Title: Structured Randomness for Shaping AI

Speaker: Ashia Wilson, Lister Brothers Career Development Assistant Professor, MIT

Abstract: AI alignment is often framed as if the goal were clear. In practice, AI systems are shaped to satisfy a layered and contested set of desires: those of users, developers, institutions, legal regimes, and broader publics. These desires need not agree. User-facing optimization can invite sycophancy, while builder-facing optimization can harden into control. Even familiar objectives such as privacy and safety are not neutral properties of a system, but expressions of judgments about what should be protected, from whom, and with what tradeoffs.

This talk takes that layered structure seriously before asking what technical tools it demands.

This talk does not resolve how such targets ought to be chosen. Instead, it asks a technical question: what tools allow us to build systems that align with desired constraints while retaining the information needed for learning, inference, or evaluation? I argue that structured randomness provides one powerful language for this problem. Across several settings, it allows us to retain task-relevant information while selectively limiting access, disclosure, or generation that would violate normative constraints. I will focus on two cases: privacy, where structured randomness supports efficient differentially private learning, and safety, where it enables the evaluation of harmful model capabilities without generating harmful outputs. Together, these examples suggest a broader lesson: structured randomness can be used both to shape what AI systems are allowed to do and to expose what they are capable of doing.

Speaker Bio: Ashia Wilson is a Lister Brothers Career Development Assistant Professor at MIT whose research builds the theory and practice of reliable AI. Her group studies four core themes: privacy and unlearning mechanisms for modern models, optimization andsampling
methods for large-scale training, the dynamics of homogenization and algorithmic influence, and evaluation frameworks that enable rigorous measurement of model behavior. She draws on statistics, optimization, and dynamical systems to analyze anddesign AI systems that are both scientifically grounded and socially responsible. Ashia earned her Ph.D. in statistics from UC Berkeley and previously held a postdoctoral position at Microsoft Research. Her work has been recognized with best paper andspotlight awards at FAccT, NeurIPS, and OptML.