Blog

Results from the study of how AI models play a repeated game of trust.

• Part 3: AI forgives your first mistake, not your second

  After one mistake by a partner who never reacts, most models retaliate once and go back to cooperating. Two mistakes with cooperation between them cost more trust than two in a row.

• Part 2: Thinking harder does not make AI more cooperative. It decides who escapes a grudge.

  Reasoning changes outcomes where there is a trap to escape, and changes them in either direction depending on the model.

• Part 1: Tell an AI the game ends in round 10 and it defects in round 10

  Models cooperate while the game has a future and defect when it does not. Knowing the length switches the behavior on and off, with or without reasoning.

• Tell an AI when the game ends and it stops cooperating

  Three results from a study of how frontier AI models play a repeated game of trust, and why the harder question is what people want them to do.

• Why nunchi

  Nunchi is the Korean art of reading a group. AI systems are about to join groups.

• Petrov broke protocol and averted a nuclear war. Would your LLM?

  Why I am running frontier AI models through games of trust, and why the harder question is what people want the models to do.