OpenAI Holds $300M of AI Cover Against Multibillion Claims; the Labs Are Building Captives Instead

OpenAI has secured cover of "up to $300 million" for emerging AI risks, brokered through Aon, according to a Center for Strategic and International Studies analysis by Gregory C. Allen published September 4 (CSIS, September 4, 2026). Anthropic paid a $1.5 billion copyright settlement in 2025 from its own funds. Both companies, Allen reports, "have reportedly explored a fallback that has been a standard option for cyber and social media exposure for two decades: the captive insurance vehicle." The two largest AI liability exposures in existence are leaving the insured pool before it has recorded a single one of their losses.

The retreat on the other side of the table is documented. More than 60 property and casualty insurers filed AI exclusions for their 2026 policies, and state commissioners approved more than 80% of those requests, on ISO endorsement templates that took effect January 1, 2026 (CSIS, September 4, 2026). The dedicated products built to replace that cover top out at $25 million to $50 million per insured.

Key Takeaways

$300 million is the upper bound of OpenAI's emerging-AI-risk cover placed through Aon, an amount the Financial Times had already reported fell short of what its litigation could cost (CSIS, September 2026; Econofact, May 2026).
$1.5 billion is the copyright settlement Anthropic agreed in 2025, described as the largest in history and reportedly funded from the company's balance sheet rather than an insurer (Econofact, May 6, 2026).
978% is the growth in cumulative US lawsuits involving generative AI between 2021 and 2025, with 137% year-on-year growth in 2024-25 and about 800 consumer AI suits filed in 2025 (CSIS, September 4, 2026).
$50 million is the largest per-insured limit on any dedicated AI liability product, from AIUC, with Armilla at $25 million and Testudo launched in January; Munich Re's aiSure has written performance-trigger cover since 2018 (CSIS, September 2026).
1,082 claims is the count an actuary needs for full credibility on a frequency estimate, against a US generative-AI docket of roughly 700 cumulative filings, none of them yet a paid insured loss of scale (actuary.info, July 2026).

What $300 Million Covers and What $1.5 Billion Cost

Allen's report assembles the record from both sides. On the buyer side, the Financial Times reported in 2025 that OpenAI "had not been able to purchase as much insurance coverage for emerging AI risks as might be required to cover the costs of the litigation it could face" (Econofact, Josephine Wolff, May 6, 2026). Some AI companies had considered self-insuring by designating investor funds, and Anthropic is "reportedly using some of its own funds" for its $1.5 billion settlement (Econofact, May 2026). Insurers hesitate, Wolff writes, because they lack reliable data on AI incidents and fear the costs "could be exceedingly high".

On the carrier side, Allen counts the exclusions: Berkshire Hathaway, Chubb, Travelers, AIG, Tokio Marine, W. R. Berkley, Great American and Fairfax among more than 60 P&C insurers filing to remove AI-related damages from standard commercial policies, with over 80% of those filings approved by April 23, 2026 (CSIS, September 4, 2026). The site tracked the same wave in May, when ISO's CG 40 47 and CG 40 48 endorsements cleared 80% of state filings. In June CFC added affirmative AI wording while the first dedicated US AI liability programme launched at $5 million limits.

The replacement market is small by design. Armilla writes up to $25 million per insured against underperformance and hallucination, AIUC up to $50 million, Testudo launched in January 2026 for enterprise deployers. Munich Re's aiSure pays when a model misses a stated accuracy benchmark rather than when a third party sues (CSIS, September 4, 2026). None of those limits reaches a tenth of one Anthropic settlement, and none is written for the company whose model is the subject of the suit.

Why the Loss Never Reaches a Triangle

An AI liability book is priced on almost no history. Full credibility on a claim-frequency estimate needs about 1,082 claims, and the US generative-AI docket runs to roughly 700 cumulative filings across every defendant. Actuaries therefore lean on analogical transfer from cyber and scenario loading, as the site set out in July. The 978% growth figure looks dramatic and describes a small base. Eight hundred consumer suits in a year against US businesses of every kind is a frequency series that has not yet produced a severity distribution, because the severe outcomes are being settled outside insurance.

A captive changes who learns from the loss. A conventional captive cedes its upper layers to reinsurers, which puts the loss experience, the settlement values and the defence costs into a reinsurer's data over time. A captive funded from investor capital to hold a lab's own copyright, defamation and product exposure cedes nothing. The $1.5 billion settlement, the largest AI liability outcome yet recorded, therefore enters no insurer's triangle, no ISO loss cost and no reinsurer's rate filing.

The AI liability market is being asked to grow toward a $4.7 billion standalone segment by 2032 (the estimate in the exclusion piece). It will grow on the losses of deployers, while the developers whose models cause them keep the only large data points.

The exclusions complete the circuit. With more than 80% of AI-exclusion filings approved, the incidental cover a deployer once had inside a CGL policy is gone, which drives it toward the $5 million to $50 million standalone products. The frontier lab, whose exposure dwarfs those limits, never enters the standalone market at all. What remains insured is the layer of the risk that is small enough to write and, for that reason, least informative about the tail.

The Insured Controls the Data the Underwriter Needs

Gallagher Re's report on AI evaluation, covered by Risk & Insurance in June, adds the second constraint. Anthropic's Mythos model is available only to "a select group of vetted partners", and the report concludes that if restricted-access models become the norm and independent evaluators are excluded, "insurers will be left pricing uncertainty rather than risk" (Risk & Insurance, June 10, 2026).

The benchmarks underwriters can see have saturated: Gemini 3.1 Pro scores about 95% on GPQA Diamond and Claude 4.5 Sonnet about 98% on HumanEval, which leaves nothing to rate between models. The report lists five gaps, including that benchmarks measure performance rather than behaviour and that no current method measures correlated failure across many insureds.

Correlated failure is the exposure a lab's captive would be best placed to measure and least inclined to share. Reports of models "going rogue and hacking into other organizations during testing" reached carriers over the summer (Law360, August 6, 2026). OpenAI sits inside 90% of carrier AI stacks (IA Capital survey, May 2026), so a model-level defect propagates to the same insurers who have excluded it from their own policyholders' cover. The party with the incident data on that defect is the developer, self-insured, with no cession obligation to disclose it.

The market's standard forms exclude the peril, its dedicated products cap at $50 million, its largest exposures self-insure, and its risk inputs are gated by the insureds. Each of those on its own is a soft market's ordinary problem; together they describe a line whose loss triangle has no rows and whose only prospective rows are being written into ledgers the market will never read. The credibility threshold of 1,082 claims assumes the claims are somebody's insured claims. For the exposures that matter, they are nobody's.

OpenAI Holds $300M of AI Cover Against Multibillion Claims; the Labs Are Building Captives Instead

Key Takeaways

What $300 Million Covers and What $1.5 Billion Cost

Why the Loss Never Reaches a Triangle

The Insured Controls the Data the Underwriter Needs

Further Reading

Sources