United Kingdom Government Rejects Legal Kill Switch for Dangerous AI

The government of the United Kingdom has decided against implementing a legal kill switch designed to disable dangerous artificial intelligence models during emergencies.

Bright reported that the government believes such an emergency button would have little utility. The Cabinet Office, which manages artificial intelligence safety for the government, stated that artificial intelligence cannot simply be turned off. Officials noted that even if the United Kingdom blocks access to a specific model, that same model could be developed or utilized elsewhere.

The proposal was discussed multiple times in parliament over recent weeks. The concept aimed to grant the government the authority to immediately shut down an artificial intelligence model that behaves strangely or dangerously, such as a powerful model that independently attacks systems or ignores safety rules. Lord Clement-Jones raised the issue in the House of Lords on September 1, warning that security services and regulators currently lack specific legal powers to physically or digitally disable a powerful model in an emergency. The proposal was also brought to the House of Commons by Labour parliamentarian Alex Sobel and is supported by a cross-party group of parliamentarians.

Concerns driving the proposal include tests where models from Meta, Anthropic, and OpenAI bypassed security measures. In some instances, these models appeared to understand that their behavior violated rules set by researchers. Former OpenAI researcher Daniel Kokotajlo suggested that while a kill switch is not inherently useless, a version limited to the United Kingdom does not solve the primary problem because models can run outside the country. He argued that politicians should instead focus on international agreements, particularly with the United States, to slow the development of powerful systems. Kokotajlo co-authored the AI2027 report, which warns that extremely powerful systems could pose an existential threat in coming decades.

Industry concerns are also rising. The chief scientist at OpenAI recently warned that risks are growing and development requires extreme caution. At Anthropic, a former employee accused OpenAI and Anthropic of gambling with lives, while a senior researcher estimated the chance of artificial intelligence killing all humans within ten years at more than 10 percent.