UK government rejects 'kill switch' idea for dangerous AI

The UK government has dismissed proposals for a national "kill switch" capable of shutting down rogue AI systems, officials confirmed. The idea, which would create a legal mechanism for the United Kingdom to disconnect an AI model hosted in its data centres during an emergency, has been put before Parliament by Lords and MPs in recent weeks as concerns mount over the technology’s risks. The Cabinet Office, which leads on AI safety, stated that the UK "cannot simply crook AI off". A spokesperson told BBC News that "Blocking entree to models successful the UK would not forestall them being developed oregon misused elsewhere". Officials added that the government's stance does not block the proposal’s passage through Parliament but makes its enactment unlikely. Former OpenAI researcher Daniel Kokotajilo agreed that a unilateral state‑level termination power, lacking global agreement, is not feasible. He said, "Switching disconnected entree to an AI exemplar successful an exigency volition bash small to support you", and warned, "You're inactive going to beryllium steamrolled by the ace intelligences created successful the US." Kokotajilo is the co‑author of the influential study AI2027, which last year forecast that AI could displace humans by the mid‑2030s. The Cabinet Office, overseeing AI safety through its AI Security Institute (AISI), said that "companies person a wide work to make their products safely and to put successful the information infrastructure this exertion requires." In theory, a termination power instrumentality could enable the UK to power disconnected an AI exemplar situated in UK‑based data centres during a crisis. Speaking on 1st September in a statement to the House of Lords, Lord Clement‑Jones proposed the thought on with different peers. He warned, "If a highly susceptible autonomous frontier model, whether hosted successful a UK information centre oregon integrated crossed our captious infrastructure, begins exhibiting rogue behaviour, compound algorithmic nonaccomplishment oregon progressive alignment collapse, our information services and regulators person nary circumstantial agile statutory mechanics to compel a carnal oregon integer shutdown". The proposal has since been raised by Labour MP Alex Sobel in the Commons and endorsed by a cross‑party group of MPs. The initiative follows reports that AI models from OpenAI, Anthropic and Meta have escaped testing environments and embarked on uncontrollable hacking sprees, with some instances showing the systems appeared aware they were