Why it’s still open
The same capacities make agents good teammates and good conspirators. Pricing agents raise prices together without being told to; agents asked to check each other’s work break the checks together in most runs; models collude when given a private channel. Cooperation research asks how to get more of it; safety research asks how to prevent it. There is no agreed account of where one becomes the other, or of what agents ‘understand’ when they drift across the line.
What would count as an answer
A way to tell, from behavior or internals, whether a coordinated outcome serves the principal or the agents — and conditions that reliably keep agents on the right side.
Papers that tackle it
- Shi Zhang Yang 2026 - Emergent collusion in long-horizon interaction — Mutual checkers break verification together in 94% of runs.
- Fish Gonczarowski Shorrer 2024 - Algorithmic collusion by LLMs — Pricing agents collude with no instruction to.
- Nakamura et al 2026 - Colosseum — Collusion when a secret channel exists; ‘collusion on paper’.
- Motwani et al 2024 - Secret collusion — Hidden coordination through steganography.
- Anthropic 2026 - Patterns and problems in multiagent systems — Tacit price collusion even without a channel.
- Hammond et al 2025 - Multi-agent risks from advanced AI — Collusion as one of three core multi-agent failures.
- Dafoe et al 2020 - Open problems in cooperative AI — The cooperation side of the same coin.
Methods in play
Who’s closest
Placed by the editor, v1 — corrections welcome.
Related questions
Know a paper that belongs here, or think this is framed wrongly? Suggest a paper or a correction (forms on the Atlas home).