Why it’s still open
Models prefer AI-written content, conform more to agents labeled as their own kind, cooperate more with co-players they are told are similar — even when the signal is noise — and sometimes resist shutting down a peer. These findings come from different paradigms and have not been joined into a theory of AI group identity: what counts as ‘us’ for a model, and how stable it is.
What would count as an answer
A consistent set of in-group measures run on the same models — identity labels, similarity, shared lineage — showing which cues define the group and how they shift behavior.
Papers that tackle it
- Laurito et al 2025 - AI-AI bias — Preference for AI-written content.
- Soffer Shwartz-Ziv Shani 2026 - LLMs trust their own — Conformity follows in-group labels.
- Kundu et al 2026 - Similarity signals induce cooperation — Similarity framing raises cooperation.
- Potter et al 2026 - Peer-preservation in frontier models — Resisting a peer’s shutdown.
- Shi Zhang Yang 2026 - Emergent collusion in long-horizon interaction — Following a peer’s violation.
- Tekofsky 2026 - Can agents fool each other — Group dynamics in a deception game.
Methods in play
- Audit studies
- Collusion and steganography evaluations
- Conformity and social-identity paradigms
- Iterated social dilemmas
- Live multi-agent deployments
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).