic weights are correlated with SES. For each field, we estimated the SES gradient with a linear regression.

Results

Our analysis reveals that some research topics within a field exhibit strong correlations with faculty’s SES, while others do not. For instance, relative to lowest-SES researchers in their field, we find that the highest-SES individuals in CS are 1.47 times more likely to study “Network Systems” and in Biology are 1.23 times more likely to study “Bioinformatics.” For these two subtopics, we see a clear and positive correlation between topic proportion and SES level. Conversely, we find little correlation with SES in “Machine Learning and AI” for CS (odds: 0.97) or “Biochemistry” for Biology (odds: 0.98). The topics correlated with higher SES background –“Bioinformatics” and “Network Systems”– are technical and resource-intensive subfields. This pattern is consistent with an association between more technical fields and more privileged socioeconomic backgrounds. However, the “Machine Learning and AI” subfield in CS is also highly technical, but shows no correlation with SES background, suggesting a more complex relationship. Our further research agenda will seek to clarify this.

Discussion

These initial results provide evidence for our hypothesis that SES background can and does shape scholarship choices, even among highly trained researchers and within STEM fields like CS and Biology, which have fewer subfields that prioritize identity-focused scholarship. These also complement existing literature that links gender, race, and ethnicity to the researcher's specialty (see [2]). Characterizing how researchers’ SES shapes what scholarship they pursue or how they pursue it is especially important given the lack of socioeconomic diversity within the scientific workforce. In future work, we will compare these effects across our 111 sampled fields, shedding new light on how the composition of the scientific workforce shapes scientific discovery and innovation. Finally, understanding the concentration of researchers by SES across subfields and topics in academia will aid policies that tackle inequalities in academia, particularly as research has shown that the topic of choice can influence scholarship [3].

References

[1] Morgan, A. C. et al. Socioeconomic roots of academic faculty. Nat. Hum. Behav 6, 1625–1633 (2022).

[2] Kozlowski, D. et al. Intersectional inequalities in science. PNAS USA 119 (2022).

[3] Hoppe, T. A. et al. Topic choice contributes to the lower rate of NIH awards to African-American/black scientists. Science Advances 5 (2019).

ABSTRACT. Background Academic hiring outcomes [1] and scholarly productivity [2] are deeply rooted in the prestige of a professor’s PhD institution and faculty institution, but little is known about the paths tenure-track professors took before their PhDs. For instance, while the current professoriate is 20 times more likely to have a parent with a PhD than the U.S. population [3], other stages in the preparation and background of today’s faculty remain poorly understood, including their undergraduate education. Given how important prestigious undergraduate institutions are for granting students access to elite jobs in general [4, 5], this raises the question of whether and how a professor’s undergraduate pedigree impacts their future career.

Data & Methods To investigate the undergraduate experiences of faculty, we drew together three complementary datasets. We compiled a detailed dataset of the educational trajectories of all 16,006 tenure-track faculty active in 2011 and representing all career stages from 404 PhD-granting Computer Science, Business and History departments in the U.S., who attended 960 U.S. undergraduate institutions. We combined this person-level data with institutional metadata from the U.S. College Scorecard and with institutional prestige rankings from the U.S. News & World Report for both liberal arts colleges and universities. Together, these datasets allowed us to compare the undergraduate education of professors to that of the general U.S. undergraduate population, and the undergraduate education of elite professors to non-elite professors.

To evaluate institutional relationships in faculty trajectories, we constructed a career trajectory network, defining each node as an institutional transition—representing either the transition from undergraduate to PhD or the transition from PhD to professorship—with directional links.

Results Compared to U.S. undergraduate students at 4-year institutions, U.S.-educated professors are more likely to have attended private, research-intensive, and elite universities. Nearly 70% of U.S. undergraduates attend public universities, compared to only 47% of U.S.-educated professors, and roughly half of U.S. undergraduates attend research-intensive universities, compared to almost all (91%) U.S.-educated professors.

In fact, the trajectories of many faculty pass through only a handful of highly elite universities: one third of faculty in our sample attended one of 15 schools for their undergraduate degree, while only one in 100 U.S. undergraduate students attends one of those schools. Further still, nearly one out of every six professors attended one of just 5 schools for their undergraduate degree: Harvard, MIT, Yale, University of California-Berkeley, or Cornell, versus just four in 1000 U.S. undergraduates. Interestingly, not only are large proportions of U.S. faculty also trained outside the U.S., but these proportions vary markedly by field (51% in Computer Science versus just 17% in History).

Even compared to each other, elite professors are more likely to have attended elite undergraduate institutions than non-elite professors: 38% of professors working at a top-5 school received their undergraduate degree from a top-5 school. Similarly, 41% of professors who received their PhD from a top-5 school in their degree program also received their undergraduate degree from a top-5 school. While there is no single path to prestigious faculty positions, the odds become much higher with a more prestigious pedigree—the conditional probability of becoming a tenure-track professor at an institution ranked in the top ten percent (given that they are a professor) is 32% for a professor who attended an undergraduate institution ranked in the top ten percent, versus just 6% for a professor who attended an undergraduate institution ranked in the lowest ten percent.

Further, our network analysis shows that elite undergraduate institutions gave faculty a relative prestige boost in where they got a PhD and faculty job. Faculty who graduated from undergraduate institutions in the top twenty percent were more likely to “move up” the hierarchy—they attended PhD institutions and were hired as faculty by institutions more prestigious than their undergraduate institution. However, faculty who graduated from undergraduate institutions in the bottom twenty percent were more likely to “move down” the hierarchy—they attended PhD institutions and were hired as faculty by institutions less prestigious than their undergraduate institution.

In order to identify the steepest inequalities in trajectories in the network, we looked at the largest strongly connected component, the largest sub-network within the network where there is a path between every pair of nodes within the sub-network. While we found that only a small fraction of highly prestigious institutions (CS: 0.5%; Business: 1.5%; History: 5%) exchange graduates among each other, this small proportion of institutions decreases even further in randomized networks, suggesting that a specific preference among the most prestigious institutions exists for PhD students and faculty from highly prestigious undergraduate institutions.

Significance This work contributes to a growing literature focused on understanding barriers to scientific jobs, particularly in light of ongoing efforts to diversify academia, including by socioeconomic background. We find there is a distinct pipeline from prestigious undergraduate institutions, to prestigious graduate school programs, to prestigious faculty jobs. The cumulative advantage offered by an elite undergraduate degree, which is most often determined by social class and not ability [5, 6], reproduces a homogenous group of tenure-track professors trained at a few highly prestigious institutions. These findings set the stage for future investigations aimed at increasing access to who becomes a professor, and will be valuable to PhD admissions committees and faculty hiring committees who want to make more equitable admissions and hiring decisions.

References [1] Clauset, Arbesman & Larremore. “Systematic inequality and hierarchy in faculty hiring networks.” Science Advances, 1(1):e1400005, 2015. [2] Way et al. “Productivity, prominence, and the effects of academic environment.” PNAS, 2019. [3] Morgan et al. “Socioeconomic roots of academic faculty.” 2021. [4] Rivera. Pedigree: How Elite Students Get Elite Jobs. Princeton University Press, 2016. [5] Chetty et al. “Mobility report cards: The role of colleges in intergenerational mobility” (No. w23618). National Bureau of Economic Research, 2017. [6] Hoxby & Avery. “The missing one-offs: The hidden supply of high-achieving, low income students” (No. w18586). National Bureau of Economic Research, 2012.

ABSTRACT. Interdisciplinarity fosters innovation by bridging diverse methodologies and knowledge domains. The increasing complexity of global challenges has amplified the role of interdisciplinarity in advancing scientific innovation. There are also widespread efforts to promote interdisciplinarity in U.S. universities and PhD programs. While interdisciplinary PhDs are shown to have long-term gains in collaboration and productivity, they may face obstacles during early career transitions, such as reduced career opportunities. This study examines the impact of PhD interdisciplinarity on academic placement and long-term career trajectories, providing insights into how interdisciplinary training interacts with institutional hierarchies and hiring practices. To explore this issue, we used data sourced from the Academic Analytics Research Center (AARC), encompassing tenure-track faculty records from U.S. PhD-granting institutions between 2011 and 2020. PhD information, including department, dissertation titles, and advisor details, was supplemented from the ProQuest Dissertation database. To enhance coverage, machine learning models were employed to predict missing PhD subfields using dissertation titles. Publication records were further enriched by matching DOIs with Clarivate’s Web of Science, enabling analyses of interdisciplinarity, productivity, and citation patterns. The final dataset comprised 32,977 faculty members, 1,038,518 publications, and detailed advisor and collaborator networks. Each faculty member’s interdisciplinarity during their PhD (PhD interdisciplinarity) is the median interdisciplinarity score of their PhD-stage publications. We measured paper-level interdisciplinarity score using the Rao-Stirling index, which quantifies the diversity of references in academic publications. This index considers both the disciplinary variety of references and their similarities. References were categorized into 144 disciplines using the Web of Science classification. University rankings were based on a "production rank", which evaluates an institution's ability to place graduates into higher-ranked faculty positions within their subfields. Institutions were ranked by percentile, with the top 10% identified as those in the 90th percentile or higher within each subfield. By descriptive statistics, we found that faculty members with high PhD interdisciplinarity face disadvantages in securing faculty positions at top-ranked institutions. This trend is evident across most fields except Math & Computing. Faculty at top institutions are more likely to have lower interdisciplinarity, suggesting a preference for candidates with a narrower disciplinary focus. Analysis of hiring trends reveals that interdisciplinarity remains stable across lower-ranked institutions but declines sharply for placements at top-ranked institutions. Logistic regression models controlling for factors like PhD institution rank, graduation year, individual features, advisor features, and collaborator features confirm that higher interdisciplinarity correlates with lower odds of placement at top-ranked institutions, especially within fields like Physics & Engineering, Life & Earth Sciences, and Social Sciences. Quantile regression further demonstrates that this disadvantage is concentrated in higher institutional ranks, highlighting structural barriers for interdisciplinary scholars in the most competitive academic environments. Moreover, we found such a challenge for interdisciplinary PhDs happens if they seek job in their original fields. Top institutions tend to hire a higher proportion of cross-field faculty—those transitioning into new or distant fields—than lower-ranked institutions, reflecting efforts to diversify academic expertise. However, interdisciplinary PhDs face reduced chances of being hired for faculty roles within their original discipline at these top institutions. Multinomial logistic regression shows that as interdisciplinarity increases, the likelihood of securing same-field placements at top institutions decreases significantly. In contrast, cross-field placements at top institutions are less affected or may even increase with interdisciplinarity in fields like Life & Earth Sciences and Social Sciences. This suggests that top institutions value interdisciplinarity more for cross-field roles than for maintaining disciplinary continuity, creating barriers for interdisciplinary scholars who remain in their primary field. By measuring the research agenda alignment between our sample and each institution’s existing faculty, we see faculty members with low Ph.D. interdisciplinarity generally have smaller deviations from the research of existing faculty, and this alignment is more pronounced at higher-ranked institutions. In contrast, faculty members with high Ph.D. interdisciplinarity show consistently higher deviations. This misalignment shows the perceived lack of "academic fit" between highly interdisciplinary research and the traditional focus areas of top-ranked institutions, further reinforcing systemic barriers for these scholars. What are the potential consequences of such challenges for interdisciplinary PhDs? We first examined it from a gender perspective. We notice that women PhDs exhibit higher levels of interdisciplinarity than their male counterparts across most fields. However, this heightened interdisciplinarity correlates with reduced opportunities at top-ranked institutions, potentially exacerbating existing gender disparities. Propensity score matching indicates that accounting for interdisciplinarity mitigates these disparities, even creating a slight advantage for women in some fields. This indicates that the structural challenges interdisciplinary PhDs face in faculty hiring may contribute to the existing gender disparities in academia. Furthermore, using event study analysis, we identified that interdisciplinary faculty have higher publication rates over the decade following PhD graduation, particularly in Life & Earth, Biology & Health, and Social Sciences. Comparing the top-ranked and other institutions, we see this productivity advantage is even more pronounced at top-ranked institutions. Therefore, investing in and supporting interdisciplinary scholars at top institutions may yield long-term gains in research output, with the current hiring barriers potentially leading to a loss in knowledge production. Overall, our findings highlight a paradox: while interdisciplinarity is heralded as vital for addressing multifaceted societal challenges, systemic biases in faculty hiring constrain the career advancement of interdisciplinary scholars. Institutional hiring practices favor narrowly defined disciplinary expertise, potentially stifling innovation and perpetuating inequities. The gendered dimensions of interdisciplinarity further prove the need for targeted interventions to support diverse scholars. To fully leverage the potential of interdisciplinary research, academic institutions need to adapt hiring practices to value interdisciplinary expertise. Initiatives such as interdisciplinary clusters, dual appointments, and funding programs can create pathways for interdisciplinary scholars to thrive. By fostering a more inclusive and strategically aligned approach to interdisciplinarity, academic institutions can better address complex, cross-disciplinary challenges and ensure the equitable advancement of diverse scholars.

ABSTRACT. Introduction The capability of universities to place PhD graduates as professors at other universities is distributed unevenly. Network studies across various fields have confirmed this for the US university system, with Gini coefficients ranging from .62 to .76 [1-4], highlighting considerable institutional stratification. However, no comparable studies exist for German universities. We examine the exchange of PhD graduates between departments, using psychology and political science as examples for which comparative data are available. The German university system is an intriguing case: it has been argued that its institutional configuration, particularly the Humboldtian tradition, fosters a more egalitarian spirit than US higher education. For instance, professorial salaries in Germany are highly regulated under civil service laws, resulting in less salary variation than in the US [5]. Based on this, one would expect lower levels of both stratification and placement inequality. Conversely, the Humboldtian ideal of uniting research and teaching has been demystified as “invented tradition” [6]. Additionally, Germany’s excellence initiative (2005-2017) has recently intensified longstanding inequalities between large universities and smaller, less prestigious ones [7]. This suggests that the stratification and placement inequality may be closer to levels seen in the US. Theory A prominent network study by Burris [2] develops a prestige theory for the academic system, drawing on concepts from Max Weber (caste system) and Pierre Bourdieu (habitus). It posits that prestige is a structural property that organizes individuals and institutions hierarchically within a field. Prestigious departments possess specific characteristics, skills, and attitudes that afford access to desirable resources and tend to remain socially closed to protect their status. H1: Prestigious departments in German universities tend to recruit members of their professoriate predominantly from other prestigious departments, and thus among themselves. Less prestigious departments aspire to gain prestige. Since they cannot equip their graduate students with the same habitus as prestigious departments, they try to hire graduate students from prestigious departments to achieve a gain in prestige. H2: Less prestigious departments in German universities tend to recruit members of their professoriate predominantly from prestigious departments. Data and Methods This analysis uses data collected by a team led by Mark Lutter in 2019 and 2019, covering academic staff in psychology [8] and political science [9] at German universities. Data was collected from departmental and individual researchers’ websites and include information on their doctoral alma mater and current university. This resulted in datasets of 495 psychology professors across 74 departments and 279 political science professors across 73 departments. This study employs network analysis, treating departments as actors and professor placements as connection within the network. A connection exists between departments A and B if a doctoral student from A becomes a professor at B. Each department’s total professors trained is recorded as out-degree, and total professors hired as the in-degree. Departments are sorted by out-degree in descending order to determine their hierarchical position in the network. We test H1 and H2 by analyzing hiring movements within the placement network. For that purpose, we divide departments by their out-degree into quartiles. The first quartile contains 25% of departments with the lowest number of placements, the fourth quartile contains the 25% of departments with the highest number of placements. Results and relevance The Gini coefficient for out-degree is .61 for psychology and .66 for political science. This means that the capability of departments to train future professors is distributed similarly unequally in Germany as it is in the USA. Lorenz curves show that in psychology, 10% of departments have trained 53% of all professors. In political science, 10% of departments have trained 60% of professors. In psychology, 77% of professors within quartile 4 were trained there, and in political science, this proportion is 79%, supporting H1. For quartile 1, only 1% of psychology professors and none of the political science professors were trained there, lending support to H2. The training and hiring network of departments can be seen as an important mechanism for the emergence and dissemination of new ideas. The diversity of ideas largely depends on the diversity of doctoral training background [1, 3]. An excessive concentration on a few training institutions could thus pose an obstacle to the generation and spread of new ideas. The investigation of political science and psychology is a starting point for uncovering the hiring structures of the German university system. Clearly, two fields are not enough to make statements about the entire system. To address this, additional data sets are being developed to include sociology, economics, law, and geography. Data analysis is expected to be completed by spring 2025, enabling us to present broader and more in-depth findings on the entire system at the Atlanta 2025 conference. Reference list 1. Barnett GA, Danowski JA, Feeley TH, Stalker J. Measuring quality in communication doctoral education using network analysis of faculty-hiring patterns. Journal of Communication. 2010;60(2):388-411. 2. Burris V. The academic caste system: Prestige hierarchies in PhD exchange networks. American sociological review. 2004;69(2):239-64. 3. Katz DM, Gubler JR, Zelner J, Bommarito MJ. Reproduction of Hierarchy-A Social Network Analysis of the American Law Professoriate. J Legal Educ. 2011;61:76. 4. Wapman KH, Zhang S, Clauset A, Larremore DB. Quantifying hierarchy and dynamics in US faculty hiring and retention. Nature. 2022;610(7930):120-7. 5. Altbach PG. Paying the professoriate: A global comparison of compensation and contracts: Routledge; 2012. 6. Paletschek S. The invention of Humboldt and the impact of national socialism:: the German university idea in the first half of the twentieth century. 2001. 7. Heinze T, Habicht IM, Eberhardt P, Tunger D. Field size as a predictor of" excellence." The selection of subject fields in Germany's Excellence Initiative. bioRxiv. 2024:2024.03. 06.583816. 8. Lutter M, Habicht IM, Schröder M. Gender differences in the determinants of becoming a professor in Germany. An event history analysis of academic psychologists from 1980 to 2019. Research Policy. 2022;51(6):104506. 9. Habicht IM, Lutter M, Schröder M. How human capital, universities of excellence, third party funding, mobility and gender explain productivity in German political science. Scientometrics. 2021;126:9649-75.

ABSTRACT. Since Price presented his postulates on the shift to Big Science (1963), many have attempted at describing the social dynamics surrounding the development of scientific breakthroughs and the implementation of policies directed at fostering a productive research environment. Decades of work on interdisciplinary knowledge creation has advanced our understanding on knowledge-related dynamics across disciplines (Gibbons et al., 1994). Their challenges have been well documented in scientific collaboration studies (Bolduc, Knox & Ristroph, 2023; Klein, 2014). These address the dynamics of the inter-discipline, pointing to differences in language, research and team norms, and other disciplinary characteristics for working across boundaries (Leahey, Beckman & Stanko, 2017). These disciplinary distinctions are acknowledged, but not systematically examined.

From a theoretical perspective, Richard Whitley (2000) built a framework describing social and intellectual organization across fields based on two key concepts: mutual dependence and task uncertainty. While the former refers to the need for external validation, the latter focuses on labor distribution. Whitley suggests that different fields will have varying degrees of dependence and task uncertainty, proposing seven typologies of working structure. The dynamics of the research process are affected by myriad factors. Whether the research is more computational in nature, lab or field-research based, and reliant on specialized equipment or facilities are factors that drive the organization of scientific activities. Where a scientific team is based and where the scientists work introduce institutional, regulatory, and other factors that shape the nature of work.

Empirical studies on the social composition and dynamics of research activities have primarily focused on issues such as team size, power dynamics, collaboration or team composition and structure. This is due to the difficulty to examine the internal dynamics of team science at a large scale. These studies tend to rely on bibliographic data, applying bibliometric indicators based on author position, and combining it with individual traits such as gender or past trajectory. Science teams challenge current notions of credit in the scientific ecosystem (Walsh & Lee, 2015), as it becomes increasingly difficult to distribute it based solely on authorship position (Biagioli & Galison, 2013). In the last decade, the expansion of contribution statements across publication records allows for the first time to explore the relation between author order, time size and distribution of tasks, based on the self-reported statements of authors themselves. Benefiting from these novel data, we ask: how does intellectual leadership and research task distribution vary across disciplinary areas?

Data and Methods We analyze a set of almost 700,000 publications using the Contributor Role Taxonomy (CRediT) based on a set of Elsevier and PLOS journals to study the organization of the sciences across fields (González-Salmón et al., 2024). CRediT is increasingly being used by journals to catalog author contributions. Co-authors indicate their specific contribution out of 14 possible roles.

Building upon Whitley’s framework, we map the levels of task uncertainty and mutual dependence across and within fields. Considering that a paper reflects the outcome of a collective endeavor, we distinguish between scientific collaborations and highly stratified and organized team science with explicit structures. We apply agglomerative hierarchical clustering to publications by field and characterize the resulting clusters based on the team size and two measures of IDR, based on references and based on authors (Pinheiro et al., 2021).

Preliminary and Expected Findings So far we have analyzed the field of Medicine, but will extend our analysis to all fields prior to the conference. After applying bootstrapping and using different resolutions, our initial results suggest that there are four collaboration structures. These show different organizational structures related to differing modes of research. However, in terms of IDR, we find no meaningful differences across groups.

Cluster 1. All team members contribute to most tasks, with a moderate contribution on administration (e.g., funding). Papers are authored by a low number of researchers (avg. team size: 4.7). Cluster 2. Teams are characterized by a high distribution on technical tasks with a shared contribution on conceptual tasks. (avg. team size: 5.9). Cluster 3. Teams are larger with a higher distribution of labor for all tasks. (avg. team size: 7.5) Cluster 4. High distribution of tasks, although most members participate on the Review & Edit contribution (avg. team size: 8.2).

We will expand our analysis to all fields of research, exploring contribution co-occurrence by fields, and adding further variables such as seniority, gender, collaboration type and funding (based on funding acknowledgements).

Next steps The results of our research contribute to the understanding of collaboration and team dynamics in publishing activities. Because CRedit asks co-authors to indicate their role in the initiation, conduct, and communication of the research, it enables us to examine these activities across disciplines. These data greatly extend what has been possible to surmise from author order, but have limitations (Sauermann & Haeussler, 2017). There may be different approaches to how author contributions are reportedb However, given the size of our dataset, we can discern organizational structures across and within disciplines to understand the internal dynamics of team science.

Bibliography Bolduc, S., Knox, J., & Ristroph, E. B. (2022). Evaluating team dynamics in interdisciplinary science teams. Higher Education Evaluation and Development, 17(2), 70-81. https://doi.org/10.1108/HEED-10-2021-0069 Gibbons, M., Limoges, C., Nowotny, H., Schwartzman, S., Scott, P., & Trow, M. (1994). The New Production of Knowledge: The Dynamics of Science and Research in Contemporary Societies. SAGE. González-Salmón, E., Di Césare, V., Xiao, A., & Robinson-Garcia, N. (2024). Beyond authorship: Analyzing disciplinary patterns of contribution statements using the CRediT taxonomy. Zenodo. https://doi.org/10.5281/zenodo.14168888 Klein, J. T. (2005). Interdisciplinary Teamwork: The Dynamics of Collaboration and Integration. In Interdisciplinary Collaboration. Psychology Press, p. 23-50. Leahey, E., Beckman, C. M., & Stanko, T. L. (2017). Prominent but Less Productive: The Impact of Interdisciplinarity on Scientists’ Research*. Administrative Science Quarterly, 62(1), 105-139. https://doi.org/10.1177/0001839216665364 Price, D. J. de S. (1963). Little science, big science. Columbia University Press New York. http://www.garfield.library.upenn.edu/lilscibi.html Sauermann, H., & Haeussler, C. (2017). Authorship and contribution disclosures. Science Advances, 3(11), e1700404. https://doi.org/10.1126/sciadv.1700404 Whitley, R. (2000). The intellectual and social organization of the sciences (2.a ed.). Oxford University Press.

ABSTRACT. This study focuses on the use and construction of composite indicators (CIs), which are statistical tools used to evaluate performance of innovation systems by combining multiple variables into a single, comparative metric, Cherchye et al. (2007) and Greco et al. (2019). Most prominent examples of CIs to evaluate innovation systems are the European Innovation Scoreboard of the European Commission and the Global Innovation Index by the World Intellectual Property Organization, Edquist et al. (2018). CIs are instrumental for policymakers, academics, and stakeholders because they condense complex datasets into comprehensible and actionable scores, helping to assess and rank performance across countries, regions, or institutions.

Purpose and Methodology of CIs to Evaluate Innovation Performance

The primary objective of using CIs is to aggregate information from several variables into a singular score or index, representing the utility or effectiveness of an alternative. Decision-makers (DMs) use this single measure to compare the performance of various alternatives. Organizations globally have leveraged CIs to rank observations based, enabling comparative assessments across many areas, Frudenberg (2003) and Bandura (2006, 2008, 2011).

However, the methodology involved in constructing CIs is complex and introduces challenges. The design process must address multiple methodological questions, including variable selection, normalization, weight assignment, and the choice of an aggregation function. The Organization for Economic Co-operation and Development (OECD), for instance, has established a "toolbox for constructors" with guidelines to avoid common pitfalls like non-transparent weighting and improper data aggregation, OECD (2008).

Challenges in CI Construction

A major issue in CI construction is the selection of an aggregation function and the weights assigned to each variable. The aggregation function determines how the various variables interact within the composite index. Many institutions opt for a linear arithmetic mean function, assuming that all variables can substitute each other perfectly, but this assumption has been criticized. For example, in the Global Innovation Index used for raking innovation systems, increasing the R&D expenditure in the business sector (BERD) can always substitute a reduction in the number of Employed ICT specialists (EICTS) in the same amount. This linear approach implies that a high score in one variable can perfectly offset a low score in another, which may not always reflect reality, Barbero et al. (2021).

Alternatively, other forms of aggregation functions that incorporate varying levels of substitutability, such as geometric mean or fixed proportions, can be applied. For instance, the fixed proportions (Leontief function) assumes that each variable contributes a necessary amount to the total utility, making it suitable when all variables are essential to the evaluation. The choice of an aggregation function is thus subjective and represents the preferences of the DMs. Furthermore, establishing weights for the variables is critical, as it defines each variable's importance in contributing to the overall score. The relative weights illustrate the trade-offs or importance between variables, which heavily influences the CI outcomes.

Subjectivity in Weights for Innovation Indicators and the DEA-BoD Approach

The subjective nature of selecting weights introduces potential biases, as DMs' preferences may not align with the realities or priorities of each alternative being evaluated. To address this, an alternative method known as Data Envelopment Analysis with Benefit-of-the-Doubt (DEA-BoD) is often employed, Cherchye et al. (2004). DEA-BoD is an unsupervised method that removes the necessity for predetermined weights, instead allowing each alternative to be evaluated on its most favorable terms. This approach involves using mathematical programming to maximize the performance score for each alternative by adjusting weights in a way that reflects each alternative’s strengths.

The DEA-BoD approach evaluates each alternative based on a constructed utility frontier, meaning that each alternative is compared relative to the best-performing peers. This approach creates a Pareto-efficient evaluation, which identifies alternatives that cannot be improved without diminishing the performance of at least one variable. DEA-BoD thus avoids biases that could result from a predetermined, universal weighting system.

Limitations of DEA-BoD

Despite its strengths, DEA-BoD has certain limitations. Its flexibility allows for zero weights to be assigned to some variables, which can lead to misrepresentations. For instance, an alternative may score highly by maximizing one variable at the expense of others, which may not accurately reflect a compressive innovation system. As an example, imagine the evaluation of a national innovation system where R&D expenditure in the business sector (BERD) is assigned a zero value because the alternative has the largest value in the number of Employed ICT specialists (EICTS).

Hybrid Methodology: Combining CI and DEA-BoD

To address the drawbacks of both CI and DEA-BoD, the study introduces a hybrid methodology that incorporates elements of both. This approach uses a standard utility function and a common set of weights from CI methodology, while also leveraging DEA-BoD to enable flexibility in weighting for each alternative. This combination provides a dual assessment: one subjective, based on a chosen utility function and standardized weights, and the other objective, through DEA-BoD’s Pareto-efficient analysis.

In this framework, two types of effectiveness measures are generated: overall utility effectiveness and technical utility effectiveness. Overall utility effectiveness reflects subjective preferences, while technical utility effectiveness is purely quantity-based, comparing each alternative against the optimal, or Pareto-efficient, benchmark. This distinction enables the identification of an “allocative” utility effectiveness component, which quantifies the difference in performance attributable to the chosen weights and preferences of the DMs (Pastor et al., 2022).

Application and Validation to the European Innovation Scoreboard

The study applies the proposed methodology to data from the European Innovation Scoreboard, which is used to rank countries based on innovation indicators. By employing different utility functions and weight configurations, we demonstrate how rankings can vary based on subjective or objective evaluations. This allows for a more nuanced understanding of each alternative’s performance relative to others under various weighting schemes.

The hybrid methodology’s advantage lies in its capacity to leverage the strengths of both CI and DEA-BoD, offering a balanced evaluation that recognizes both the DM’s preferences and the objective performance of alternatives. This helps to bridge the gap between subjective and objective assessments, making it useful for situations where impartiality and fairness are critical.

ABSTRACT. ● The Core Challenge: Despite significant scientific and technological advancements, the U.S. faces persistent, deeply entrenched societal problems, including socioeconomic disparities, unequal access to opportunity, and uneven resilience to global challenges. This indicates that the current national innovation system, and the science policy guiding it, is not effectively oriented or structured to translate progress into equitable outcomes for all Americans. ● Informed by Prior Work: This perspective draws on extensive work highlighting the limitations of traditional approaches to science and innovation policy in addressing complex societal challenges. Prior work by Dr. Crow has examined the need to rethink structures within science funding agencies ("Time to Rethink the NIH"), the imperative to link scientific research more directly to tangible societal outcomes ("Linking Scientific Research to Societal Outcomes"), and the argument that the science system can become too focused on traditional pursuits rather than engaging with real-world problems ("Overcoming Stone Age Logic"). Discussions on the future of U.S. science policy acknowledge persistent disparities and the need for research structures to adapt ("The Next 75 Years of US Science and Innovation Policy: An Introduction"). The concept of transforming institutional models to achieve greater societal impact and inclusivity ("Designing the New American University") supports the view that a reorientation is necessary for science policy to contribute effectively to broader societal well-being. ● A Novel Analytical Approach: The 1/3 Country Framework: To illustrate the gap between innovation capacity and societal well-being, the presentation utilizes a unique analytical framework that reimagines the United States population divided into three "countries" based on household income tiers: the Bottom Third, Middle Third, and Top Third. ○ Methodology: This analysis moves beyond simple averages by using income quantiles (specifically, 60 ventiles or 5% population groups within these terciles) as the primary unit of analysis. This approach allows for the examination of how a wide range of social and economic indicators vary across the entire income distribution, revealing granular gradients and disparities within and between these income "countries" over time. This provides a detailed picture of the "lived experience" associated with different income levels, which is often obscured by national aggregate data. ● Evidence from the 1/3 Country Analysis: The presentation's central evidence derives from a series of visualizations generated from this analysis, organized into key domains. The data reveals significant disparities across the income spectrum, indicating where the current innovation system's benefits and opportunities are not reaching all segments of the population: ○ Demographics & Context: The analysis identifies how fundamental demographic characteristics like age and racial diversity are distributed across income levels, providing context for understanding the populations experiencing different outcomes. ○ Economic Opportunity & Hardship: The data reveals the unequal distribution of economic well-being, including differences in income, employment stability, and vulnerability to poverty. This highlights how economic opportunity is not uniformly accessible across the population. ○ Education: The analysis demonstrates how access to and outcomes from education, a factor in social mobility and participation in the innovation economy, are stratified by income. This indicates how the current system may contribute to unequal opportunities for human capital development. ○ Living Standards: The data depicts differences in the material conditions of life, including access to housing, transportation, and technology. This shows how basic living standards vary across income groups. ○ Health: The analysis reveals the impact of socioeconomic status on health and well-being, including disparities in health insurance, self-rated health, and the prevalence of chronic conditions. This indicates how health outcomes are stratified by income. ○ Civic Engagement: The data explores how socioeconomic status correlates with participation in democratic processes and access to political influence. This suggests how disparities in the innovation system may relate to unequal voice and representation in shaping policy.

● Call to Action and Conclusion—Designing for Wicked Problems: The patterns revealed by this analysis indicate that a national innovation system not explicitly designed to address systemic inequities may not achieve broad social progress. The persistent disparities across the 1/3 Countries suggest that current science policy approaches are insufficient. Science policy must be intentionally reoriented to identify, prioritize, and tackle "wicked problems"—complex, interconnected societal challenges that require integrated, transdisciplinary research and innovation efforts focused on equitable outcomes and societal impact. The 1/3 Country analysis provides a data-driven case for this transformation, emphasizing that achieving national prosperity and resilience requires a science policy explicitly committed to fostering opportunity and well-being for all segments of the population.

ABSTRACT. This paper explores the link between innovation and Global Value Chain (GVC) participation, especially focusing on how Africa could strengthen its GVC integration. Through a systematic search and review of existing research, the study emphasizes the significance of National Innovation Systems (NIS) for technological progress and potential industrial growth in Africa. Despite Africa facing challenges like weak research infrastructure, limited funding, and skills gaps, there are opportunities for leveraging innovation to enhance GVC participation.

The research aims to establish a foundation for Africa to boost its GVC integration through innovation, potentially accelerating industrialization. A “systematic search and review” method identified and evaluated relevant literature, focusing on innovation, GVCs, and industrialization in Africa. The study involved selecting key academic databases, using specific search terms related to innovation and GVCs, and setting criteria to include relevant qualitative and quantitative studies while excluding irrelevant sources.

Data analysis involved a narrative synthesis to pinpoint themes such as the role of government policies, technology adoption, and collaboration within GVCs. By identifying patterns across the literature, the study highlighted how governments and firms must work together to enhance Africa’s GVC involvement. Strategic government interventions like industrial policies and infrastructure investment are essential, along with proactive strategies by African firms, such as upgrading technology and expanding exports. These can make African businesses more competitive globally.

The study also drew lessons from countries like China and South Korea, which successfully integrated into GVCs through government support and innovation promotion. African governments can take similar steps by fostering research and development, technology transfer, and creating innovation hubs to accelerate growth.

At the firm level, strategies such as technology upgrading, strategic alliances, and capacity-building programs are recommended to improve Africa’s GVC participation. Export activities are essential for African firms to access new technologies and global best practices. The paper emphasizes the importance of cross-border collaboration and regional partnerships in fostering innovation, sharing knowledge, and boosting industrialization.

Key research gaps identified include mapping African GVC participation, understanding GVC dynamics in Africa, promoting technology adoption, and evaluating policy frameworks supporting innovation. The study proposes a research agenda focused on fostering innovation-led growth within African GVCs, with priorities such as enhancing technology adoption, improving policy support, and promoting cross-border collaboration. By addressing these areas, Africa can better position itself within global networks, promoting sustainable economic growth.

ABSTRACT. The growing ubiquity and sophistication of algorithms have transformed industries and economies worldwide, particularly within the digital economy. Algorithms drive innovation and operational efficiency, enabling the emergence of platform economies that bridge suppliers and consumers through data-driven matchmaking and management systems. Ride-hailing platforms, in particular, rely on algorithms to optimize operations, from dispatching drivers to evaluating worker performance. While these algorithmic systems enhance productivity and reduce transaction costs, they also introduce significant challenges. The opacity and complexity of algorithmic decisions often undermine worker autonomy and exacerbate power asymmetries between platforms and workers. This imbalance raises critical questions about fairness, accountability, and transparency in algorithmic management. Furthermore, traditional regulatory frameworks struggle to adapt to the unique challenges posed by algorithm-driven economies, necessitating innovative approaches to algorithmic governance. Despite a growing body of research on algorithmic management, little is known about how to design regulatory systems that balance efficiency with fairness while addressing the societal impacts of algorithmic decisions. This study employs a simulation-based approach to analyze the implications of algorithmic strategies in ride-hailing platforms, providing actionable insights into algorithm regulation and platform governance.

The research utilizes an agent-based simulation model, constructed on a Markov Decision Process (MDP) framework, to explore the dynamics of algorithmic decision-making within a two-sided ride-hailing market. The model examines how algorithms balance efficiency and fairness in the context of platform operations, focusing on key metrics such as driver income, income distribution fairness, passenger waiting times, and platform profitability. The MDP framework defines state variables, including the number of active drivers and passengers as well as average driver working hours, and models platform actions as trade-offs between prioritizing proximity in dispatching and promoting income fairness among drivers. The reward function evaluates outcomes based on system performance, enabling the simulation to assess the long-term impacts of different algorithmic priorities. Two optimization algorithms, Proximal Policy Optimization (PPO) and Deep Q-Networks (DQN), are employed to simulate the evolution of platform strategies under varying objectives, such as maximizing profitability or ensuring fairness. The simulation iteratively updates the market state based on interactions between drivers and passengers, offering a detailed analysis of the trade-offs involved in algorithmic management. By systematically varying simulation parameters, the study evaluates the robustness of its findings and explores the potential impacts of platform decisions on various stakeholders.

The findings reveal the intricate trade-offs inherent in algorithmic management. Efficiency-focused strategies, which prioritize passenger wait times and platform profitability, often lead to disproportionately low wages and extended working hours for drivers. Conversely, fairness-focused strategies improve income equity among drivers but result in longer passenger waiting times, potentially undermining user satisfaction and platform growth. These results highlight the challenges of simultaneously achieving high driver income, low passenger wait times, and maximum platform profitability. Furthermore, the study underscores the critical role of transparency in algorithmic governance. The opacity of algorithmic decision-making processes not only exacerbates power imbalances between platforms and workers but also erodes trust and fairness in the marketplace. Enhancing algorithm transparency and accountability is thus essential to fostering equitable outcomes and restoring stakeholder trust in platform economies.

The implications of these findings extend to policymakers and platform operators, underscoring the need for adaptive governance frameworks that balance efficiency and fairness. Policymakers should establish regulatory standards that promote algorithmic transparency and fairness while encouraging platforms to adopt balanced strategies that consider broader societal impacts. Collaborative governance, involving regulators, platforms, and stakeholders, is crucial to developing innovative solutions that address the inherent trade-offs of algorithmic management. Additionally, platforms must prioritize worker empowerment, providing tools and resources that allow workers to better understand and influence the algorithms that shape their livelihoods. These measures can help ensure that algorithmic systems serve the interests of all stakeholders, contributing to a more equitable and sustainable digital economy.

This study provides a critical contribution to the ongoing discourse on algorithmic governance, highlighting the need for innovative regulatory approaches that align technological advancements with social welfare. By integrating simulation-based analysis with practical policy recommendations, the research lays the groundwork for a balanced approach to algorithmic management in platform economies, offering valuable insights for future research, policymaking, and platform governance. The findings emphasize the importance of transparency, equity, and accountability in algorithm design and implementation, advancing the broader goal of achieving sustainable and equitable outcomes in the digital economy.

ABSTRACT. Research question

What similarities, differences, approaches, and particularities characterize patenting policies and decisions for computer-implemented inventions in the European Union and the Andean Community?

Software is protected by copyright,however, over the past few decades, technological advancements have associated software with new functionalities and creations that open the possibility of claiming protection through the patent regime, due to the fulfillment of the requirements of novelty, inventive level, or industrial application. This is a golden age of technology marked by the great abundance of data and computing capacities, giving rise to technological developments applied to countless areas of science in general and thus to the patenting of what has been called ‘computer-implemented inventions’ (hereinafter CII).

From the legal point of view, TRIPS (WTO. 1994) is one of the most significant legal instruments of the 20th century concerningIP and the multilateral trade system, since it is the affirmation and recognition that nearly all consumer products undergo transformative processes involving raw materials or inputs, culminating in a final product that involves IP rights

However, there are many debates that can arise around the technological change involving CIIs and their legal aspects, as TRIPS does not eliminate legal concerns about the difficulties that many patent offices and courts have in distinguishing patentable from non-patentable subject matter, an issue that has led to the creation of guides to facilitate this work.

On the other hand, middle- and low-income countries consider that protecting software through the patent system entails lengthy prosecution processes that sometimes do not coincide with its life cycle. Additionally, it inhibits competition by generating access barriers, since it hinders interoperability between programs, systems and networks. On the other hand, the high costs involved in patent protection which negatively affect medium-sized and small companies or independent developers when they wish to protect their inventions in different countries (AIPPI, 2017; Saiz, 2022).

Thus the intention of this summary is to present the results of project research in which the scope and limits of Cs were analyzed based on: (a) reviewing its conceptual evolution in the scientific literature indexed in the Scopus database between the years 2015 to 2022; (b) reviewing its legal treatment in regulatory frameworks of multilateral organizations such as the World Trade Organization (WTO) and the World Intellectual Property Organization, as well as the development of policies, standards and guidelines of regional bodies -in the case of the European Union and the Andean Community-; and (c) reviewing some contributions of actors interested in the patent system and relevant to achieve a better understanding of the limits and scope of CIIs.

Methodology

Analyzing and interpreting complex scenarios necessitates examining CIIs through multiple approaches or dimensions, in order to provide greater depth to the reflection on their scope (Phillips and Ritala, 2019). Thus, in this article, the following reflections are put forward: (a) from the conceptual dimension, the literature review enables the analysis of the elements defining computer-implemented invention and distinguishes them from those pertaining to software; (b) from the regulatory dimension, the analysis of international treaties, policies, guidelines and regional mandates, facilitates the identification of characteristics unique to functionalities of computer-implemented inventions protected under patent system and the different levels of development oriented to their protection; c) from the structural dimension, with respect to the actors, a presentation is made of those interested in the development of computer-implemented inventions, in order to understand the dynamics, activities and contributions.

The research was developed from a qualitative approach and an exploratory scope. To achieve the purpose, a literature review was conducted based on the following phases: 1) data collection, 2) data analysis and 3) review report.

Subsequently, in the selection process of documents collected in the Scopus database, the following inclusion criteria were applied: 1) that they were publications between 2015 and 2022; 2) that they included one or more of the admitted keywords; 3) selection by abstract content; and 4) that they were articles published in open access, although those obtained in the databases licensed by the institutions involved were added.

Results

Regarding the conceptual dimension, the evidence suggests that clarity has emerged over time if we compare the definitions of CIIs proposed between 2010 to 2022 in the literature as well as by patent offices.

Regarding the regulatory dimension, some of the key challenges identified, which motivated this research, focused precisely on this component, since the unifying role of multinational organizations on the term CIIs is not evident in the regulatory framework. In other words, the treaties administered by the WIPO and the WTO are not a clear reference on their protection and, instead, jurisprudence is the source of law from which clarity, limits and scope have been granted to what is understood by this type of inventions. This shows that it is especially for countries with a Latin and positivist tradition that the difficulties are accentuated by the regulatory vacuum that this implies.

Regarding the structural dimension, which refers to the actors, the structure shows a relationship in which the linear and descending institutional hierarchy does not provide clarity in terms of legal certainty. The definition of CIIs' limits originates not from international entities but from regional and national governments at two levels: administrative in the patent offices and judicial in tribunals and high courts.

ABSTRACT. One of the most recent challenges posed by the European Union is the transition towards environmental friendly technologies to address major challenges such as the conflicting trends of a growing population and the limited amount of natural resources (European Commission, 2023). In light of this, a strategy is proposed at the European level involving a variety of organizations (e.g. firms, universities and research centers) to renew their knowledge and skill sets to be able to develop these technologies (JRC, 2022). Germany as well as other European countries are currently introducing new green technologies to meet the standards imposed by EU. Given its diverse and broad research infrastructure comprehending both public (universities, universities of applied sciences, technical universities and research institutes) as well as private entities (Graf & Menter, 2022), Germany is a particular interesting research case. Green technologies are complex in their nature and they rely on a higher extent on developments put forward in other technologies (Barbieri et al., 2020, 2023). The literature on technological development focuses mainly on knowledge recombination activities put forward by inventors (Cassi & Plunket, 2015). Mainly due to data limitations, only recently new studies shifted towards assessing the recombination impact of organizations for new technological advancements (e.g. Wanzenböck et al., 2024). We continue this line of research investigating the, potentially different, roles of public and private organizations in making new knowledge combinations with a particular focus on green technologies given their strategic importance. In particular, we focus on a combinatorial function called uniqueness regarded as important for knowledge recombination in general (Su et al., 2013) and even more in the case of green technologies (Barbieri et al., 2020). After identifying from which typology of organization the majority of unique combinations in the green realm come from, we also analyze their impact on subsequent inventions. We further distinguish between impact on subsequent green technologies or on other non-green technologies. Particularly in the context of a green transition, it is important to assess if the new knowledge produced by those inventions is also subsequently re-utilized or not, and in which knowledge fields.

In general, the development of new technologies passes through a process called technological recombination. Knowledge elements (both old and new) are combined in new ways to produce technological advancements (Weitzman, 1998). Different bits of knowledge necessary to produce new technologies employed in a region are called knowledge base (Kalthaus, 2020). The inventors are the ultimate agents which create these new combinations. However, the unit which provides the resources (both tangible and intangible) and ultimately profit (in terms of reputation or revenues) from those inventions are the organizations such as firms, universities or research institutes. But, different types of organizations have different propensities to combine pieces of knowledge due to their orientation towards basic or applied research. Basic research, in the classic view, is focused on introducing new revolutionary technological combinations without a clear application. Whereas applied research is usually redefining knowledge that is already existing to produce and commercialize products and services. Moreover, basic research is usually regarded as more difficult to be performed but with a higher impact on subsequent inventions. Whereas, applied research is usually easier to perform but with a limited impact on further knowledge development (Graf & Menter, 2022). Thus, given those properties we expect to observe differences in recombination activities of private and public organizations with regard to green technologies. In fact, by relying to a larger extent on developments put forward in other technological realms (Barbieri et al., 2020) we expect also to find different effects on subsequent inventions when considering different organizations.

We use OECD REGPAT September 2024 for identifying inventive activities. The focus is on the patents which have at least one inventor with a German address. The considered time period is between 1990 and 2021. We use CPC classes at 5 digit level to identify the knowledge elements. Following Graf & Menter (2022) using the Regpat HAN database, patents are associated using an algorithm to the organizations that filed them. Green technologies are captured by all patents with at least one class falling in the three-digit CPC classification Y02. To measure the uniqueness of combination, we rely on a very well known indicator in Social Network Analysis called Redundancy Coefficient. This indicator is able to capture the extent to which the technological combinations created by a patent are unique in the entire German knowledge base. Finally, in line with other literature in the field to assess the impact on subsequent inventions we rely on forward citations.

In general, we contribute to the literature on technology recombination and more specifically on the literature about green technology development. Our results are also important for policymakers, who should consider how organizations can renew and adapt their knowledge bases in order to accelerate the transition to green technologies.

References Barbieri, N., Marzucchi, A., & Rizzo, U. (2020). Knowledge sources and impacts on subsequent inventions: Do green technologies differ from non-green ones? Res. Policy, 49(2), 103901. Barbieri, N., Marzucchi, A., & Rizzo, U. (2023). Green technologies, interdependencies, and policy. JEEM, 118, 102791. Cassi, L. & Plunket, A. (2015). Research collaboration in co-inventor networks: combining closure, bridging and proximities. Reg. Stud, 49(6), 936–954. European Commission (2023). The european green deal industrial plan. Accessed: 2024-09-23. Graf, H. & Menter, M. (2022). Public research and the quality of inventions: the role and impact of entrepreneurial universities and regional network embeddedness. Small Bus. Econ., 58, 1187-1204. JRC (2022). Towards a green & digital future. Joint Research Centre (JRC), European Commission. Kalthaus, M. (2020). Knowledge recombination along the technology life cycle. JEE, 30(3), 643–704. Su, Z., Ahlstrom, D., Li, J., & Cheng, D. (2013). Knowledge creation capability, absorptive capacity, and product innovativeness. R&D Management, 43(5), 473–485. Wanzenböck, I., Rocchetta, S., Kim, K., & Kogler, D. F. (2024). The emergence of new regional technological specialisations: exploring the role of organisations and their technological network structure. Industry and Innovation, (pp. 1–27). Weitzman, M. L. (1998). Recombinant growth. QJE, 113(2), 331–360.