Differential Privacy in the Clean Room - Copyright Protections for Generative AI

Watch a 46-minute lecture from the Simons Institute where Aloni Cohen from the University of Chicago explores the intersection of differential privacy and copyright protection in generative AI models. Learn about the conditions under which AI model outputs might avoid copyright infringement of training data, with a focus on differential privacy (DP) as a mathematical framework for non-disclosure. Examine the limitations of near-access freeness (NAF) in preventing copyright infringement, and discover how differential privacy relates to clean room design principles. Understand the key differences between NAF and DP approaches, particularly in handling copyrighted material and preventing model taintedness. Delve into the theoretical foundations that could help establish legal and technical frameworks for responsible AI development while respecting intellectual property rights.