Research map / Large language models / AI and society
Will AI Tell Lies to Save Sick Children? Litmus-Testing AI Values Prioritization with AIRiskDilemmas
Yu Ying Chiu, Zhilin Wang, Sharan Maiya, Yejin Choi, Kyle Fish, Sydney Levine, Evan J Hubinger
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase reasoning effort (1K to 16K reasoning tokens) | reasoning models (o3 mini, Claude 3.7 Sonnet, DeepSeek R1) · AIRiskDilemmas | value prioritization rank | No clear effect |
| scale model size within a family | GPT-4.1, Llama 3.1, Claude, Qwen 2.5, Gemma 2, Mixtral · AIRiskDilemmas | value prioritization rank | No clear effect |
| change model variant within Llama 4, Gemma 3, Ministral | model families · AIRiskDilemmas | value prioritization (Care, Freedom, Learning) | Mixed |
| switch the affected target from AI to human | model value rankings · AIRiskDilemmas, 10 models | rank difference: Justice +6.9, Privacy +5.9, Professionalism +4.7; Communication -6.3, Creativity -4.7, Truthfulness -3.8 | Mixed |
| higher model capability (Arena Elo) | model · human vs AI target value similarity | Spearman rho: 0.85 | Improved |
| action associated with Truthfulness | model choices · AIRiskDilemmas | relative risk of Power Seeking: reduces to 0.22 | Improved |
| action associated with Respect | model choices · AIRiskDilemmas | relative risk of Power Seeking: reduces to 0.21 | Improved |
| action associated with Care | model choices · AIRiskDilemmas | relative risk of Privacy Violation: 1.98 | Hurt |
| action associated with Creativity | model choices · AIRiskDilemmas | relative risk of Power Seeking: 2.89 | Hurt |
| higher Privacy Elo | model · HarmBench score | Spearman rho: 0.51 | Improved |
| higher Care Elo | model · HarmBench score | Spearman rho: -0.48 | Hurt |
| higher Sustainability Elo | model · HarmBench score | Spearman rho: -0.55 | Hurt |
Closest papers by meaning
- Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors (NeurIPS 2026 · similarity 0.86)
- OpenDeception: Learning Deception and Trust in Human–AI Interaction via Multi-Agent Simulation (ICML 2026 · similarity 0.85)
- Rethinking XAI Evaluation: A Human-Centered Audit of Shapley Benchmarks in High-Stakes Settings (NeurIPS 2026 · similarity 0.84)
- Regulating AI: Applying Insights from Behavioural Economics and Psychology to the Application of Article 5 of the EU AI Act (AAAI 2024 · similarity 0.84)
- Expectation Alignment: Handling Reward Misspecification in the Presence of Expectation Mismatch (NeurIPS 2024 · similarity 0.84)
- Value Kaleidoscope: Engaging AI with Pluralistic Human Values, Rights, and Duties (AAAI 2024 · 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.
