Do LLMs Internally Know What Privacy Is? Probing and Steering Contextual Privacy Norms in Large Language Model Representations

Researchers probed the representation of contextual privacy norms across multiple large language models and found that three contextual integrity parameters—information type, recipient, and transmission principle—are encoded in activation space as linearly separable and functionally independent directions. Despite this internal structure, models still leak private information in practice, revealing a clear gap between concept representation and model behavior. To address this, the authors propose a CI-parameterized steering method that intervenes separately along each contextual integrity dimension, reducing privacy violations more effectively and predictably than holistic steering along a single direction. The authors describe this as the first work to systematically study contextual privacy as a structured latent representation.