One leading AI chief executive reportedly told a Berkeley computer scientist that he did not expect serious regulation of his industry until there was a “Chornobyl-scale disaster”. That single line, buried in a Guardian column by Stuart Russell, is the most honest summary of where the world has landed: the people building the most powerful systems in history assume the rules will arrive only after something has gone catastrophically wrong. Britain has spent two years positioning itself as the country that takes AI risk seriously. It is now worth asking, plainly, whether the UK is regulating ahead of the danger or quietly waiting for the disaster like everyone else.
The wager nobody voted for
Russell’s argument is uncomfortable because it is structural rather than alarmist. AI risk rises in lockstep with AI capability, he writes, and capability is rising fast. He cites the June events around Anthropic as symptoms: the company’s own published warning about early signs of recursive self-improvement, its suggestion that the world should “slow or temporarily pause frontier AI development”, and a reported case in which a frontier model could run end-to-end cyber-attacks with no human in the loop. The chief executives, in his telling, openly describe a meaningful chance of human-level catastrophe, and governments reply by offering subsidies and fast-tracked permits.
The strategic point for a UK audience is not whether the most extreme scenarios are right. It is that the prevailing regulatory posture — across most of the democratic world — is to wait. Wait for evidence, wait for an incident, wait for the politics to turn. That is a wager. And it is being placed on behalf of every business, public body and citizen that now depends on these systems, by people they did not elect and cannot lobby.
Strategic Reality: “Wait and see” is not a neutral position. It is an active bet that the first serious AI failure will be survivable, recoverable and instructive rather than catastrophic. Most safety regimes in history were written the other way round — before the technology was allowed to scale, not after it failed.
Russell’s proposed alternative is deliberately unexciting: a licensing regime that requires a minimum safety standard before a system can be built and released. As he points out, this is how nuclear power, aircraft, buildings, lifts, hairdressers and sandwich makers are already handled. The provocation is the contrast. We license the person making your lunch, but not the company claiming to build the most dangerous technology in history.
How Britain got out in front, then drifted
For a brief period, the UK had the strongest claim to leadership on AI safety of any government. The numbers tell the trajectory.
| Milestone | Date | What it signalled |
|---|---|---|
| AI Safety Summit, Bletchley Park | November 2023 | 28 nations plus the EU sign the Bletchley Declaration on frontier risk |
| AI Safety Institute founded | November 2023 | First state-backed body for independent frontier-model evaluation |
| Seoul AI Summit (UK co-hosted) | May 2024 | 16 companies sign voluntary Frontier AI Safety Commitments |
| AI Opportunities Action Plan | January 2025 | Emphasis pivots from safety to growth and adoption |
| Institute renamed “AI Security Institute” | February 2025 | ”Safety” drops out of the name; focus shifts to security and crime |
| Paris AI Action Summit | February 2025 | The agenda becomes “action”; the UK declines to sign the final declaration |
Read top to bottom, the table describes a country that convened the world on safety and then, within fifteen months, recast the conversation around opportunity. None of that pivot is irrational. Growth is a legitimate national priority, and the UK genuinely has assets here — chiefly the institute, which retains real technical credibility and pre-deployment testing relationships with the major labs. But the binding rules never came.
Critical Context: The UK still has no statute governing frontier AI. The 2023 white paper chose a “pro-innovation” model with five principles applied by existing sector regulators and no new powers. The frontier legislation promised in the 2024 King’s Speech — rules for “the most powerful artificial intelligence models” — has been consulted on and deferred rather than enacted.
So Britain occupies an unusual position. It has the world’s most developed institutional capacity to assess frontier risk, and among the least developed legal capacity to do anything about what it finds. The institute can test a model and form a view. It cannot withhold a licence, because there is no licence to withhold.
What the gap actually costs
A regulatory vacuum is rarely free, even for the firms it is meant to favour. It simply moves the cost around.
The first cost is reactive whiplash. Russell’s column opens on exactly this: a US administration that spent years deregulating, then issued an abrupt export directive that pulled deployed models off the market with little notice. Regulation deferred does not mean regulation avoided. It means regulation arriving suddenly, in a crisis, in whatever shape the panic dictates. For businesses, a hard rule announced calmly with a two-year runway is vastly easier to absorb than the same rule imposed overnight after an incident.
The second cost is divergence. The EU AI Act is now in force and phasing in obligations on high-risk and general-purpose systems. Any UK organisation operating across the Channel already builds to that standard. A UK firm that designs only to the lighter domestic regime risks doing the compliance work twice — once for reality, once for the rules that eventually catch up.
Hidden Cost: The absence of UK rules does not reduce your compliance burden if you trade into the EU, sell to enterprises that demand assurance, or operate in a regulated sector. It just removes the clear domestic benchmark you would otherwise design against — and replaces it with guesswork about what Westminster will do under pressure.
The third cost is trust. Buyers, insurers and procurement teams increasingly want evidence that an AI system has been evaluated, not just shipped. Where the state sets no floor, that burden falls entirely on the vendor to invent and on the buyer to assess. A licensing floor, whatever its frictions, gives everyone a shared baseline.
Three regimes, three bets
| Jurisdiction | Posture | The implicit bet |
|---|---|---|
| EU | Binding, risk-tiered (AI Act in force) | Front-load the cost; accept slower deployment for predictability |
| UK | Principles-based, non-statutory, institute-led | Stay flexible; legislate later, ideally before a crisis forces it |
| US | Deregulatory by default, episodic intervention | Move fastest; reach for blunt instruments (export controls) when alarmed |
The UK bet is the most finely balanced of the three. It can become the best of both worlds — agile rules informed by genuine technical evaluation — or the worst, if the technical capacity to see the risk never connects to the legal capacity to act on it.
Why “wait” feels rational and usually isn’t
The case for waiting is seductive and worth stating fairly. The technology is moving so fast that any rule risks being obsolete on arrival. Premature regulation could entrench incumbents and push frontier work offshore. And the genuinely catastrophic scenarios remain contested among serious experts. These are not stupid arguments.
But each has a quieter rebuttal. Rules that set a standard of care rather than a fixed technical specification age well — aviation safety law does not name the components of a jet engine, it requires the engine to be demonstrably safe. The offshoring fear cuts both ways: a credible licensing regime can become a mark of quality that buyers prefer, much as the EU’s data rules became a de facto global benchmark. And “the experts disagree” is an argument for a precautionary floor, not against one, precisely because the downside is asymmetric.
Reality Check: Russell’s sharpest line is that the recent reversal in US policy suggests the trigger for real regulation might not be a Chornobyl after all — it might only take a “Three Mile Island”. A contained, frightening, non-fatal incident may be enough to flip the politics overnight. Plan as though the rules could harden on short notice, because that is now the most likely path.
The honest reading is that “wait and see” usually means “wait until forced, then overcorrect”. For a business, the worst outcome is not regulation. It is regulation that lands abruptly, written in anger, with no transition period — and finds you unprepared.
What UK leaders should do before the rules arrive
National policy will move on its own timetable, and probably not the one you would choose. The variable you control is how ready your organisation is for a regime that is currently a matter of when, not if. The work splits by maturity.
For organisations early in adoption, build a basic AI inventory now. Know which systems you operate or rely on, which are frontier-grade, and which touch anything a regulator would call high-risk — decisions about people, money, health, safety or critical infrastructure. You cannot govern what you have not catalogued, and most teams have never drawn this map.
For organisations scaling AI, design to the stricter neighbour. Where the EU AI Act sets a clear high-risk standard, treat it as your working baseline rather than waiting for a lighter UK equivalent that may never materialise or may later converge upward. Document your evaluation, testing and human-oversight arrangements as if you already had to show them, because the direction of travel says you eventually will.
For organisations where AI is core to the product, build a safety case as a first-class artefact, not a compliance afterthought. Be able to state, in writing, what your system can and cannot do, how you tested it, and what would make you pull it. That is precisely the evidence a future licensing regime — UK, EU or via your largest customers — will ask for, and the firms that have it will move while competitors scramble.
Take Action: For your most consequential AI system, write the one-page answer to “what is our minimum safety standard, and how do we know we meet it?” If that page cannot be written today, you have found your most urgent governance gap — and a head start on whatever the rules turn out to require.
The challenges leaders tend to underestimate
Four difficulties surface only once an organisation tries to govern AI seriously, and each has a practical response.
The first is the evidence vacuum. Teams discover, often late, that they cannot actually describe how a critical model behaves at the edges, because no one tested for it. The fix is to make adversarial and edge-case testing a release gate, not an optional extra.
The second is regulatory mirroring. Firms wait for the UK to act, then design narrowly to whatever lands — and find it is weaker than what their EU operations or enterprise customers already demand. Design to the strictest binding standard you are realistically exposed to, and the domestic rules become a formality rather than a scramble.
The third is governance theatre. An AI ethics policy on the intranet is not a control. The mitigation is to attach real decision rights to it: a named owner who can stop a deployment, and a documented threshold at which they must.
The fourth is timing risk. Because the trigger for hard regulation may be a single incident rather than a planned consultation, the runway could be far shorter than the usual policy cycle implies. Treat readiness as a standing capability, not a project you start when a bill is published.
Warning ⚠️: “There are no rules yet” is not a strategy. It is a description of the calm before the rules — and the firms that use the calm to prepare will outlast the ones that mistake it for permission to defer.
The strategic takeaway
Russell’s column is addressed to governments, but its most useful lesson is for everyone downstream of the decisions governments are deferring. The world is running a live experiment in whether powerful AI can be governed reactively — regulated after the first serious failure rather than before it. Britain is a sophisticated participant in that experiment, with better instruments than most and less legal authority than its rhetoric implies. Whether that gap closes gracefully or violently is not something any single business can decide.
What each organisation can decide is its own posture toward the rules that are coming. Three commitments separate the prepared from the exposed: knowing exactly which of your AI systems would attract scrutiny under any serious regime; designing to the strictest standard you are genuinely exposed to rather than the most convenient; and holding a written, current account of why your most important systems are safe enough to run. None of that requires a new law to begin. All of it gets harder, and more expensive, if you wait for one.
Success Factor: The organisations that come through the next regulatory turn well will not be the ones that guessed the timing. They will be the ones who built so that the timing did not matter — ready for the rules whether they arrive by design or by disaster.
The unsettling possibility in Russell’s piece is that the question is no longer whether AI will be regulated, but what it will cost to get there. A licensing floor written calmly, in advance, is cheap. The same floor written in the aftermath of a Three Mile Island moment is not. Individual firms cannot choose which path the country takes — but they can make sure that, either way, they are not the ones caught building on the assumption that the rules would never come.
Source and attribution
This analysis draws on “Will it take a ‘Chornobyl-scale disaster’ for us to regulate AI?” by Stuart Russell, published in The Guardian on 17 June 2026. Russell is a distinguished professor of computer science at the University of California, Berkeley, president of the International Association for Safe and Ethical Artificial Intelligence, and a Guardian US columnist. The column argues that AI regulation lags catastrophic risk and proposes a pre-deployment licensing regime. The UK regulatory framing, the timeline of British AI policy, the cross-jurisdiction comparison and the business recommendations are original to Resultsense.
Resultsense provides analysis to help UK professionals and businesses make sense of AI developments. For more, explore our insights and news coverage, or get in touch to discuss how these shifts affect your organisation.