On 18 November the House of Commons Business, Innovation, Science and Trade Committee published a report, Guardrails for Growth: Making AI Safe, and the line that travelled was the one about pulling the plug. The committee wants a regulator with the statutory power to step in while an artificial-intelligence model is still being built: to halt high-risk training, to block a deployment, or to suspend a service already running if its risks cannot be contained. Its chair, Liam Byrne, put it in the register that guarantees a headline. We cannot give today's AI leaders unfettered freedom to behave like modern-day Dr Frankensteins, he said, and added that a regulator must have the power to intervene when risks cannot be controlled.

I want to take the report seriously, because it deserves it, and because the instinct behind it is sounder than the headline makes it look. But I also want to be honest about where the instinct meets the engineering, because the single most quotable power in the report is the one least likely to work as advertised, and the committee, to its credit, half-knows this. Byrne said it himself, in the same breath: once some models are released, they cannot reliably be recalled. That sentence is the whole problem in miniature, and it is worth slowing down over, because an enormous amount of AI policy is being written as though it were not true.

What the report actually asks for

Strip away the Frankenstein and the report is a measured document, and a more interesting one than the coverage suggested. It identifies what it calls ten gaps in Britain's current approach, and the current approach is, essentially, trust. The four firms the committee questioned — OpenAI, Google DeepMind, Anthropic and Meta — all told MPs they would not release a model they considered unsafe, and I have no reason to doubt that any of them meant it. The difficulty, as the committee observed, is that they do not use the same yardstick for "unsafe", that no two of them draw the line in the same place, and that nothing in British law requires any of them to obtain an independent opinion before shipping. The safety of the most capable systems on earth currently rests on the private judgement of the companies that profit from shipping them, each marking its own homework to a standard it sets itself.

Against that, the committee proposes a handful of concrete things. It wants an AI Safety and Security Bill in the next parliamentary session. It wants the AI Security Institute — the government body that has done the actual technical work of testing frontier models, and done it well — kept in that testing role but reinforced, with a separate regulator holding the enforcement powers, so that the people who evaluate are not the people who police. It wants developers of designated high-risk systems to produce evidence for their safety claims rather than assert them, and to let independent researchers examine the systems, with any significant change to a model's capabilities or safeguards triggering fresh scrutiny. It wants mandatory reporting of AI incidents and near misses. And it wants stronger protection for whistleblowers, after former employees told the committee that restrictive contracts and financial disincentives discourage staff from raising concerns.

That is a programme about scrutiny, evidence and reporting, and almost all of it is sound. Notice that the enforceable, sensible core of it happens before release: independent testing, evidence for safety claims, researcher access, scrutiny when capabilities change. The dramatic power — switching off a model already in the world — sits at the end of a chain whose earlier links are the ones that actually bear weight.

Why you cannot recall a model

The headline power runs into a hard technical fact, and it is the fact Byrne himself acknowledged. A regulator can halt a training run it knows about, inside a jurisdiction it controls, operated by a company with a British legal presence and a reputation to protect. That part is real, and for the frontier labs it may even be sufficient, because there are only a handful of them and they are all, for now, large organisations with lawyers and shareholders. But "block deployment" and "suspend the service" assume the model stays where the regulator can reach it, and a great many models do not.

The moment a model's weights — the numbers that are the model — are released openly, as Meta's Llama family and a long line of others have been, they are copied, mirrored, fine-tuned and re-hosted beyond anyone's recall. There is no switch. Meta was in the room for this inquiry, and open-weight release is its stated strategy, so the committee cannot have missed the point. You can prohibit a British company from releasing such weights in the first place, which is a pre-release control and a real one. What you cannot do is un-release them afterwards, any more than you can recall a published book by buying back every copy. A regulator who announces, after the fact, that a widely mirrored open model is now prohibited has issued a press release, not an intervention.

This is why the framing matters, and why I keep returning to the word "before". The committee's instinct — that protection must begin while systems are being built and tested, as Byrne put it — is exactly right, and it is right precisely because the after is so weak. The useful version of this regime is the one that makes release conditional on independent evidence. The version that promises to reach into the world and switch off a model that is already loose is selling a capability that physics does not grant. I would rather the political energy went into the part that works.

"Guardrails for growth" is the right name

It would be easy, for someone in my trade, to treat this as the familiar contest between safety and innovation, and to pick the safety side because it is the one my profession is paid to care about. I think that framing is wrong, and the report's title is a quiet argument against it. The committee did not call this Brakes on Growth. It calculated, using OECD productivity estimates, that AI could leave the British economy up to £400 billion larger over the next decade, and it argued — this is the part worth dwelling on — that stronger safeguards would encourage adoption, particularly among smaller firms, rather than retard it.

That is not a throwaway line, and it matches what I see in the field. The organisations holding back from adopting AI in any serious way are overwhelmingly not the ones worried about existential risk. They are ordinary businesses that cannot tell whether a given tool is safe to put near their data, their customers or their regulated processes, and who have no way to acquire that assurance and no appetite to carry the liability of guessing. For them, an absence of regulation is not freedom; it is fog. A credible, independent testing regime is the thing that would let a mid-sized company adopt a model with the same shrug of confidence it brings to buying a piece of certified electrical equipment. The committee has grasped something that the louder parts of the debate miss: for the mass market, trust is an enabler of adoption, not a tax on it. We are, to borrow a phrase I am fond of, better off selling good mornings than nightmares, and a safety regime that unlocks the cautious majority is a good-mornings policy.

The industry is not of one mind, and that tells you something

The most revealing context for this report is not in the report at all. It is that the AI companies themselves are now openly split on whether they want to be regulated, and the split is instructive.

In September, OpenAI's head of policy for Europe, the Middle East and Africa told reporters that the United Kingdom should act now to set up a stronger set of safety rules, and that the company was in a position to support durable, mandatory, capability-based requirements — the first time it had called for binding British rules on frontier AI. Its proposed scope was narrow and self-serving in a way the company did not trouble to hide: national-security and cyber risks only, the handful of frontier labs only, with explicit reassurance that startups building less powerful systems should be left alone. In the same period Anthropic, according to the Financial Times, skipped the AI Security Institute's pre-release testing for one of its Mythos models, which is precisely the voluntary-cooperation gap the committee is now trying to close. And the government, through the Business Secretary, warned against being hyperbolic about AI risk and cautioned that rules cutting Britain off from frontier developments would obviously make us less safe.

Hold those three together and the picture is honest. One large lab wants binding rules, scoped tightly enough to bless its own position and raise the drawbridge behind it. Another declines a voluntary test when it suits its schedule. The government fears being left behind more than it fears the models. When the incumbents start asking to be regulated, a careful observer does not simply applaud; they ask who the proposed rules would entrench, because a safety regime written around the capabilities and processes of today's three or four giants is also, inescapably, a moat around those three or four giants. The committee's insistence that the AI Security Institute keep testing while a separate body enforces is, read in this light, rather wise. It keeps the evaluator independent of both the industry and the politics, which is the only arrangement under which "independent testing" means anything.

The precedents hold, up to a point

The committee reaches, as these arguments always do, for aviation, pharmaceuticals and nuclear power: industries where we do not take the manufacturer's word that the product is safe, but require independent evidence before it reaches the public, and where a regulator can ground a fleet or pull a drug. The analogy is useful and the committee is right to use it. We did not decide that aircraft were too economically important to regulate; we decided that their importance was the reason to.

But the analogy breaks in one place that matters, and it is the place we have already been. A grounded aircraft stays grounded; a withdrawn drug leaves the shelves. Those products are physical, scarce and controlled at the point of manufacture, and the regulator's writ runs because the thing cannot be trivially copied. A model's weights can be copied perfectly, instantly and without limit. The right lesson from aviation and pharma is therefore the pre-market one — nothing reaches the public without independent evidence — and not the recall one, because the recall depends on a scarcity that software does not have. A regime built on the first lesson is enforceable. A regime that leans on the second will work for the compliant British frontier lab and fail for everyone and everything else, which is to say it will work exactly where voluntary cooperation already mostly works, and fail exactly where it is needed.

What this means if you are not a frontier lab

Almost no one reading this builds frontier models, and it would be reasonable to ask why any of it matters to an ordinary business. It matters because regulation of this kind never stays at the top of the stack. The committee's proposals are aimed at the handful of companies training the largest systems, but the evidence requirements, the incident-reporting duties and the testing expectations will flow downhill through procurement and contract long before they flow through statute. Within a year or two of any such Bill, the question can you demonstrate that the AI in this product was independently evaluated will appear in enterprise purchasing questionnaires, because that is how every upstream safety regime has ever reached the businesses below it. I have watched the Cyber Resilience Act and DORA do exactly this to companies that were nominally out of scope, and AI will be no different.

So the practical posture for an ordinary organisation is neither to campaign on AI safety nor to ignore it, but to get ahead of the paperwork it will shortly be asked for. Know which AI systems you actually use, including the ones embedded in software you bought for other reasons, because you cannot govern what you have not enumerated. Keep the evidence you already have about them — model cards, vendor safety documentation, the results of any independent testing — in one place rather than scattered across inboxes. Decide, now, who in the organisation owns the question of whether a given AI tool is fit to put near your data and your customers, because when the questionnaire arrives it will need an owner. None of that waits on a Bill, none of it is expensive, and all of it is simply good practice that the regulation, if it comes, will later make mandatory. I have written before about the usage policy you actually need and about how the regulators are already paying attention to frontier AI in critical infrastructure; this report is the same tide reaching the next part of the shore.

For boards

Three questions, none of which needs a technical answer.

Do we know every AI system in use across our organisation, including the ones embedded in tools we bought for something else, and who owns the decision that each is safe to use? If no single person can answer, that is the gap the regulation is designed to expose, and it is cheaper to close it now than under a statutory deadline.

When a supplier tells us their AI is safe, do we accept their word, or do we ask for the evidence — and would we recognise good evidence if we saw it? This is the exact question the committee is asking of the frontier labs, and it will be asked of your suppliers next.

If a model we depend on were found to be dangerous and withdrawn tomorrow, what in our business would stop working, and do we have an alternative? The recall the committee wants may be hard to enforce against the lab, but a model becoming unavailable — withdrawn, deprecated, or simply switched off by its maker — is a supply risk you can plan for regardless of what Parliament does.

The closing observation

The report is better than its headline, and the headline is the part that will not work. That is not a reason to dismiss it; it is a reason to read it properly. Britain does not need, and cannot have, a regulator that reaches into the world and switches off a model that has already escaped into it. What it can have, and what this committee has actually argued for underneath the Frankenstein language, is a regime in which the most capable systems cannot reach the public without independent evidence that they are what their makers claim — the same bargain we struck with aircraft and medicines, and for the same reason, which is not that the technology is too dangerous to allow but that it is too important to leave to the unverified word of the people selling it.

The power to pull the plug makes the better speech. The power to insist on the evidence before the plug is ever plugged in is the one that would actually make us safer, and it is the quieter half of what these MPs have proposed. I hope the Bill, if it comes, keeps the quiet half and is honest about the loud one. The worst outcome would be a law that promises the public a switch that does not exist, while leaving the pre-release scrutiny that does the real work as an afterthought. We have seen that film with other technologies. It is better, as ever, to build the guardrail than to promise the rescue.