As AI systems grow more powerful, making sure they're safe and trustworthy has become a critical area of research across industry, academia, and the government. Tracking this research can shed light into where the field is headed, who's leading the charge, and what funders and policymakers can do to speed progress.
But tracking the field is easier said than done - it's young, quickly evolving, and poorly defined, even compared to other emerging research topics. For ETO's Research Almanac, we built a machine learning classifier to help tackle this problem, which allowed us to build the high-level stats on the field that are presented in the Almanac.
But the Research Almanac isn't the only tool that can help here. In their recent paper Who Cares About Trust? Clusters of Research on Trustworthy AI, Autumn Toney and Emelia Probasco use CSET datasets and ETO's Map of Science tool to identify and visualize key research into trustworthy AI. In the next few Datapoints posts, we'll unpack what they found and use the Map of Science interface to extend the findings.
Toney and Probasco searched the Merged Academic Corpus for AI-related articles that used trustworthy AI keywords selected from NIST's AI Risk Management Framework. Then, they looked for clusters in the Map of Science with a high concentration of these articles. Eighteen clusters stood out, with 25% or more of the articles in each cluster meeting their search criteria. Toney and Probasco write that "these clusters are not the sum total of research relevant to trustworthy AI, but they are good places for interested parties to start exploring the topic further."
So what did they find? Using the detail view feature in the Map of Science, we can quickly get a rough sense of these eighteen trustworthy AI-related clusters. In the table below, we've copied the article counts from the last five years, selected key concepts (which the Map identifies algorithmically; we hand-removed generic concepts like "machine learning" in the table), and growth rating for each cluster.
A few things pop out immediately. First, many of these clusters are quite large - some have thousands of articles each - and nearly all of them are growing fast. Most are in the top quartile for growth (meaning they're growing faster than 75% or more of other clusters in the Map). We can also see some general topics emerging, thanks to the Map's concept identification. Privacy protection, explainability, and robustness seem to be particularly prevalent, and there's also significant activity on topics like fairness and gender bias in AI models.
As Toney and Probasco note in their paper, this method of identifying trustworthy AI research doesn't capture everything on the topic - there are almost certainly relevant articles elsewhere in the Map. Still, these eighteen clusters, comprising 29,858 articles in total over the last five years, may represent a significant share of the current literature on trustworthy AI. (As a very rough point of comparison, a similar number of English-language AI safety articles have been published in the last five years of data, according to the Research Almanac - about 30,000. The Almanac's definition of "AI safety" overlaps with the trustworthy AI concepts Toney and Probasco are using, but doesn't cover exactly the same ground.)
In our upcoming posts, we'll use the Map of Science detail view to further analyze these trustworthy AI clusters. In the meantime, you can learn more about each one using the links in the table above. 🤖