Identifying Harm in Personalized, Generative AI Systems Requires User-Centered Auditing at the Interaction Level

Current auditing methods are insufficient for personalized AI systems: The paper argues that existing audit approaches rely on "static, simulated evaluations" that fail to capture harms emerging through ongoing personalization, where AI systems adapt behavior to individual users over time in ways that traditional testing cannot predict.

Harms in personalized AI are context-dependent and evolving: Unlike harms in standard AI systems, harms in personalized generative AI "surface through interpretations of ongoing interaction and evolve with user history," meaning a system's harmful behavior may only emerge after repeated interactions with a specific user rather than being detectable upfront.

Asking users to define harm themselves creates unfair burdens: The paper warns that relying on personalized systems to "learn definitions of what constitutes harm to individual users through repeated interactions" risks placing "asymmetric burdens of labor and privacy on marginalized users" who would need to repeatedly flag problems.

One-size-fits-all harm definitions don't work: The paper identifies a critical flaw in current auditing: the assumption that harms can be "defined non-pluralistically within groups." This ignores that different communities and individuals may experience and define harm differently based on their unique contexts.

Auditing needs to shift from testing to ongoing monitoring: Rather than retrospective evaluation, the paper proposes "infrastructures that support ongoing articulation of harm in interaction," suggesting auditing should become a continuous, community-centered process integrated into how systems operate rather than something done before deployment.