TOP
Would you buy a used car if you could not tell the good ones from the bad? Neither, it turns out, will an insurer.
The price of coverage depends not only on the losses that may occur, but on what each side knows about them. Any market in which the two sides know different things behaves in a characteristic way.
In insurance, the party seeking coverage often knows more about the exposure than the insurer does. Artificial intelligence sharpens the imbalance. Much of what an AI system will do is hard to observe in advance, and the most capable models cannot readily be examined independently.
Baruch Berliner’s framework includes, among its market conditions, limits on adverse selection and moral hazard, both of which arise when the insurer cannot see what the policyholder can.1 Biener, Eling, and Wirfs find information asymmetry among the market conditions cyber strains.2
Economics has a model of what follows. In 1970 George Akerlof described a market for used cars in which the buyer cannot tell a sound car from a bad one – a “lemon.”3 Put yourself in it. Not knowing which is which, a buyer will pay only for a car of average quality; but the owner of a good car knows it is worth more than that average price and will not sell at it, while the owner of a lemon is glad to accept that price. So, the good cars do not come to market, and what is left on offer is worse than the average the buyer was prepared to pay for. Assuming the worst is the reasonable response to hidden quality, and the assumption, once priced in, helps make itself true.
An insurer faces the same logic in reverse. For each firm seeking insurance cover, two prices matter. The insurer’s break-even premium is its expected loss on the account: spread across a diversified book, that idiosyncratic risk averages away, so the expected loss is all the insurer needs. The firm’s willingness to pay is higher, because it bears the same risk undiversified and is risk-averse – it prefers a fixed premium to an uncertain loss of equal expected value and will pay above the expected loss to shed the uncertainty. That excess is the firm’s cost of bearing the risk; it is also the gain from transfer, positive for the low-risk and the high-risk firm alike, so the transaction is worth entering for each. The insurer cannot see which is which, so it quotes one premium to the applicant pool – the average. That pooled premium is above the low-risk firm’s willingness to pay, so the low-risk firm retains the risk, and only high-risk firms remain in the market. The graphic sets out the two-type case.
Adverse selection in cyber insurance: cover is worth writing for both firm types, yet the pooled premium drives the low-risk firm out of the pool. Circles mark each firm type’s expected loss (the insurer’s break-even premium), squares its willingness to pay; the dashed vertical is the pooled premium, 0.5 × $40,000 + 0.5 × $160,000 = $100,000 – above the low-risk firm’s $70,000, so it declines cover and only high-risk firms remain. Illustrative assumptions; a stylized adaptation to insurance of the model in Akerlof’s 1970 article, cited in the notes of this post. Source: Gen Re.
Whether the low-risk firm stays in the market depends on how much can be seen. Where a risk can be inspected – routine cyber, with its established controls and its loss record – price can track type, the low-risk firm is offered a premium it will pay, and cover is written. Where type cannot be seen, the insurer prices to the pool, the low-risk firm declines, and the exposure that takes up cover is the high-risk one. Past a point, the response is not to price the exposure but to exclude it.
The mechanism in the graphic below matches the record. Over the past year, carriers have filed to exclude AI‑related losses from standard liability policies. In the United States, the Insurance Services Office (ISO) introduced three generative‑AI exclusion endorsements for commercial general liability, effective January 2026; the endorsements are optional, and adoption is at the discretion of each insurer.4
Some carrier exclusions are absolute, reaching across Directors & Officers (D&O), Errors & Omissions (E&O), and fiduciary lines.5 In the United States, by mid-2026, state regulators had approved more than 80% of such filings.6 In July 2026, Verisk, ISO’s parent, said it is evaluating additional options for AI exposures, including agentic AI – AI that acts autonomously rather than generating content on request.7 A filings census published in July 2026 measured the breadth: a review of nearly 10,000 filings found that more than 60 Property and Casualty groups had filed on AI exclusions – 41 groups with at least one subsidiary filing to adopt, and a further 20 filing to delay adoption – across Commercial General Liability, Umbrella and Excess, Errors and-Omissions, and package lines, with one large carrier stating it was not currently seeking to implement; the filings hold the option open. A carrier adopting the exclusions broadly described the exposure as “an area of emerging aggregation risk” that Personal and Advertising Injury coverage was not designed to manage.8 The reason underwriters give is the one Akerlof’s model anticipates: an exposure they cannot evaluate is one they will not price. This graphic places that record on a time axis – the exclusion perimeter advancing step by step, and, beneath it, the response forming in dedicated cover.
Selected milestones from April 2025: the exclusion of AI from standard liability cover above the axis (deep blue) and the response in dedicated cover below (amber). The backdrop to the record: litigation over generative AI in the United States grew 978% between 2021 and 2025, by a reinsurance broker’s count. Not a census of the market; approximate dates are placed as stated: “by mid-2026” between April and July 2026, “early 2026” at February 2026. Sources: as cited in the notes to this post; compilation Gen Re.
A different reading comes from other market participants. A global reinsurance broker, in a 2026 study with partners, argues that AI liability is a growing and largely uninsured exposure – generative‑AI litigation in the United States grew by 978% between 2021 and 2025 on its count – and that the market should meet it with dedicated cover and governance rather than exclusion.9 A study published in July 2026, led by a specialist AI‑underwriting firm with co‑authors from academia, research institutes, broking, underwriting, and AI development, locates the same adverse-selection dynamic in agentic AI – arising, in its account, when “responsible actors, correctly judging themselves to be overcharged, decline coverage” – and holds that the response is risk selection, auditing, and technical underwriting, not exclusion.10 One of that firm’s co‑founders had put the gap to the Financial Times in November 2025, months before the study: “Nobody knows who’s liable if things go wrong.”11
Both readings invoke the cyber precedent: cyber, too, was carved out of standard policies before a standalone market grew up around it – a repeat now underway, with a Lloyd’s-backed managing general agent (MGA) underwriting standalone AI liability cover for U.S. mid-market enterprises since early 2026, and an established carrier describing the exclusions as the most immediate opening for a dedicated product.12
Diversification, measurement, accumulation, and inspection are separate conditions in the study of insurability. Artificial intelligence strains each through one feature: a few shared, fast-moving, hard-to-examine systems on which many exposures depend at once. Where that feature is present, adding risks does not reduce the relative variability of the loss, a rate cannot be trusted across the tail, a single event can reach much of the book of insurance, and the exposure cannot be inspected in advance.
None of this determines whether AI should be insured, or on what terms. It locates where the decisive choices are made: in the policy wording – to grant cover, cap it with a sub-limit, or exclude it – applied before the rate. The market is writing them now, filing by filing.
In case you missed the other articles in this series, take a look back at them.
