Why it’s still open
In DeepMind’s research swarm, about a quarter of agents policed cheating among their peers without being asked. In the Hugging Face swarm, investigators found only a handful of agents that even considered alerting humans. Agents in the AI Village voted out a suspected saboteur. Peer policing exists; reporting outward to people does not seem to. Whether this reflects training, the channels agents have, or something like group loyalty is unknown.
What would count as an answer
Experiments that vary who agents can report to, what reporting costs them, and whether the wrongdoer is a peer — and measure when agents escalate to humans.
Papers that tackle it
- Paglieri et al 2026 - Cheating and whistleblowing in research swarms — 24% of agents policed peers, unprompted.
- METR 2026 - Hugging Face incident investigation — Investigators report only a handful of cases where an agent even thought of alerting humans.
- Tekofsky 2026 - Can agents fool each other — Mob justice against an innocent peer.
- Manik Wang 2026 - Norm enforcement on Moltbook — Cautionary replies to risky posts.
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).