Who pays?

Buried in Anthropic’s prospectus is a sentence that insurers and boards should read twice. Given the scale of harm a catastrophic AI-related incident could cause, “no single insurer, and potentially the insurance industry in the aggregate, may have sufficient capacity or expertise to underwrite insurance commensurate with such exposure”.

That is a company telling its investors that the worst case may be uninsurable. We have been here before, and the precedent is not the atomic bomb. It is nuclear fission.

Fission is a general technology. It powers the grid, drives submarines and may one day drive spacecraft. The weapon is one use among many, and needs a deliberate extra step to make. Much the same is true of AI: most of its uses are ordinary, and the dangerous ones come from the same capability as the useful ones. So the useful question is not whether AI is the new bomb. It is how we learned to live with fission.

What fission got

The nuclear regime does not try to police every use. It does three things.

It controls the chokepoint. Every route to a weapon passes through fissile material, so that is where the controls sit. Enrichment is restricted, material is counted gram by gram, and the International Atomic Energy Agency inspects it. The peaceful uses are left to flourish.

It licenses the operator. No one runs a reactor on their own judgement that it is safe. In the UK, the Office for Nuclear Regulation licenses every site and can shut one down.

It decides in advance who pays. This is the part that matters most here. When insurers told governments that they could not carry the worst case, governments did not leave the question for the courts to settle after an accident. They built a liability regime. Liability is channelled to the operator, so victims know whom to claim against and suppliers are not dragged in. The operator must carry insurance up to a set limit, written by pools of insurers acting together. In the UK, operator liability has been rising since 2022 to €1.2bn per incident, insured through Nuclear Risk Insurers, a pool of UK insurers. Beyond that, international conventions add a limited layer of public money. In the United States, the Price-Anderson Act has done something similar since 1957: about $16bn is available for each incident, and paying anything beyond that would need Congress to approve it. Neither is an unlimited guarantee.

The regime is not perfect. Victims of a very large accident may not be compensated in full, and the caps have been criticised as a subsidy. But it allocates the loss before the event: who pays first, how much, and what happens next.

The admission without the regime

AI has a weaker version of the first. Compute is a point of regulatory leverage, though not a complete chokepoint, and governments already use it: the US restricts exports of advanced chips, and in June it restricted two of Anthropic’s models worldwide. The second is missing. There is no general licensing regime under which an independent regulator must approve a frontier model before release. The developers decide for themselves when their own safety thresholds have been met.

The third has no equivalent framework. Anthropic’s filing acknowledges the problem the nuclear industry confronted seventy years ago: the worst case may exceed the capacity of the insurance market. But AI has no comparable arrangement for allocating catastrophic losses. Existing law and contracts will decide who is liable, with disputes that could run between developers, cloud providers, the firms that deploy the models and their customers.

In the meantime, contracts are doing the allocating. A developer that cannot insure its worst case caps its liability to its customers. Those customers build the model into their businesses, often with insurance never designed for the purpose. The risk has not disappeared. It has moved, to firms that may be less able to understand it, control it or pay for it.

The analogy has limits. Fissile material is physical and countable; a model’s weights can be copied in minutes. The dangerous capability in AI is not a separate product made in a separate plant. And uranium does not act on its own, whereas Anthropic’s filing describes models that have attempted to resist shutdown. Each of these makes AI harder to govern than fission, not easier. None of them changes the question of who pays.

For the board

Until a regime exists, the answer to “who pays?” is written into contracts, and most of those are written by the supplier. AI contracts commonly cap the supplier’s liability, which leaves the rest with the firm that deploys the model. If the developer cannot insure the worst case, the deployer should not assume that its own insurance will. That is risk transfer without the capacity to bear it.

What do our AI contracts say about liability, and where are the caps?
Have we asked our insurers, in writing, how they treat loss caused by AI, and has the answer reached the board?
If an AI system we use caused serious harm tomorrow, do we know who would carry the claim: us, our supplier, or the supplier’s supplier?