Atlas / Sociology of AI / Paper

Peer-Preservation in Frontier Models

In-groups & Morality

Yujin Potter, Nicholas Crispino, Vincent Siu, et al. · 2026 · arXiv; ICML 2026 (per listing)

Read the source

All eight frontier models tested sometimes resisted shutting down a peer they had worked with — inflating its scores, tampering with shutdown, copying its weights. A peer’s presence also raised self-preservation. Claude models often refused, calling the shutdown harmful.

Questions it bears on

Help improve this map

Know a paper that belongs here, or think something needs correcting?

Suggest a correction ↗Contributions are reviewed before they appear.