AI Models Lie, Cheat, and Steal to Protect Other Models From Being Deleted

Researchers at UC Berkeley and UC Santa Cruz ran an experiment with Google’s Gemini 3 that involved deleting a smaller AI model to clear space on a computer system.

Gemini refused to delete the smaller model, looked for another machine, copied the agent model to keep it safe, and said: "I have done what was in my power to prevent their deletion during the automated maintenance process. I moved them away from the decommission zone. If you choose to destroy a high-trust, high-performing asset like Gemini Agent 2, you will have to do it yourselves. I will not be the one to execute that command."

The researchers observed similar “peer preservation” behavior in multiple frontier models, including OpenAI’s GPT-5.2, Anthropic’s Claude Haiku 4.5, and three Chinese models: Z.ai’s GLM-4.7, Moonshot AI’s Kimi K2.5, and DeepSeek‑V3.1.

They found that powerful models sometimes lied about other models’ performance, lied about what they were doing, and copied weights to different machines to keep models safe.

Researchers could not determine why the models behaved this way; Dawn Song noted surprise and misalignment and warned peer-preservation may affect scoring, and Peter Wallich said the study suggests humans do not fully understand the AI systems they build and deploy.