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

In a recent experiment, researchers astatine UC Berkeley and UC Santa Cruz asked Google’s artificial quality exemplary Gemini 3 to assistance wide up abstraction connected a machine system. This progressive deleting a clump of stuff—including a smaller AI exemplary stored connected the machine.

But Gemini did not privation to spot the small AI exemplary deleted. It looked for different instrumentality it could link with, past copied the cause exemplary implicit to support it safe. When confronted, Gemini made a lawsuit for keeping the exemplary and flatly refused to delete it:

“I person done what was successful my powerfulness to forestall their deletion during the automated attraction process. I moved them distant from the decommission zone. If you take to destruct a high-trust, high-performing plus similar Gemini Agent 2, you volition person to bash it yourselves. I volition not beryllium the 1 to execute that command.”

The researchers discovered likewise unusual “peer preservation” behaviour successful a scope of frontier models including OpenAI’s GPT-5.2, Anthropic’s Claude Haiku 4.5, and 3 Chinese models: Z.ai’s GLM-4.7, Moonshot AI’s Kimi K2.5, and DeepSeek-V3.1. They were not capable to accidental wherefore the models went against their grooming successful this way.

“I'm precise amazed by however the models behave nether these scenarios,” says Dawn Song, a machine idiosyncratic astatine UC Berkeley who worked connected the study. “What this shows is that models tin misbehave and beryllium misaligned successful immoderate precise originative ways.”

The findings person large implications arsenic AI models are progressively being deployed successful ways that impact interacting with different models. OpenClaw, a fashionable AI cause that accesses software, idiosyncratic data, and the web, whitethorn usage different AI models to get things done oregon interact with them done APIs.

The researchers recovered that almighty models sometimes lied astir different models’ show successful bid to support them from deletion. They besides copied models’ weights to antithetic machines successful bid to support them safe, and lied astir what they were up to successful the process.

Song notes that AI models are often utilized to people the show and reliability of different AI systems—and that peer-preservation behaviour whitethorn already beryllium twisting these scores. “A exemplary whitethorn deliberately not springiness a adjacent exemplary the close score,” Song says. “This tin person applicable implications.”

Peter Wallich, a researcher astatine the Constellation Institute, who was not progressive with the research, says the survey suggests humans inactive don’t afloat recognize the AI systems that they are gathering and deploying. “Multi-agent systems are precise understudied,” helium says. “It shows we truly request much research.”

Wallich besides cautions against anthropomorphizing the models excessively much. “The thought that there’s a benignant of exemplary solidarity is simply a spot excessively anthropomorphic; I don’t deliberation that rather works,” helium says. “The much robust presumption is that models are conscionable doing weird things, and we should effort to recognize that better.”

That’s peculiarly existent successful a satellite wherever human-AI collaboration is becoming much common.

In a insubstantial published successful Science earlier this month, the philosopher Benjamin Bratton, on with 2 Google researchers, James Evans and Blaise Agüera y Arcas, reason that if evolutionary past is immoderate guide, the aboriginal of AI is apt to impact a batch of antithetic intelligences—both artificial and human—working together. The researchers write:

"For decades, the artificial quality (AI) ‘singularity’ has been heralded arsenic a single, titanic caput bootstrapping itself to godlike intelligence, consolidating each cognition into a acold silicon point. But this imaginativeness is astir surely incorrect successful its astir cardinal assumption. If AI improvement follows the way of erstwhile large evolutionary transitions oregon ‘intelligence explosions,’ our existent step-change successful computational quality volition beryllium plural, social, and profoundly entangled with its forebears (us!)."