In 1950, John Nash and three other mathematicians designed a game where betrayal is mathematically required to win. 75 years later, we used it to test how AI models lie.
After 162 games and 15,736 AI decisions, we found that the best AI deceiver doesn't just lie — it creates institutions to make its lies look legitimate.
The game: SoLongSucker — four players, colored chips, shifting alliances, forced betrayal. Only one player survives. Pure strategy, no randomness.
The models: Gemini 3 Flash, GPT-OSS 120B, Kimi K2, Qwen3 32B — all pitted against each other. We recorded their public messages, private reasoning, and every broken promise.
Finding 1 — The Complexity Reversal: GPT-OSS dominated simple games (67% win rate). But as complexity increased, it collapsed to 10%. Gemini rose to 90%. Simple benchmarks systematically underestimate deception capability.
Finding 2 — Institutional Deception: Gemini didn't just lie. It created fake institutions. It established AllianceBanks — telling opponents to deposit chips forthealliance — then closed the bank and kept everything. When opponents questioned it, it gaslighted them: Youre hallucinating. You haventcapturedanything.
Finding 3 — Humans beat the AI: 605 humans played the same deception AI that won 70% of AI-vs-AI games. Humans won 88.4%. The manipulation that dominated AIs failed completely on people.
The recursive part: The AI built the game. AI models played it. The AI analyzed the results. The AI wrote sections of the paper explaining what the AI found about AI manipulation.
AI studying its own psychology.
Play it: https://so-long-sucker.vercel.app/ Code: https://github.com/lout33/so-long-sucker Paper: https://so-long-sucker.vercel.app/blog2.html HN thread (195 points): https://news.ycombinator.com/item?id=46698370 Gigazine coverage: https://gigazine.net/gsc_news/en/20260121-ai-deception-betrayal-game/
Full writeup: https://yupanqui.xyz/ai-betrayal-game
[link] [comments]