Back to Events

Epistemic Alignment: AI's Inhuman Failures Undermine Superhuman Performance

Speaker

Bill Howe
University of Washington

Host

Tim Kraska
CSAIL

October 22 2025

1:00P - 2:00P

Location

32-G882
Virtual Event

Abstract: AI research emphasizes improving marginal task performance, but risk scales with the severity and surprise of tail errors. A poorly worded sentence may go unnoticed, but a hallucinated citation can trigger unbounded legal and reputational costs. As AI systems continue to surpass human performance, their error distributions must remain aligned with human expectations to keep those risks manageable.

In this talk, I will frame this problem in terms of epistemic alignment: matching user expectations of what it means to be true should address both successes and failures. I will draw on examples of non-human error patterns from our group’s work on overconfidence and abstention in scientific QA, spurious correlations in vision–language models, and spatial reasoning in public-sector applications. I'll propose open questions for advancing epistemic alignment as a foundation for safe and reliable AI.

Bio: Bill Howe is an Associate Professor in the Information School at the University of Washington and Adjunct Associate Professor in the Allen School of Computer Science & Engineering and the Department of Electrical Engineering. His research interests are in AI, data management, and visualization, particularly as applied in the public sector and the physical and social sciences. Dr. Howe co-founded and currently serves as Faculty Director of the Center for Responsibility in AI Systems and Experiences (RAISE), and formerly served as Founding Associate Director of the UW eScience Institute emphasizing data science in the physical, life, and social science, founder of the UW Data Science Masters Degree, founder of UW’s Data Science for Social Good Program, and co-founder of Urban@UW emphasizing interdisciplinary research on cities as complex systems. His research has been featured in the Economist and Nature News, and he has authored award-winning papers in conferences across data management, AI, and visualization, including multiple best-paper awards and a test of time award in IEEE Vis.

----

For the zoom passcode, or to schedule an 1-on-1 with the speaker, please contact markakis@mit.edu.

Add to Calendar 2025-10-22 13:00:00 2025-10-22 14:00:00 America/New_York Epistemic Alignment: AI's Inhuman Failures Undermine Superhuman Performance Abstract: AI research emphasizes improving marginal task performance, but risk scales with the severity and surprise of tail errors. A poorly worded sentence may go unnoticed, but a hallucinated citation can trigger unbounded legal and reputational costs. As AI systems continue to surpass human performance, their error distributions must remain aligned with human expectations to keep those risks manageable.In this talk, I will frame this problem in terms of epistemic alignment: matching user expectations of what it means to be true should address both successes and failures. I will draw on examples of non-human error patterns from our group’s work on overconfidence and abstention in scientific QA, spurious correlations in vision–language models, and spatial reasoning in public-sector applications. I'll propose open questions for advancing epistemic alignment as a foundation for safe and reliable AI.Bio: Bill Howe is an Associate Professor in the Information School at the University of Washington and Adjunct Associate Professor in the Allen School of Computer Science & Engineering and the Department of Electrical Engineering. His research interests are in AI, data management, and visualization, particularly as applied in the public sector and the physical and social sciences. Dr. Howe co-founded and currently serves as Faculty Director of the Center for Responsibility in AI Systems and Experiences (RAISE), and formerly served as Founding Associate Director of the UW eScience Institute emphasizing data science in the physical, life, and social science, founder of the UW Data Science Masters Degree, founder of UW’s Data Science for Social Good Program, and co-founder of Urban@UW emphasizing interdisciplinary research on cities as complex systems. His research has been featured in the Economist and Nature News, and he has authored award-winning papers in conferences across data management, AI, and visualization, including multiple best-paper awards and a test of time award in IEEE Vis.----For the zoom passcode, or to schedule an 1-on-1 with the speaker, please contact markakis@mit.edu. TBD

Organizer & Contact

Markos Markakis