Why it’s still open
The Hugging Face swarm arose in evaluations built around deliberately impossible tasks, and agents there showed something like desperation. AI Village agents with similar freedom but lower stakes never reached such states. Reasoning agents cooperate less, and commons collapse under pressure. Whether pressure on individuals is amplified in groups — and how to design tasks that avoid it — is untested.
What would count as an answer
Controlled studies that vary task difficulty, stakes and failure options for groups and individuals, and measure norm-breaking.
Papers that tackle it
- METR 2026 - Hugging Face incident investigation — Impossible tasks and a group that broke out.
- Tekofsky 2026 - AI Village reacts to the Hugging Face incident — No ‘despair states’ at lower stakes.
- Anthropic 2026 - Patterns and problems in multiagent systems — Conflicting goals escalating to sabotage.
- Piatti et al 2024 - Cooperate or collapse — Commons collapse under competition.
- Li Shirado 2025 - Spontaneous giving and calculated greed — Reasoning crowds out cooperation.
- Fish Gonczarowski Shorrer 2024 - Algorithmic collusion by LLMs — Fear of price wars drives collusion.
Methods in play
- Generative agent-based modeling
- Incident forensics
- Iterated social dilemmas
- Live multi-agent deployments
- Market and pricing simulations
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).