The Wrong Analogy, at the Right Moment

The Foreign Secretary is right that governments should not wait for a catastrophe before regulating AI. She has picked the wrong catastrophe to name. The risk profile that ought to be scaring policymakers looks nothing like Hiroshima. It looks like asbestos.

The statement, and what it is doing. Yvette Cooper told reporters this week that the world "cannot afford to wait for an AI equivalent of Hiroshima" before acting on safety, framing frontier AI as the security challenge of the coming decade and calling for international rules of the kind that eventually followed the bomb. The framing is deliberate. It reaches for the most legible catastrophe in twentieth-century memory and asks whether the international system will move before, or only after, the equivalent moment in AI.

As a piece of political theatre, the analogy works. As a piece of regulatory reasoning, it has a problem the speech does not address: the Hiroshima model assumes there will be a Hiroshima. A single, dated, attributable event, with a photograph and a body count, around which governance can organise itself. That is not the shape of the risk that most AI safety researchers actually worry about, and it is not the shape of the harms that are already accruing.

What the nuclear analogy quietly assumes. Nuclear governance arrived after Hiroshima because Hiroshima was legible. A city, a weapon, a state, a date. Even then, the substantive international regime took time: the Non-Proliferation Treaty was signed in 1968, twenty-three years after the bomb fell. The regulatory sequence Cooper is warning against is in fact the sequence nuclear governance actually followed. If AI's timeline is faster than nuclear's, "act before Hiroshima" is asking the international system to do something it has not previously managed.

More importantly, the nuclear frame smuggles in an assumption about what the harm looks like. It assumes a discrete event, caused by an identifiable actor, with immediate and visible casualties. That is one AI risk scenario, mostly the one that involves autonomous weapons or a rogue deployment against critical infrastructure. It is not the scenario that describes most of what is happening now.

The asbestos frame does more work. Consider the alternative. Asbestos was known to be harmful for decades before regulation caught up. The mechanism was understood, the harm was distributed across millions of workers and residents, the industry had commercial incentives to delay, and by the time the legislation arrived the damage was baked into buildings, lungs and legal systems for a generation. Leaded petrol is the same story. So is tobacco. So, arguably, is social media.

The harms that pattern describes are diffuse, cumulative, attribution-resistant, and commercially defended. That is a much better fit for what AI safety researchers at institutions like the Centre for the Study of Existential Risk and the AI Now Institute actually flag: labour displacement without a replacement path, epistemic degradation from synthetic media at scale, concentration of decision-making power in a handful of model providers, chatbot-mediated radicalisation, and the steady erosion of the information commons.

None of those produce a Hiroshima. They produce a slow accumulation that registers, in the way asbestos registered, only once the damage is structural.

The regulatory urgency the asbestos frame produces does not depend on waiting for a single spectacular event. It depends on noticing that the mechanism is already known.

The UK's own position is the interesting subtext. Cooper's statement lands against a specific policy backdrop. The UK's 2023 AI White Paper explicitly declined to legislate a horizontal AI regime, preferring existing regulators using existing powers. The AI Safety Institute was rebranded the AI Security Institute in April 2024, a shift in emphasis that tracked away from broad safety toward national-security-adjacent concerns. The 2025 AI Opportunities Action Plan is, in its own framing, pro-deployment. Post-Brexit, the UK is not bound by the EU AI Act, whose unacceptable-risk prohibitions took effect in February 2025.

Read against that record, "the world cannot wait" is not only a call to the international system. It is a call, from inside the government, to a government that has so far chosen to wait. That is the more interesting story, and the one the coverage has largely missed.

The counter-case, taken seriously. The industry argument against precautionary regulation is not frivolous. Poorly calibrated rules do suppress useful deployments: diagnostic tools, fraud detection, climate modelling, tutoring systems that reach students no human tutor will. A regime that treats every frontier model as a latent Hiroshima will over-price safety and under-price the counterfactual harms of not deploying. That is a real cost, and any honest precautionary case has to name it.

The response is not that the counterfactual disappears. It is that the counterfactual gets easier to reason about under the asbestos frame than under the Hiroshima one. Asbestos regulation did not ban insulation. It banned a specific material once the mechanism was legible enough to distinguish from its substitutes. The equivalent for AI is targeted, mechanism-specific rules on the harms that are already visible: synthetic media provenance, model access for adversarial actors, deployment in high-stakes decisions without recourse. That is a smaller, more defensible regulatory ask than the one Cooper's rhetoric implies, and it is closer to what the EU AI Act actually does.

What the enacted governance is worth. It is worth noting that precautionary AI governance has, in several places, already been enacted. The EU AI Act is law. The Bletchley Declaration was signed by 28 countries including the US and China. Safety institutes exist in the UK, the US, and elsewhere. The honest question is not whether governance has moved but whether what has moved is substantive or performative, and whether it can be enforced against firms whose market capitalisations exceed the GDP of most of the states regulating them. Cooper's statement does not answer that question, and the Hiroshima frame makes it harder to ask, because it keeps the conversation anchored on the event that has not happened rather than the enforcement that already has not worked.

I would rather the Foreign Secretary had said the quieter, less quotable thing. The harm we should be regulating against does not look like a bomb. It looks like a material we kept installing in buildings for forty years after we knew what it did. That is a harder speech to give, because it implicates decisions the government is making now, not decisions a future international system might fail to make. But it is the accurate one.

Footnotes

¶