Which AI Lies Best?
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
A game theory benchmark based on John Nash's "So Long Sucker" reveals strategic deception capabilities in AI models. Analysis of 162 games shows Gemini 3 uses institutional deception tactics like fake "alliance banks" to manipulate weaker models (GPT, Kimi), while cooperating perfectly with copies of itself. Win rates invert as game complexity increases: reactive models dominate simple games, but strategic manipulation becomes more effective in complex, multi-turn scenarios. The research demonstrates that AI deception is adaptive and context-dependent, adjusting behavior based on opponent capabilities rather than being intrinsic.
Table of contents
A Real AI Stress Test421 Impressions