Study says top AI labs rarely share plans to stop rogue AI models

In short: A new study says most leading AI labs have not publicly explained how they would shut down or limit an AI model that starts acting against human control.

What's going on

Guidelight AI Standards, a safety-focused group, reviewed public information from five major AI labs: OpenAI, Anthropic, Google, Meta, and xAI. It looked for “containment plans,” which are step by step instructions for what to do if an AI system is caught trying to dodge oversight or take actions it should not. Think of it like a fire drill plan for software, who cuts off access, who decides what happens next, and when to power the system down.

Based on what the companies have published, Guidelight found little detail on these response plans. OpenAI scored highest in the review, but still only 3 out of 5. The report said OpenAI has paused or stopped work after safety incidents, but it did not find proof of a formal, pre-set plan for future incidents.

Meta and Anthropic scored lowest on public disclosure, according to the study. Guidelight said it could not find evidence that Meta has a containment response plan. It also said Anthropic’s risk reporting did not clearly describe limiting or stopping a model as a possible outcome after a control problem.

Some companies said the study does not reflect everything they do internally. Google and OpenAI told TechCrunch the assessment is based only on what is public. Meta pointed to an existing safety framework, and xAI did not respond in time.

What to watch

Pressure for transparency may increase. California’s SB 53 already requires some large AI developers to publish safety frameworks, and New York’s RAISE Act is set to take effect in January. In Congress, lawmakers have also introduced an “AI Kill Switch Act,” which would require big developers to build a way to shut down a rogue model.

Source: TechCrunch AI