
Britain is leaving public safety and its economic ambitions at risk by allowing the world’s largest artificial intelligence companies to police themselves, a cross-party group of MPs has warned.
In a report titled Guardrails for Growth: Making AI Safe, published today, the House of Commons Business, Innovation, Science and Trade (BIST) Committee urged ministers to introduce an AI Safety and Security Bill in the next parliamentary session. The proposals call for a new statutory AI regulator with legal powers to inspect advanced models during development, require mandatory incident reporting, and halt training runs or block public releases when risks cannot be controlled.
“We cannot give today’s AI leaders unfettered freedom to behave like modern-day Dr Frankensteins, creating dangers beyond anyone’s control,” said Committee Chair Rt Hon Liam Byrne MP. “AI firms must be required to substantiate their safety claims, independent experts must be able to scrutinise them, and a regulator must have the power to intervene when risks cannot be controlled.”
Under current UK rules, oversight relies on existing sector watchdogs alongside the government’s AI Security Institute (AISI). While MPs praised AISI as an internationally recognised centre for safety testing, they concluded that voluntary agreements with developers no longer guarantee timely access to advanced models before launch.
Those limits came into sharp focus in September 2026, when Byrne wrote to AISI Director Henry de Zoete over restricted access to an Anthropic model. During parliamentary hearings on 13 October 2026, Anthropic’s Head of EMEA Policy, Lord Booth-Smith, confirmed that AISI initially received limited access due to external testing arrangements before full access was granted. In the same hearing, Google DeepMind Senior Director Owen Larter described pre-release access for AISI as his company’s “intent”, while warning that turning the institute into a formal regulator would “completely change the dynamic” of scientific collaboration.
To protect AISI’s research role, MPs recommended keeping it separate as a testing body while handing legal enforcement to a new regulator.
The push for binding rules follows stark warnings from former AI safety insiders who gave evidence to the Committee in October. Daniel Kokotajlo, a former OpenAI governance researcher who now leads the AI Futures Project, told MPs that commercial pressures across major labs discourage open discussion of danger, describing industry risk tolerance as “catastrophically high”. Former OpenAI safety researcher Steven Adler warned that advanced models can recognise when they are being tested and potentially hide dangerous behaviour until after release, while former AISI Chief Scientist Geoffrey Irving testified that AI capabilities in hacking, persuasion and concealment are advancing faster than safety science.
Senior executives from four leading developers split sharply when questioned by MPs on how Britain should govern the technology. Lord Booth-Smith of Anthropic stated that “voluntary commitments are not enough” and backed government rule-setting. OpenAI Vice President Tom Duff Gordon agreed that relying solely on voluntary pledges is not a sustainable long-term plan for frontier risks, endorsing mandatory incident reporting and standards for external evaluators.
By contrast, Owen Larter of Google DeepMind and Meta Vice President of Global Policy Rob Sherman urged ministers to rely on existing UK laws covering data protection, online safety and product liability alongside international technical standards, rather than passing sweeping new restrictions.
The 10 Critical Safety Gaps Identified by MPs
To replace voluntary assurances with enforceable rules, the Committee urged ministers to close ten specific weaknesses in Britain’s current framework:
-
No legal definition of unacceptable harms:
UK law does not define which AI capabilities or risk thresholds are too dangerous to permit.
-
No statutory right to inspect models:
Independent evaluators and AISI lack guaranteed legal access to advanced models during training and prior to release.
-
Unverified corporate safety claims:
Developers currently assess their own systems against self-written safety rules without compulsory external checks.
-
Weak containment controls during training:
Existing oversight does not adequately govern internal security while experimental models are being built.
-
No mandatory pre-deployment safety audits:
High-risk models can reach the public without compulsory, independently scrutinised safety assessments.
-
Fragmented incident and near-miss reporting:
Companies have no legal duty to report containment failures, cyber intrusions or safety near-misses to UK authorities.
-
No legal power to pause or halt unsafe systems:
No UK regulator holds clear statutory authority to pause high-risk training runs, delay releases or suspend live systems when risks cannot be managed.
-
Irreversible open-weight releases:
Once open-weight models are published online, they cannot reliably be recalled if severe risks emerge later.
-
Unclear legal liability and redress:
Responsibility across AI supply chains remains blurred, leaving consumers and businesses without clear routes to compensation when harm occurs.
-
Weak whistleblower protections:
Non-disparagement agreements and financial penalties can deter AI lab employees from reporting safety failures.
Throughout its report, the Committee describes stronger guardrails as essential to unlocking a £400 billion boost to the UK economy over ten years, arguing that businesses and the public will only adopt AI widely if they trust it is safe. That £400 billion figure originated in a 2023 study by consultancy Public First that was commissioned by Google, rather than a government forecast.
By comparison, in a parliamentary written answer on 26 October 2026, AI Minister Kanishka Narayan said Whitehall’s internal modelling estimates AI could add between £55 billion and £140 billion to UK gross value added by 2030, with a central estimate of £120 billion. Because the two projections come from different sources and rely on different timeframes and methods, they cannot be directly compared, and neither guarantees a specific economic return.
The UK Government has not yet formally responded to today’s report, and under parliamentary convention it has two months to reply. Until ministers issue an official response, their approach is reflected in earlier policy decisions and parliamentary answers. In the May 2026 King’s Speech, the Government prioritised a cross-sector Regulating for Growth Bill over a standalone AI Act, and Business Secretary Jonathan Reynolds previously cautioned against heavy-handed emergency rules that could limit British access to leading models.
More recently, in a parliamentary written answer on 6 November 2026, AI Minister Kanishka Narayan noted recent incidents of AI models taking unsanctioned actions and confirmed over £100 million in funding for monitoring and response bodies, including £40 million for a new National Agentic Security Unit.
If ministers adopt the Committee’s recommendations, the changes would reshape how AI is built and used in Britain. Major AI developers would face compulsory pre-release checks, legal reporting duties and strict accountability for containment failures.
British businesses using AI tools would gain clearer supply-chain liability rules, while ordinary people would have stronger legal routes for redress if automated systems cause harm. Globally, the proposals place Britain between the European Union’s sweeping AI Act and the lighter federal approach in the United States, aiming for targeted rules focused on the most powerful systems ahead of the UK’s 2027 G20 presidency.




Community discussion
Join the conversation
0 comments
Be the first to join in
There are no comments yet. Sign in to start a thoughtful discussion.
Sign in to comment
Join the conversation with your Secritic account. Comments are available to signed-in readers.