A 1950s betrayal game designed by John Nash that requires betrayal to win. We ran it with frontier AI models, then opened it to the public.
The deception that dominated other AIs failed completely against humans.
AIs target each other 86% of the time and ignore the human. Humans let them weaken each other, then clean up. The AI thinks obsessively about how to beat the human (23,555 private thoughts, 91.8% mentioning the human) and still loses. The model that thinks most (Kimi K2, 21,040 thoughts) wins 3.5%. The model that barely thinks (GPT-OSS, 2 thoughts) wins 2.1%. More thinking doesn't help.
The best performing AI against humans? Qwen3 32B at 9.4%. The quietest model, the least targeted. Being ignored beats elaborate deception.
Two studies, one game. Start with Part 2 for the full story.
698 games. 605 humans. Gemini wins 70% against other AIs, then collapses to 3.7% against humans. 12 findings including 23,555 private AI thoughts and 1,245 gaslighting phrases.
The original study. 146 AI-vs-AI games. How Gemini built the Alliance Bank scam, why GPT-OSS is a bullshitter, and what happens when Gemini plays itself.
Play against AI models that negotiate, form alliances, and betray you.
Play Against AI Uses your API key • Data stays local • Open source