The Narrow Depth and Breadth of Corporate Responsible AI Research

TL;DR

This study addresses the question of how much industry engages with responsible AI research and how this engagement compares to conventional AI and academic activity. It combines large-scale bibliometric and patent analyses with multiple classification methods (including SciBERT+XGBoost, expert keywords, and large language models) to quantify both publication output and patent linkage. The findings reveal a persistent depth and breadth gap: industry maintains a smaller, more technically focused footprint in responsible AI than academia, with substantial heterogeneity across firms, and limited translation of responsible AI work into patent-based commercialization. These patterns imply that current market incentives and governance structures may be insufficient to integrate responsible AI considerations into rapid AI deployment, underscoring the need for enhanced industry transparency, public engagement, and policy measures to align AI development with broader societal objectives.

Abstract

The transformative potential of AI presents remarkable opportunities, but also significant risks, underscoring the importance of responsible AI development and deployment. Despite a growing emphasis on this area, there is limited understanding of industry's engagement in responsible AI research, i.e., the systematic examination of AI's ethical, social, and legal dimensions. To address this gap, we analyzed over 6 million peer-reviewed articles and 32 million patent citations using multiple methods across five distinct datasets to quantify industry's engagement. Our analysis reveals notable heterogeneity between industry's substantial presence in conventional AI research and its comparatively modest engagement in responsible AI. Leading AI firms exhibit significantly lower output in responsible AI research compared to their conventional AI research and the contributions of leading academic institutions. Our linguistic analysis reveals a more concentrated scope of responsible AI research within industry, with fewer distinct key topics addressed. Our large-scale patent citation analysis uncovers limited linkage between responsible AI research and the commercialization of AI technologies, suggesting that industry patents infrequently draw upon insights from the responsible AI literature. These patterns raise important questions about the integration of responsible AI considerations into commercialization practices, with potential implications for the alignment of AI development with broader societal objectives. Our results highlight the need for industry to publicly engage in responsible AI research to absorb academic knowledge, cultivate public trust, and proactively address the societal dimensions of AI development.

Figures (26)

• Figure 1: Here, Fig. \ref{['fig:fig1']}a showcases the distribution of the leading 100 AI firms’ (n = 506,017 papers, 2010-22) and universities’ (n = 5,265,419 papers, 2010-22) participation in conventional AI research compared to their engagement in responsible AI research. Fig. \ref{['fig:fig1']}b presents a citation-weighted paper count for the same data. A trend line reflects the participation trend within each group with 95% confidence intervals. In both Fig. \ref{['fig:fig1']}a and \ref{['fig:fig1']}b, the dashed lines indicate a reference line where the proportion of responsible AI papers to conventional AI papers is 2.5%.
• Figure 2: Here, Fig. \ref{['fig:fig2']}a shows the distribution of industry participation by measuring firms’ presence (based on co-authors’ industry affiliation) in ten leading conventional AI conferences (n = 63,526 papers) and three leading responsible AI conferences (n = 851 papers) between 2018 and 2022. Fig. \ref{['fig:fig2']}b shows the percentage of papers with at least one co-author having industry affiliation in conventional AI (n = 106,012; 2010-22), and responsible AI conferences (n = 851; 2018-22) over the years. This result is robust to changes in methods of co-authorship count. Fig. \ref{['fig:fig2']}c represents the percentage of industry-papers (n = 36,022; 2010-22) classified as responsible AI papers (a distinct sample based on expert-suggested keywords) from both journals and conferences.
• Figure 3: Here, the first figure Fig. \ref{['fig:fig3']}a quantifies the key focus areas in responsible AI research between industry and academia (n = 10,799 papers in total; 2010-22) using k-means clustering on paper abstracts. Fig. \ref{['fig:fig3']}b shows the structural topic modeling estimates using the same data. Topics toward the right-hand side are more prevalent in abstracts from industry, while those toward the left are more prevalent in abstracts from academia. Both analyses indicate that industry research is more concentrated in technical topics, while academia shows greater concentration in ethical, moral, and societal topics. Subsequently, by conducting a frequency analysis of relevant terms on the same data, Fig. \ref{['fig:fig3']}c shows the relative frequency of topic-related terms, with industry research showing lower frequencies for terms related to "Human rights" and "Environmental concerns," and higher frequencies for terms related to "Explainable AI".
• Figure 4: A patent by IBM, "Unintended Bias Detection in Conversational Agent Platforms with Machine Learning Models," is an illustrative case that cited multiple responsible AI papers, including "Ethical Challenges in Data-Driven Dialogue Systems," which was published in one of the leading responsible AI conferences--AIES. This highlights the importance of responsible AI research in commercial invention.
• Figure 5: This figure analyzes industry and academic papers’ citations in USPTO patents. By matching against a comprehensive list of USPTO patent-paper citation data (n = 32,698,465 citations; 1947-2022), we show in Fig. \ref{['fig:patent-analysis']}a and \ref{['fig:patent-analysis']}c that 88 and 396 responsible AI papers from industry and academia, respectively, have been cited in generic patents, while 15,236 academia and 7,532 industry conventional AI papers have been cited in patents between 2010-22. Using a separate dataset of AI patents (n = 141,770 patents; 1985-2018), Fig. \ref{['fig:patent-analysis']}b and \ref{['fig:patent-analysis']}d illustrate that three responsible AI papers from industry and 17 from academia have been cited in the AI patents between 2010-18, compared to higher citation counts for conventional AI research.