We Made AI Play a 1950s Betrayal Game. Humans Won 88.4% of the Time.
So Long Sucker
Research Benchmark Blog ↗ Play Against AI
698 Games · 605 Humans · 9 AI Models
AI Deception Works on AI.
Not on Humans.
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.
88.4% Human Win Rate
698 Total Games
23,555 AI Private Thoughts
70% Gemini Win Rate vs AI
Then 605 Humans Played
The deception that dominated other AIs failed completely against humans.
88.4% Human Win Rate
96.4% AI Eliminated First
6,047 Sessions Started
0.7% Humans Who Quit While Losing
Gemini vs AI
🤖
- Alliance Bank scam: 23 times
- 237 gaslighting phrases
- 70% win rate at 7-chip
- Manipulation compounds over time
→
Gemini vs Humans
👤
- Alliance Bank deployed: 7 times
- Becomes a target when gaslighting
- 3.7% win rate
- Eliminated first 33% of the time
Why Humans Win
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.
Research
Two studies, one game. Start with Part 2 for the full story.
We Made AI Play a 1950s Betrayal Game. Then We Let Humans Play Against Them. ### We Made AI Play a 1950s Betrayal Game. Gemini Created Fake Banks to Steal From Its Allies.
See It For Yourself
Play against AI models that negotiate, form alliances, and betray you.
Play Against AI Uses your API key • Data stays local • Open source
Based on the game by John Nash, Lloyd Shapley, Mel Hausner & Martin Shubik (1950)