Get the latest news, advances in research, policy work, and education program updates from HAI in your inbox weekly.
Sign Up For Latest News
This seminar pairs a case study of AI diffusion with an organizational approach to technological changes in the workplace.
This seminar pairs a case study of AI diffusion with an organizational approach to technological changes in the workplace.
Sessions run Wednesdays from 4:30–5:30 PM in CoDa E160. Each session features a different Stanford speaker; talk titles are announced by the organizers.
Sessions run Wednesdays from 4:30–5:30 PM in CoDa E160. Each session features a different Stanford speaker; talk titles are announced by the organizers.
Talk Title: Classical Statistics in the Age of AI
Abstract: Researchers increasingly use generative AI to create “digital twins” and conduct synthetic experiments. Recognizing that LLMs often fail to capture complex real-world behavior, a growing literature has proposed novel methods for combining synthetic and ground-truth data. This talk illustrates the continued relevance of classical statistics for this challenge through two projects. In the first project, we argue that off-the-shelf linear regression is a natural approach for incorporating AI predictions into experiments, building on the randomization inference framework dating back to Fisher and Neyman. Unlike many recent proposals, standard linear regression inherits a “do no harm” property in that the adjusted estimator automatically reverts to the unadjusted difference in means when AI predictions are uninformative. In the second project, we examine how to improve the AI predictions themselves using “activation steering,” a technique from mechanistic interpretability that modifies internal LLM activations to shift behavior toward a target concept. We show that many steering methods implicitly estimate the average gradient of an outcome regression, a quantity with a long history in causal inference and econometrics. This connection immediately enables more flexible models, including Neyman-orthogonal estimators, which in turn lead to improved AI predictions for the first project’s regression framework. Together, these results show how classical statistical ideas continue to provide both conceptual clarity and empirical gains at the intersection of causal inference and generative AI.
Speaker: Avi Feller, Assistant Professor in the Goldman School of Public Policy and the Department of Statistics, UC Berkeley
Speaker Bio: Avi Feller is an assistant professor in the Goldman School of Public Policy and the Department of Statistics at UC Berkeley. His methodological research centers on learning more from social policy evaluations, especially randomized experiments. His applied research focuses on working with governments on using data to design, implement, and evaluate policies. Prior to his doctoral studies, Feller served as Special Assistant to the Director at the White House Office of Management and Budget and worked at the Center on Budget and Policy Priorities. Feller received a Ph.D. in Statistics from Harvard University, an M.Sc. in Applied Statistics as a Rhodes Scholar at the University of Oxford, and a B.A. in Political Science and Applied Mathematics from Yale University.
The logistics details for each session will be provided to registered participants. Sign up here to be included on the mailing list and learn about future events.
This talk is co-sponsored by USF's Master's in Applied Economics and the Stanford Causal Science Center. For additional information and abstracts from past talks, please click here
