Anthropic warns AI liability could exceed insurance capacity

Anthropic has warned that AI liability could exceed the insurance market's capacity to cover it, according to a report from insurance trade outlet Program Business. The warning puts a long-deferred question at the centre of the AI business conversation: when a model causes harm, who pays, and how much can actually be recovered.

The candidate material carries only that core claim. It does not say where or in what form Anthropic made the warning, and it offers no liability figures or premium estimates. The safer reading is directional: a leading model developer is signalling that the existing liability insurance market may not be able to absorb the exposure AI creates.

The concern is not abstract. When companies wire agents into customer service, code, finance and supply chains, the question legal teams ask most often is not whether the model is accurate but who carries the loss when it is not. The chain of responsibility between model vendor, deployer and end user is still governed largely by bespoke contracts.

Insurance is the quiet infrastructure underneath all of this. Without a policy that responds, many enterprise procurement processes simply cannot close; and to price a policy, an insurer has to be able to bound the risk — which turns engineering details like explainability, log retention and human review points into contract language.

If capacity really is the constraint, the first to feel it will not be the model companies but the enterprises attempting large-scale deployment. Projects get slowed by legal and risk review, or pushed toward captives and other self-insurance structures. That friction rarely makes headlines, but it sets the pace of adoption.

Three things to watch: whether Anthropic publishes a fuller account, whether insurers bring dedicated AI liability products and wordings to market, and whether regulators step into the question of how liability is allocated. Until then, the durable point in this report is the structural gap it names, not any specific number.