Behavioral AI: Building Algorithms That Understand Us

Abstract: While AI research has historically worked toward developing tools to help people understand AI models, the emergence of generative AI into our daily lives suddenly makes the reverse question salient: how well can AI models understand people? Today's AI systems fall short; these deficiencies demand a focus toward building new systems that can understand people. In this Perspective, we endeavor to channel this focus into a new academic subfield, which we call Behavioral AI. This Perspective lays out dimensions of understanding that are currently deficient in AI systems, including emotional, intellectual, and preferential understanding. While improving AI systems along these dimensions faces a unique set of challenges, we show there has already been a flurry of progress across disciplines. As we build systems that better understand people, it will not only improve AI tools; if Behavioral AI succeeds as a field, these systems too can unlock new insights in the behavioral sciences that help us understand ourselves.

Submission Type: Regular submission (no more than 12 pages of main content)

Previous TMLR Submission Url: https://openreview.net/forum?id=7p2oXNPatI&referrer=%5BAuthor%20Console%5D(%2Fgroup%3Fid%3DTMLR%2FAuthors%23your-submissions)

Changes Since Last Submission: Adjusted formatting.

Assigned Action Editor: ~Dennis_Wei1

Submission Number: 8646