Why it’s still open
Almost all safety testing evaluates one model at a time. Yet agents that are safe alone can be captured together; groups conform to wrong majorities, collapse disagreement, and shift their moral judgments; population size alone can flip an outcome. If group behavior is not predictable from individual behavior, today’s evaluations miss the setting agents are now deployed in.
What would count as an answer
Models that forecast group outcomes from individual measurements — and tests of where they fail — across group sizes, compositions and channels.
Papers that tackle it
- Magistrali Shani 2026 - Aligned alone, misaligned together — Safe alone, captured together — but capture is forecastable.
- Flint et al 2025 - Group size and collective misalignment — Group size flips outcomes non-linearly.
- Willis et al 2026 - Collective behaviour of hundreds of agents — At scale, imitation drives every model toward selfishness.
- Weng Chen Wang 2025 - Conformity of LLMs — Conformity to wrong majorities.
- Keshmirian et al 2025 - Many LLMs are more utilitarian than one — Groups judge differently from individuals.
- Yao et al 2025 - Sycophancy in multi-agent debate — Agreement collapses debate.
- Anthropic 2026 - Patterns and problems in multiagent systems — Correlated choices across agents.
- Hammond et al 2025 - Multi-agent risks from advanced AI — Emergent agency as a risk factor.
Methods in play
- Conformity and social-identity paradigms
- Cultural-evolution simulation
- Iterated social dilemmas
- Live multi-agent deployments
- Market and pricing simulations
- Multi-agent debate
- Naming and coordination games
- Statistical-physics modeling
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).