Summary
- The Business, Innovation, Science and Trade Committee has identified ten gaps in the UK's approach to advanced AI safety and security.
- MPs want an AI Safety and Security Bill and a statutory regulator, separate from the AI Security Institute, with powers to restrict unsafe development and deployment.
- The committee links stronger safeguards to a possible £400bn economic opportunity over ten years, but that figure is a scenario rather than guaranteed growth.
MPs are calling for a new statutory regulator with powers to restrict dangerous artificial intelligence systems, arguing that Britain’s reliance on voluntary safety commitments leaves critical decisions about advanced technology in the hands of the companies developing it.
The House of Commons Business, Innovation, Science and Trade Committee published its report, Guardrails for growth: AI safety and security, on 18 November. It identifies ten gaps in the UK’s oversight arrangements and recommends an AI Safety and Security Bill in the next parliamentary session.
Its proposals include a regulator able to require independent verification of safety claims, impose additional safeguards and restrict the development or deployment of systems presenting unacceptable risks. The committee also wants clearer rules on incident reporting, legal responsibility and protections for people who disclose serious safety concerns.
These recommendations have not been adopted by the government. Parliament would need to approve legislation before the proposed authority could exercise the powers described in the report.
Research does not provide enforcement powers
Britain has invested in specialist technical research through the AI Security Institute, which examines the capabilities and risks of increasingly advanced models. Its evaluations can identify concerning behaviour, investigate safeguards and provide evidence about emerging threats.
However, the committee found that the institute’s research function does not give it authority to prohibit deployment or compel developers to change their systems. Companies retain considerable discretion over whether and how they act on safety findings, including decisions to delay a release.
Under the committee’s preferred structure, technical research would continue through the institute, while a separate statutory regulator would exercise enforcement powers. Researchers could provide evidence about model behaviour without being responsible for every legal judgement concerning acceptable risk.
The distinction would introduce external accountability into decisions currently made largely through company safety frameworks. An independent authority could intervene when a developer’s proposed safeguards were insufficient, rather than depending on the same organisation to evaluate and approve its own response.
Such powers would require clear legal thresholds. Advanced models differ in their capabilities, access to tools and potential applications, making it difficult to determine unacceptable risk through a single technical measurement.
The committee recommends defining circumstances in which the government could delay, restrict or prohibit training, testing or deployment in the UK. Those interventions would need to be proportionate and subject to appropriate accountability, particularly where technical evidence remains uncertain.
Independent testing would face legal requirements
Safety evaluations conducted by developers can identify weaknesses before models become available, but the committee heard concerns about relying exclusively on assessments performed by organisations with commercial interests in releasing a product.
Independent evaluators can challenge a developer’s assumptions, examine how safeguards perform and investigate behaviours that internal testing may have overlooked. To do so effectively, they need meaningful access to relevant systems, documentation and technical information.
That access raises commercial and security questions. Model developers may hold sensitive research data and proprietary information, while unrestricted distribution of advanced models could create risks of its own. A statutory framework would need to define who can conduct assessments, what information they may obtain and how it must be protected.
The committee also argues that important disagreements between developers and independent assessors should be documented rather than resolved through private arrangements that leave regulators without a clear account of the evidence.
Even thorough evaluation before release cannot establish how a model will behave under every subsequent condition. Developers may connect systems to new applications or tools, while users may encounter circumstances not represented in the original tests.
Continuing monitoring would therefore form part of the proposed oversight model. Its effectiveness would depend on whether developers and deployers can identify relevant failures and provide evidence that corrective action has been taken.
Reporting failures would become compulsory
The committee has identified inconsistent reporting of serious AI incidents as another weakness. Without common reporting duties, public authorities may struggle to identify whether related failures are occurring across several systems or organisations.
A mandatory framework could require developers and other relevant businesses to report incidents meeting defined thresholds. The legislation would need to establish which events qualify, who is responsible for notification and how sensitive technical details should be handled.
Whistleblower protections would provide another route for significant concerns to reach regulators. Researchers and employees may become aware of unsafe practices or unresolved problems, but disclosing that information can carry professional risks, particularly where contractual confidentiality obligations apply.
Protected reporting procedures would not make every allegation accurate. They would give the relevant authority a means of receiving and assessing information without requiring individuals to publish confidential technical material themselves.
The committee also wants clearer liability arrangements and provisions addressing harmful misuse. These would need to distinguish the responsibilities of model developers from those of organisations deploying AI within particular products or services.
A £400bn opportunity remains conditional
The report places its proposed safeguards alongside the government’s ambitions for AI-related economic growth. Drawing on OECD productivity estimates, the committee calculates that AI could leave the British economy up to £400 billion larger after ten years.
That figure is a possible economic outcome, not an amount of investment already secured or a guarantee that AI adoption will deliver the projected gains. Productivity improvements depend on businesses introducing useful technology, adapting processes and achieving benefits greater than their implementation costs.
Stronger safety arrangements could support adoption where uncertainty about reliability, legal exposure and organisational responsibility has delayed purchasing decisions. However, regulation also introduces obligations that must be designed carefully if smaller developers are to compete with larger companies.
The committee identifies AI assurance as a potential British technology industry, building on existing expertise in security research, model evaluation and technical testing. Companies supplying those services could benefit from clearer standards, although the commercial opportunity will depend on the legislation and customer demand that eventually emerge.
Its report also opens a discussion about international rules during the UK’s 2027 G20 presidency. Advanced models are developed and distributed across borders, making national oversight more difficult where services can be accessed internationally.
The committee is seeking evidence about a possible global convention on AI safety and security. No such agreement has been concluded, and international negotiations would be separate from establishing enforceable powers within Britain.
The immediate decision rests with ministers, who must determine whether to introduce the recommended legislation and what authority any new regulator should receive. Until then, the AI Security Institute’s technical research continues without the statutory intervention powers the committee believes are necessary.












