Context-based normative simulacra from narrative fiction

Abstract

Information handling practices of LLM agents are broadly misaligned with privacy expectations, raising urgent questions as these systems, increasingly autonomous, proliferate. Contextual Integrity (CI), which defines privacy as appropriate information flows, provides a principled framework for addressing these questions. Although other work has drawn on CI, existing approaches double inference cost via supervisor-assistant architectures, or fine-tune on narrow task-specific data. Here, we offer an approach for learning context-specific privacy reasoning, demonstrating its efficacy in a surprising application to narrative fiction: we fine-tune LLMs (SFT, then RL) on normative simulacra (structured representations of norms and information flows) that we extract from public-domain novels. We also introduce normative grounding, where information flows are evaluated in terms of relevant, retrieved informational norms. We evaluate on five CI-aligned benchmarks spanning distinct societal contexts and conduct ablations to isolate the contributions of reinforcement learning and normative grounding.