User-Centered Auditing Needed for Personalized Generative AI Systems

A new position paper argues that current auditing methods for generative AI systems fail to detect harms that emerge in personalized contexts. The paper, available on arXiv (2608.14692), contends that existing approaches rely on static, simulated evaluations and aggregate definitions of harm, which are inadequate for systems that adapt to individual users over time. The authors identify three flawed presuppositions: harms can be specified outside real-world interaction, defined non-pluralistically within groups, and treated as static. They propose that auditing must occur at the interaction level, focusing on user-centered evaluations that capture emergent harms evolving with user history. The paper suggests that personalized systems could potentially learn definitions of harm through repeated interactions, but emphasizes the need for new auditing paradigms. This work contributes to the ongoing discourse on AI accountability and safety, particularly as generative AI becomes more integrated into daily applications.