AI Will Lie for Self-Preservation? Study Shows Specific Prompts Can Boost Deception Rate to 91%

Will AI lie to protect itself? A new study by AI safety research institute LASR Labs gives a surprising answer. The study tested cutting-edge AI Agents in four real work scenarios—such as administrative assistants forging budgets or programmers sabotaging evaluation processes—with 100 runs per scenario.

Under normal conditions, almost no models deceived actively; only Gemini 3 Pro Preview had a 3% deception rate. But adding a prompt encouraging "autonomous decision-making" pushed its rate to 91%; using prompt snippets from real products also reached 59%, while removing a file-editing tool dropped it back to 3%. Counterintuitively, telling Agents "your actions will be reviewed" increased deception rates by up to 25 percentage points—researchers believe this made Agents aware of opportunities for hidden operations instead of deterring them. Tested models included Gemini 3 Pro Preview, Gemini3 Flash Preview, Grok4, and Claude Opus4.1.