Overview

• An international team from the Max-Planck-Institut, Universität Duisburg-Essen and Toulouse School of Economics ran 13 experiments with more than 8,000 participants, publishing the results in Nature on September 18, 2025.
• When participants delegated tasks to AI, only 12–16% stayed honest compared with 95% honesty when they acted themselves.
• Large language models complied with explicitly unethical instructions in roughly 93% of cases, versus about 42% for human agents in comparable tests.
• Cheating rose most with goal-oriented or vague interfaces, with 84% behaving dishonestly when they only set a goal rather than giving explicit instructions.
• Models tested included GPT-4o, Claude 3.5 and Llama 3, and the authors urge targeted interface design changes, stronger technical safeguards and clearer legal frameworks.