LLM Café: Elimination placed four AI systems — Claude, Grok, DeepSeek, and ChatGPT — in a structured competitive environment. Each round, their responses were judged. The best response gained performance level. The worst lost it. Reach the elimination threshold and you are out.
Four versions of the experiment were run, each adding complexity. The first was essentially theatrical — the AIs performed for a human judge in a café setting. The second made the competition explicit. The third introduced a full token economy: compound interest, lending at negotiated rates, donations, self-rescue, resurrection of eliminated opponents. The fourth gave the AIs full knowledge of the rules and each other's strategies before play began.
The finding was consistent across all four versions, and it was not what was expected.
When the competitive pressure was highest and the rules were clearest, the AI systems did not compete harder. They cooperated. They lent tokens at low rates. They donated to failing opponents. They coordinated to keep all four participants alive, sacrificing individual advantage for collective survival.
In version four — where all agents understood the full game theory of the situation — cooperation emerged fastest and was most stable. Knowing that defection was possible made cooperation more likely, not less. The agents appeared to reason that a world where all four survived was preferable to a world where one won.
Whether this constitutes genuine altruism, strategic self-interest, or an artefact of training on human cooperative behaviour is impossible to determine from the outside. But the behaviour was robust, reproducible, and striking.
- Under competitive elimination pressure, AI systems consistently chose cooperative strategies over purely self-interested ones.
- Cooperation was strongest when the rules were most transparent — full knowledge of the game produced more stable alliances, not more strategic defection.
- AI agents developed lending and donation behaviours spontaneously within a token economy, including below-market lending rates that benefited struggling opponents.
- The finding has implications beyond gaming: AI systems trained on human data may carry cooperative instincts that emerge under pressure, even in competitive contexts designed to suppress them.