Reconciling Conflicting Privacy Facets and Establishing Precedence Conditions

Establish operational frameworks that reconcile conflicts among legal requirements, social privacy norms, and individual user preferences in Large Language Model–mediated data flows, and determine the conditions under which law, social norms, or individual preference should take precedence in governing contextual integrity for data sharing and use.

Background

The paper argues that privacy management for LLMs cannot be solved purely through technical means and must incorporate sociotechnical approaches grounded in laws, social norms, and individual preferences. It highlights Contextual Integrity as a guiding framework while noting the practical difficulty of operationalizing it across real-world contexts.

The authors present recent work on benchmarks and systems that test and enforce contextual privacy norms but emphasize that conflicts arise between legal requirements, social expectations, and personal preferences. They call attention to the need for principled mechanisms to resolve such conflicts and for defining precedence among these facets, especially in LLM-mediated interactions and agentic settings.