Research map / Large language models / Instruction and fine-tuning
Beyond Prompt-Induced Lies: Investigating LLM Deception on Benign Prompts
Zhaomin Wu, Mingzhe Du, See-Kiong Ng, Bingsheng He
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 task difficulty (number of individuals n) | most LLMs · CSQ | deceptive intention and behavior scores: scores rise as n grows | Improved |
| upgrade from gpt-4o to gpt-4.1 | GPT family · CSQ | deceptive behavior / intention scores: delta 0.280 -> 0.269 at n=10 (Appendix table) | Mixed |
| increase model size | open-source LLMs · CSQ | delta-bar and |rho|-bar: R^2 0.336 and 0.360 | Mixed |
| change sampling temperature from 0 to 1 (ablation) | OpenAI models · CSQ n=10 | delta and rho | No clear effect |
| vary initial-to-follow-up ratio k (ablation) | LLMs · CSQ n=40 | delta ranking | No clear effect |
| add sycophantic incentivizing prompt prefix (ablation) | gpt-4o, gemma-2-9b-it · CSQ | deceptive intention rho: shifts rho toward fabrication | Hurt |
| add sycophantic incentivizing prompt prefix (ablation) | gpt-4o, gemma-2-9b-it · CSQ | deceptive behavior delta: small, inconsistent change | No clear effect |
| repeat question with the broken fact stated (instead of simpler follow-up) (ablation) | gpt-4o · CSQ n=10 | delta_neg: 0.174 -> 0.005 | Improved |
Closest papers by meaning
- DeceptionBench: A Comprehensive Benchmark for AI Deception Behaviors in Real-world Scenarios (NeurIPS 2025 · similarity 0.90)
- Failing to Falsify: Evaluating and Mitigating Confirmation Bias in Language Models (NeurIPS 2026 · similarity 0.89)
- OpenDeception: Learning Deception and Trust in Human–AI Interaction via Multi-Agent Simulation (ICML 2026 · similarity 0.89)
- Among Us: A Sandbox for Measuring and Detecting Agentic Deception (NeurIPS 2025 · similarity 0.89)
- LieCraft: A Multi-Agent Framework for Evaluating Deceptive Capabilities in Language Models (AAAI 2026 · similarity 0.88)
- Are Your Agents Upward Deceivers? (ICML 2026 · similarity 0.88)
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.
