Abstract: Policymakers and key stakeholders place significant emphasis on pre-deployment evaluations and model-based safeguards for general-purpose AI systems. However, these evaluations suffer from a wide range of challenges. In particular, they do not provide any guarantees about downstream model behaviors, especially under an adversarial model. This talk will explore these challenges, as well as paths forward.

Bio: Peter Henderson is an Assistant Professor at Princeton University with appointments in the Department of Computer Science and the School of Public and International Affairs. Previously, he received a JD and PhD in Computer Science at Stanford University. His research focuses on the intersection of AI and law. This includes building AI systems that work for the public good, making AI systems safe, and making sure that the law shapes positive outcomes for AI. His work has been covered by major media outlets like the New York Times and The Wall Street Journal. And has been cited by policymakers and courts.