What Is UK AI Law? AI Safety Bill Explained

There is no dedicated AI law in the United Kingdom yet. Artificial intelligence in Britain is currently governed by a patchwork of older laws that were not written with AI in mind, and by regulators working within their existing powers. That position is now under serious challenge from Parliament itself. In the space of two months, two parliamentary committees have concluded that the current framework does not protect people well enough, and both have called for a dedicated AI law and a new independent regulator.
This explainer sets out what UK AI law is as it stands on 20 November 2026, what the two committee reports actually recommend, what an AI Safety and Security Bill would do, and what happens next. Nothing here is law yet: the proposals are recommendations to the Government, and ministers have not committed to introducing a Bill.
Does the UK have an AI law?
No. The UK has no statute that deals specifically with artificial intelligence. Instead, AI systems are governed indirectly through laws that apply to particular harms or sectors. Data protection rules, including the UK GDPR, apply when AI processes personal data. Equality law applies when an AI system discriminates. Product safety law, consumer protection law and sector rules in areas such as finance and healthcare also bite, depending on how and where a system is used.
The Government has so far preferred what it calls a regulator-led approach: existing regulators, such as the Information Commissioner's Office, the Competition and Markets Authority and the financial regulators, apply their own rules to AI within their sectors, guided by cross-cutting principles rather than a single AI statute. Supporters of that approach say it is flexible and avoids burdening a fast-moving technology with rigid rules. Critics, including the two parliamentary committees, say it leaves gaps, because no single body is responsible for AI as a whole and because the existing laws were designed for harms that look very different from those AI can cause.
The Joint Committee on Human Rights put the point bluntly in September 2026, concluding that nowhere in the world, including the UK, has a legislative and regulatory approach to AI that is fit for purpose. Its chair, Alex Sobel MP, said the UK's current framework does not offer adequate protection against the human rights risks that AI systems pose.
What did MPs recommend in November 2026?
The most recent intervention came from the House of Commons Business, Innovation, Science and Trade Committee. Its report, Guardrails for growth: AI safety and security, was published on 18 November 2026 as the Eleventh Report of Session 2026-27, HC 689. The committee took evidence from AI researchers, former industry employees and senior representatives of OpenAI, Google DeepMind, Anthropic and Meta, and it identified ten critical gaps in the UK's framework for AI safety.
The committee calls for an AI Safety and Security Bill in the next parliamentary session, as part of a clear legislative roadmap. It says the roadmap should combine targeted action, duties to prevent harm, a clear liability regime, criminalising the misuse of AI for harm, and an international convention to enshrine global rules. It also wants a statutory AI regulator, separate from the AI Security Institute, but drawing on the institute's expertise, with secure multi-year funding.
The report is the first in a series from the committee's long-term inquiry into AI adoption, safety and governance. It is a committee's view, not government policy, and no Bill exists yet. The Government must now decide whether to accept the recommendations.
What are the ten gaps in UK AI safety rules?
The ten gaps identified by the committee cover the whole lifecycle of a powerful AI model, from training to deployment and beyond. The first is the absence of agreed definitions of which harms are unacceptable and where the thresholds for acceptable risk should sit. Without those definitions, the committee argues, nobody can judge whether a system is safe enough to release.
The second and third gaps concern access and verification. The committee found that the UK lacks statutory powers to require access to AI models for safety testing, and that there is no independent verification of developers' own safety claims. In practice, that means the public is largely asked to take the word of the companies building the most powerful systems.
The fourth and fifth gaps concern what happens before a system is released. The committee found insufficient controls on internal deployment, training and containment, and an absence of safety evidence before release. It wants mandatory, independently scrutinised safety assessments before designated high-risk systems are deployed.
The remaining gaps concern what happens after release and when things go wrong. There is no mandatory reporting of incidents and near misses. There are no powers to restrict or halt unsafe development or deployment. There are insufficient safeguards for models whose release cannot be reversed, such as open-weight models published in full. Liability arrangements and routes to redress for people harmed by AI are inadequate. And whistleblowers inside AI companies do not have enough protection to raise concerns safely.
The committee's chair, Liam Byrne MP, said the hearings had exposed weaknesses in independent testing, incident reporting and powers to intervene. The committee argues that closing the gaps matters for growth as well as safety: drawing on OECD productivity estimates, it calculates that AI could leave Britain's economy up to 400 billion pounds larger after ten years, while the UK's recognised expertise in AI testing and evaluation could create a new export industry in AI assurance.
What would the AI Safety and Security Bill do?
The proposed Bill would be the first step, in the committee's words, towards a comprehensive framework for AI safety and security. It has not been drafted, so its exact contents are not known, but the report sketches its architecture. It would put duties to prevent harm on a statutory footing, set out a clear liability regime so that it is clear who is responsible when an AI system causes harm, and criminalise the misuse of AI for harm.
Crucially, it would create a statutory regulator with powers the current system lacks. The committee wants the regulator to be able to delay, restrict or prohibit the deployment, training or testing of high-risk models. That power to stop an unsafe system, which the committee calls the power to stop, is the single most consequential proposal in the report, because no UK body currently has it.
The regulator would also be able to require safety evidence before release, demand incident reports, and order remedies for individuals without requiring them to incur prohibitive costs. The Joint Committee on Human Rights made a similar proposal in September, calling for a single independent statutory AI oversight body that could test and evaluate AI systems, including a prior approval regime for high-risk uses, prohibit systems from being launched or deployed if unacceptable risks were identified, order their withdrawal from the market, publish mandatory codes of practice and sanction noncompliant actors.
What is the AI Security Institute, and would it become the regulator?
The AI Security Institute, known as AISI, is the UK body that tests advanced AI models for safety. It was established in November 2023 as the AI Safety Institute and renamed in February 2025. It currently operates on a voluntary basis with no statutory powers, which means it relies on AI developers agreeing to submit their models for evaluation.
Both committees want that changed, but they differ on the structure. The Joint Committee on Human Rights recommends putting AISI on a statutory footing, with developers of powerful AI models required to submit new models to it for review, evaluation and testing, and suggests ministers consider merging AISI's functions with those of the new regulator to avoid duplication. The Business Committee takes a different view: it says the statutory regulator should be a separate body from AISI, but one that draws on AISI's technical expertise.
Either way, the direction of travel is the same: from a voluntary testing body to a statutory system in which safety evaluation is a legal requirement rather than a favour granted by developers.
What did the human rights committee say?
The Joint Committee on Human Rights published its report, Human Rights and the Regulation of AI, on 14 September 2026. It followed an inquiry launched in July 2025, with ten oral evidence sessions and 74 written submissions. The report's main conclusion is that the UK's existing legal framework for AI is fragmented and difficult to navigate, with most legislation designed for specific harms and sectors, leaving gaps in human rights protection.
Among the rights the committee found at risk were privacy, freedom of expression, equality and non-discrimination, and the right to a fair trial. It noted that current law applies to AI mainly at the point of deployment, placing the burden of responsibility on the organisations that deploy AI systems rather than on the developers that design and build them. Its recommendations include a dedicated AI Bill, a statutory oversight body, risk-based regulation proportionate to the danger posed, and a timetable for ratifying the Council of Europe Framework Convention on Artificial Intelligence, Human Rights, Democracy and the Rule of Law, which the UK has signed but not yet ratified.
The Government was expected to respond to the human rights committee within two months of publication, by 14 November 2026. That deadline has now passed, and the Business Committee's report adds further pressure for a substantive reply.
How does the UK compare with the EU?
The UK currently has no equivalent to the EU AI Act, the European Union's comprehensive AI law. The Joint Committee on Human Rights noted that the EU framework reaches AI developers as well as deployers and users, and signalled that the UK could move towards a structure more closely aligned with the EU. The Business Committee also calls for an international convention on AI safety and security, arguing that global rules are needed because the most powerful models are developed and deployed across borders.
That comparison matters for British businesses. A UK company selling AI products into the EU already has to comply with the EU AI Act, whatever the UK does. A UK law that diverges sharply from the EU framework could therefore create two sets of compliance duties, while a law that aligns with it could reduce friction. The committees argue that regulation should be proportionate and should not place unjustified burdens on businesses, especially those using the lowest-risk AI systems, and should not overly restrict innovation and growth.
What happens next?
The immediate next step is the Government's response. Ministers must decide whether to accept the call for an AI Safety and Security Bill in the next parliamentary session, whether to create a statutory regulator, and whether to give the AI Security Institute statutory powers. The Business Committee has asked ministers to set out the legislative sequence they envisage, and it will continue its inquiry with further reports and a call for evidence on a global convention on AI safety and security.
For the public, nothing changes immediately. There is no new law to comply with and no new regulator to complain to. But the direction of travel is now clear: two separate parliamentary committees, from different angles, have reached the same conclusion that the UK's voluntary, regulator-led approach is no longer enough. The question for 2027 is whether the Government agrees, and what kind of AI law it is prepared to introduce.
What does it mean for businesses and the public?
For businesses, the reports signal that the era of purely voluntary AI safety commitments may be coming to an end. Companies building or deploying powerful AI systems should expect mandatory safety assessments, incident reporting and clearer liability rules, and should start keeping the kind of safety documentation a regulator would want to see. For the public, the proposals would mean a named body to complain to, powers to stop dangerous systems before they cause harm, and remedies that do not require expensive legal action.
Both committees stress that AI also brings large benefits, from scientific discovery to productivity growth, and that regulation should be proportionate to risk. The debate is not about whether AI should be developed, but about who gets to judge whether it is safe, and what happens when the answer is no.