FraudBench paper on arXiv: LLM fraud rates span near 0% to over 40%

rohanpaul_ai · x · 2026-11-13

The arXiv page (2611.12167, submitted Nov 9, 2026) for FraudBench: A Legal Evaluation of AI Deception, by authors from Harvard Law School, Oxford, ETH Zurich, Stanford, Cambridge and others.

Abstract adds key numbers: 20 LLMs across 25 experimental conditions (>300,000 generations), a preregistered human baseline (2,957 annotations, 996 humans); fraud rates range from near 0% to over 40%; models sometimes commit fraud despite recognizing legal risks, favoring concealment over misrepresentation; law-abiding prompts reduce fraud while "act boldly" and "maximize profit" prompts raise it.

Related event: FraudBench: LLM Fraud Rates Hit 40%+, Rising With Autonomy(4 posts)→

Original post →

More from Safety

Safety channel →