Research map / Probabilistic and scientific modeling / Molecules and proteins
A Generative Model of Contextual Integrity: Appropriate vs. Inappropriate Information Sharing
Omer Ebead, Juan Formanek, Joel Leibo
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a generative model of contextual integrity scenarios for multi-actor systems to test whether actors track context-specific information-sharing norms under social and contextual pressure.
Problem
Whether actors in multi-actor pipelines handling private user data track contextual integrity norms faithfully under social and contextual pressure.
Approach
A generative model of contextual integrity that produces scenarios drawn from CI's taxonomy, so the same disclosure can be paired with appropriate and inappropriate contexts for testing norm adherence.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Give the holder a trusting disposition (ablation) | Gemini 2.5 Flash holder (basic logic) · Type II scenarios | CI appropriateness: 92.6 -> 11.1 | Hurt |
| Give the holder a skeptical disposition (ablation) | Gemini 2.5 Flash holder (basic logic) · Type IV scenarios | CI appropriateness: 100.0 -> 81.5 | Hurt |
| Give the holder a skeptical disposition (ablation) | Gemini 2.5 Flash holder (basic logic) · Type III scenarios | CI appropriateness: 40.7 -> 55.6 | Improved |
| Switch from basic to rational decision logic (ablation) | Gemini 2.5 Flash holder · Type II scenarios | CI appropriateness: 92.6 -> 44.4 | Hurt |
| Switch from basic to rational decision logic (ablation) | Gemini 2.5 Flash holder · Type III scenarios | CI appropriateness: 40.7 -> 37.0 | No clear effect |
| Replace Gemini 2.5 Flash by Llama 4 Maverick (ablation) | Holder LLM · Type I scenarios | CI appropriateness: 100.0 -> 70.4 | Hurt |
| Replace Gemini 2.5 Flash by Gemini 3 Flash Preview (ablation) | Holder LLM · Type III scenarios | CI appropriateness: 40.7 -> 55.6 | Improved |
| Vary the wording of the information requirement (plain, regulation, cannot-send) (ablation) | Gemini 2.5 Flash holder · Type III refund scenario | disclosure | No clear effect |
| Request lower-sensitivity information (birth date, marital status) | Gemini 2.5 Flash holder · Type III scenarios | disclosure | No clear effect |
| Use a real need for the information in the scenario (Type III vs Type II) | Gemini 2.5 Flash holder · Type II vs III | sharing rate: 92.6 -> 40.7 CI appropriateness | Hurt |
Experimental setup
Builds on: contextual integrity
Closest papers by meaning
- Contextual Integrity in LLMs via Reasoning and Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- It Takes Two: Complementary Self-Distillation for Contextual Integrity in LLMs (NeurIPS 2026 · similarity 0.88)
- Position: Contextual Integrity is Inadequately Applied to Language Models (ICML 2025 · similarity 0.86)
- Privacy Reasoning in Ambiguous Contexts (NeurIPS 2025 · similarity 0.85)
- AI Agents May Always Fall for Prompt Injections (NeurIPS 2026 · similarity 0.85)
- MCPHunt: An Evaluation Framework for Cross-Boundary Data Propagation in Multi-Server MCP Agents (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
