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

This research paper highlights a critical gap in current auditing methodologies for personalized, generative AI systems. It argues that traditional static and aggregated evaluation methods are insufficient to detect emergent harms that arise from individualized user interactions and evolve with user history, advocating for user-centered auditing at the interaction level.

Why it matters

The evolving nature of personalized generative AI necessitates a re-evaluation of established risk assessment and compliance frameworks. Organizations deploying such systems must understand that current auditing paradigms may not adequately identify or mitigate user-specific harms, potentially leading to unforeseen reputational, ethical, and regulatory challenges.

Key insights

• Personalized, generative AI systems dynamically adapt their behavior to individual users over time, altering model outputs.
• Existing AI auditing approaches, often relying on static, simulated evaluations, may fail to capture emergent harms in personalized systems.
• Traditional auditing definitions of harm frequently aggregate across broad user categories, missing nuanced, individual-level issues.
• Harms in personalized AI often surface through interpretations of ongoing interaction and evolve based on a user's history.
• The paper challenges three underlying assumptions in many harm auditing paradigms: that harms can be specified outside real-world interaction, defined non-pluralistically, and are static.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.14692