The UK government has rejected proposals to establish a legal “kill switch” that would allow authorities to shut down dangerous artificial intelligence systems or data centres during a major AI-related emergency.
The proposal was put forward as an amendment to the Cyber Security and Resilience Bill, currently being considered by Parliament. Supporters of the amendment argued that the government should have explicit legal powers to intervene if an AI system caused a serious security or operational emergency.
The proposed powers would have allowed the Secretary of State to direct the shutdown of AI systems deployed on a substantial scale or, in certain circumstances, data centres hosting them. The amendment defined an AI security or operational emergency as a situation involving a compromise of relevant network and information systems that was caused or contributed to by an AI system and presented a “catastrophic risk.” This included large-scale disruption to critical infrastructure or essential services, significant damage to national security or defence capabilities, or severe, large-scale harm to human life.
Liberal Democrat peer Lord Tim Clement-Jones was among those supporting the proposal in the House of Lords. The amendment also received support from other parliamentarians, including Baroness Kidron, Baroness Harding of Winscombe and Lord Hunt of Kings Heath. It included safeguards requiring the government to report to Parliament after using the proposed emergency powers.
However, the government has opposed creating a specific statutory AI “kill switch”. A Cabinet Office spokesperson said that the UK “cannot simply turn AI off” and argued that preventing access to AI models in Britain would not stop those models from being developed or misused elsewhere. The government also said that AI companies have responsibility for developing their products safely and investing in the security infrastructure needed to support the technology.
The government has also pointed to powers that already exist within the Cyber Security and Resilience Bill. According to the government’s position, existing direction-making powers could allow authorities to require regulated organisations, including data-centre operators, to take or stop particular actions when their systems create a qualifying national-security risk.
Baroness Lloyd of Effra, the Department for Science, Innovation and Technology minister handling the legislation in the Lords, argued that directing an operator to stop using a particular AI model could be more proportionate than shutting down an entire data centre. She noted that AI systems can be distributed across several data centres and jurisdictions, meaning that shutting down individual facilities would not necessarily stop an AI system from operating elsewhere.
The debate comes as governments and researchers examine how to respond to increasingly capable AI systems. The UK’s AI Security Institute has been involved in evaluating emerging AI risks, while the government has said it will continue taking a science-led approach to understanding and preparing for those risks. Ministers have also indicated that they are considering whether additional targeted interventions may be necessary.
The proposed amendment also sought to introduce technical requirements for AI providers and data-centre operators. These included maintaining secure communication channels with the government, having infrastructure capable of complying with emergency directions, conducting regular emergency exercises and carrying out post-incident reviews before operations resumed.
The proposal was not adopted. The relevant Lords amendment was listed as “not moved,” meaning the House was not asked to take a decision on it. A separate Commons amendment concerning similar last-resort powers was also not called for a vote.
The government’s decision leaves the UK without the specific statutory AI shutdown mechanism sought by the amendment. Instead, its current approach relies on existing regulatory powers, responsibilities placed on AI companies and targeted government intervention, while continuing to assess the security risks associated with increasingly capable AI systems.
