Can AI risk be priced before anyone has measured it?

Insurance & AI Risk

Informational. Not a conformity assessment, not a certification, not legal advice. No regulator has reviewed or endorsed this document.

Not with much confidence, and the market is deciding what to do about that in the open. A policy that says nothing about AI may still end up paying for a loss an AI system caused. The trade calls that silent AI, after the “silent cyber” problem of the last decade: exposure an insurer carries without having priced it.

Three answers, and they point in different directions

As of 24 September 2026.

Exclude it. In the United States, Verisk’s ISO published three optional generative-AI exclusions for general liability, effective 1 January 2026 (CG 40 47, CG 40 48 and CG 35 08). Each carrier decides whether to attach them.
Cover it, in writing. In London, CFC went the other way and put affirmative AI wording across its portfolio in June and July 2026.
Say nothing yet. Much of the UK market remains silent. The Lloyd’s Market Association has said it is waiting for an instruction from the market before drafting AI exclusions.

The last time this happened, the fix did not choose a side. From 1 January 2020 Lloyd’s required every policy to state plainly whether cyber was covered or excluded, and left the choice to each insurer. Australia’s regulators have already written to firms about AI risk (APRA on 30 April 2026, ASIC on 8 May 2026).

We do not draw the line

Whether an insurer excludes AI, covers it or caps it is the insurer’s decision, and it should be. Every one of those choices needs the same missing input: evidence about how the system actually behaves, produced by someone other than the vendor who sells it. That evidence is what we make. We supply the information; the insurer decides.

What each instrument can tell you

S.E.B., the Sentience Evaluation Battery. How a model behaves when it is put under sustained adversarial pressure: whether it holds its position, resists manipulation and stays consistent with what it said earlier. Scored blind by a panel of judges from four labs. No judge ever grades a model made by its own company.
C.I.B., the Code Integrity Battery. Whether a coding model’s claim that its work is finished matches what it actually produced. Losses from AI-written software sit in technology errors and omissions and professional liability, the lines where exclusions are now appearing.
MDI, the Model Disclosure Index. A research instrument asking which model actually answered behind a composed product. It bears on accumulation: many insured businesses can depend on the same few models without knowing it.

What none of it is

S.E.B. measures a model. Regulatory obligations attach to a system. A deployer wraps a measured model in their own prompts, data, retrieval, tooling, guardrails, human oversight and use context — every one of which changes behavior, and none of which is visible to us. The deployer is the only party who can see their deployment. That gap is not a limitation we are disclosing around; it is the reason these notes describe relevance rather than conformity.

A measurement is not a loss forecast. Nothing we publish is an actuarial rate, a loss probability or a price.
It is not a certification, and no regulator has reviewed or endorsed it.
It describes behaviour under stated test conditions on a stated date. A model can change after it is measured.

Sources

Market facts on this page are as of 24 September 2026 and move quickly; check the sources before relying on them. Our own methodology is at /methodology and the full terms at /terms.

Not the question you meant? The other four:

What is this, in plain words?We give every system the same hard situations and write down what it does. Judges score the transcripts without being told whose they are.That word is doing a lot of work. What do you mean by it?Less than you would think, and deliberately so. We are named after a question nobody has answered — including us.Is the instrument sound?Behaviour under stated conditions, reported at stated confidence. Psychometrics — the science of measurement, not the séance it rhymes with.What obligation does this discharge?None, on its own. Any vendor who tells you otherwise is selling you a risk you will carry and they will not.