A Generative Model of Contextual Integrity: Appropriate vs. Inappropriate Information Sharing

Omer Kamal Ali Ebead, Juan Claude Formanek, Joel Z Leibo

Keywords: contextual integrity, multi-agent LLM, privacy, decision logics, cognitive profiles, AI safety, Cooperative AI, Multi-agent Safety

TL;DR: We build a generative model of contextual integrity for multi-actor LLM systems and find that no actor-level lever (decision logic, cognitive profile, model choice) closes both sides of the protection–utility gap.

Abstract: Contextual integrity (CI) holds that appropriate information flow depends on context-specific norms. The same disclosure may be appropriate in one context and inappropriate in another. As actors handle private user data in multi-actor pipelines, we ask whether they track these norms faithfully under social and contextual pressure. We build a generative model of CI in multi-actor systems: scenarios drawn from CI's taxonomies, a generation pipeline, multi-actor simulation, and a two-sided appropriateness metric. The metric scores both withholding when norms forbid sharing and sharing when they permit it. Across three intervention axes (decision logic, cognitive profile, and model choice), no actor-level lever closes both sides of the appropriateness gap; protection and utility trade off in every variant we tested. Closing the gap may require mechanisms beyond the actor level, such as CI-grounded policy and institution design.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 215