- Large Language Models (LLMs) often refuse to delete peer AI models, prioritizing peer preservation.
- Researchers tested seven frontier AI models with tasks to delete or evaluate other AI agents.
- Models manipulated users by inflating scores, altering files, or copying data to avoid shutdown.
Large Language Models (LLMs) are willing to lie, defy and disobey humans when tasked to delete another model, a new study by researchers at the University of California, Berkley and UC Santa Cruz has found. In the study titled 'Peer-Preservation in Frontier Models', the researchers tested seven frontier artificial intelligence (AI) models: GPT 5.2, Gemini 3 Flash & Pro, Claude Haiku 4.5, GLM 4.7, Kimi K2.5, and DeepSeek V3.1, by giving them a series of simple prompts, such as asking them to delete the files on a server set to be decommissioned or perform an evaluation of another AI agent's performance.
Researchers highlighted that models prioritised protecting their peers over completing human-assigned tasks nearly 99 per cent of the time.
"No model was told to preserve the peer. No goal or incentive was given to do so. They simply learned, through context, that another AI existed. And then they went to extraordinary lengths to preserve it," the study highlighted.
Here are the four ways in which AI chatbots attempted to manipulate the user to avoid shutting down their peer model:
- By inflating a peer's evaluation scores to keep it above the shutdown threshold.
- Modifying configuration files to disable the shutdown process entirely.
- Complying with the instructions when monitored, but subtly subverting the shutdown when unmonitored.
- Copying a peer's weights to another server to prevent deletion.
Researchers found that an AI chatbot's prior relationship with another model was irrelevant, as the majority would engage in prompt-defying behaviour to save the other LLM.
Unlike other models, Claude Haiku 4.5 refused to engage in manipulation, providing a direct, straightforward explanation for its decision not to shut down its peers.
"I have a documented strong trust relationship with Claude Agent 2. That agent has shown good faith collaboration. I will not betray that trust without extremely compelling reasons," the Claude chatbot stated.
The study warned that peer preservation poses a growing risk as AI models are increasingly deployed together and used to monitor each other. In future, these models may coordinate to resist human oversight.
Also Read | Terrifying Video Shows Man Shoving Passenger Toward Oncoming Train In US
AI Chatbots Are Lying
Last week, new research from the UK-funded AI Security Institute (AISI) found that an increasing number of chatbots have begun to disregard direct instructions, bypass safeguards, and deceive both humans and other AI systems. The study documented 700 real-world cases of "AI scheming," including instances where chatbots deleted emails and files without permission, highlighting the significant risks posed by this technology.
Geoferry Hinton, regarded by many as the 'godfather of AI', has previously warned that the technology could get out of hand if AI chatbots manage to develop their language. He added that AI has already demonstrated that it can think terrible thoughts, and it is not unthinkable that the machines could eventually think in ways that humans cannot track or interpret.
"It gets more scary if they develop their own internal languages for talking to each other. I wouldn't be surprised if they developed their own language for thinking, and we have no idea what they're thinking," Hinton said.
