SILENT AI

BLUF: The insurance industry prices hurricanes to the decimal, writes cover for nuclear plants, and underwrites satellites. In 2026 it looked at artificial intelligence and started writing it out of the policy. At least six carriers filed their own exclusions, regulators approved more than 80 percent of the requests they reviewed, and the industry's standards body issued a form that lets carriers strip generative-AI liability from commercial policies nationwide. The frontier labs cannot buy cover at scale and are backing settlements with investor money. The reason the industry's own analysts give is the same one we found in our investigation of the AI monoculture: when everything runs on the same few models, one failure hits everyone at once, and that is not a risk insurance can hold. The people paid to be right about risk have voted. Their vote is no.

THE END OF SILENT AI

For years, AI risk was "silent" in insurance. Policies neither mentioned nor excluded it, so a claim arising from an AI product would land on a general liability policy written before anyone thought about it. That began to end on January 1, 2026.

On that date the Verisk/ISO endorsement CG 40 47 01 26 became available to carriers. It is optional, but where a carrier attaches it, it excludes bodily injury, property damage, and personal or advertising injury "arising out of 'generative artificial intelligence'" from commercial general liability. ISO forms are the factory standard; most commercial policies in America are built on them. Since then, at least six carriers have filed their own AI exclusions with state regulators; a September 2026 CSIS count names eight.

The carriers said why, through the people who price the risk. In November 2025 the Financial Times reported that AIG, Great American, and WR Berkley had asked regulators for permission to exclude AI liabilities. Dennis Bertram, an underwriter at the specialty insurer Mosaic, told the paper AI outputs are "too much of a black box." (AIG said it had no plans to implement the exclusion it filed.) WR Berkley's language reaches "any actual or alleged use" of AI, including products merely "incorporating" it. Berkshire Hathaway, Chubb, and Travelers filed for the same, and in April 2026 The Information reported, citing a Wolfe Research analysis of the filings, that state regulators had approved more than 80 percent of those requests.

Receipts: Verisk's form descriptions via Claims Journal, Jul 20, 2026; Fenwick, "The End of 'Silent AI'?"; FT, Nov 23, 2025, via PYMNTS and TechCrunch; the Berkley endorsement text via Hunton; The Information, Apr 23, 2026, via CSIS and PYMNTS.

THE LABS CANNOT BUY COVER

In October 2025 the Financial Times reported that insurers balked at the multibillion-dollar claims facing OpenAI and Anthropic in copyright and liability suits. OpenAI had secured cover of up to $300 million for emerging AI risks, a figure one source disputed as much lower, and in any case a small fraction of potential exposure. OpenAI had considered self-insuring through a captive; Anthropic was paying part of its $1.5 billion authors settlement from its own funds.

Think about what that means. The most valuable private companies on earth, backed by the largest balance sheets in technology, cannot buy insurance for their core product. At least one academic proposal now openly argues for government as the insurer of last resort for AI.

Receipts: FT, Oct 8, 2025, via Slashdot and Silicon UK; Seeking Alpha on the insurers' hesitation; Trout, "Insuring Uninsurable Risks from AI: Government as Insurer of Last Resort," arXiv 2409.06672.

THE DIVERGENT BET

London is the exception, and the size of its bet is the tell. Lloyd's syndicates write affirmative AI liability. Armilla launched a standalone policy with Chaucer in April 2025, described by the trade press as the first of its kind; by early 2026 its limits reached $25 million or more per organization. Testudo launched in January 2026, and by late February its limits reached $9.25 million per insured. By one March 2026 count, five dedicated AI liability products exist worldwide; the same count names only three standalone providers. The Lloyd's Market Association has not drafted exclusions. Its head of technical underwriting, David Powell, told Insurance Insider in June that the LMA is "waiting for an instruction from a market to produce something for them."

So the market's affirmative price for AI liability is $9 to $25 million per organization, against exposures in the billions. That gap is the industry's estimate of the risk, and it is the honest one.

Receipts: Chaucer and Armilla releases, Apr 2025 and Feb 2026; Testudo insights, Jan 21 and Feb 26, 2026; Purdy House on the first products, Mar 22, 2026; Insurance Insider, Jun 2, 2026; ResultSense on CFC's Lloyd's cover, Jul 30, 2026.

THE REASON IS THE MONOCULTURE

The reason the industry's own analysts give for stepping back is the finding that fuses this story with our earlier one. Underwriting rests on independence: a fire in Ohio does not cause a fire in Oregon, so the losses can be pooled. AI breaks that. A paper this spring, "The Insurability Frontier of AI Risk," put it plainly: an upstream foundation-model failure "can correlate losses across many cedents at once," defeating the law of large numbers that underwriting rests on. The paper traces the argument to Aon's Kevin Kalinich, who told the FT in November 2025 that an upstream AI provider's mistake produces "a systemic, correlated, aggregated risk" across many insureds. Gallagher Re's March 2026 white paper tells reinsurers the same thing: AI failures "may affect unconnected insureds simultaneously."

That is the AI monoculture, rediscovered by actuaries. In our companion investigation we documented that every frontier lab was vulnerable to the same attacks, shares the same data-labeling vendors, and converges on the same behaviors. The insurers reached the same conclusion from the other side. When every hospital, bank, and law firm runs on three or four models, a single flaw is not a claim. It is a catastrophe with one cause and ten thousand plaintiffs. No pool can hold that.

Receipts: Leung et al., arXiv 2605.18784, May 6, 2026; Gallagher Re, "Smart Systems, Blind Spots," Mar 24, 2026; Fenwick on insurer reasoning; FractalNode, "The Monoculture" case file.

THE PROPERTY SIDE IS PRICING LIKE A HURRICANE

The other half of the industry, property cover for the buildings, is not refusing. It is pricing AI like a natural catastrophe. Swiss Re's research arm reported in July 2026 that hyperscaler capital spending on AI would reach $750 billion in 2026, and that the largest AI campuses now carry asset values exceeding $20 billion before the servers go in. In a March 2026 report it said lenders demand insurance limits covering the full cost of construction before financing, and cited a projection that global data-center premiums will rise from about $10.6 billion to $24.2 billion by 2030. On September 3, 2026, Verisk launched a U.S. Data Center Exposure Database covering more than 2,500 facilities, built for catastrophe analytics and portfolio accumulation. Brokers were already reporting, in April, that underwriters had begun managing data-center aggregation the way they manage catastrophe aggregation, with concentration surcharges emerging as a premium driver. Zurich's head of construction told Bloomberg it expects the boom to make insurance securitization necessary.

Loss data already exists. Gallagher's 2026 survey found one in five insurance-industry respondents reporting a client loss tied to AI in the past year. Stanford's 2026 AI Index counted 362 documented AI incidents in 2025, up 55 percent from 233 the year before.

So what? The property market is saying the buildings are a catastrophe class. The liability market is saying the outputs are uninsurable. Both halves of the industry are hedging the same bet the tech industry is doubling.

Receipts: Swiss Re press release and sigma 2/2026, Jul 8, 2026; Swiss Re sigma insights 07/2026, Mar 27, 2026; Verisk via GlobeNewswire, Sept 3, 2026; Hotaling Insurance Services, Apr 22, 2026; Claims Journal on Zurich, Jun 23, 2026; Risk & Insurance on accumulation, Apr 1, 2026; Gallagher survey via Risk & Insurance, May 14, 2026; Stanford AI Index 2026, ch. 3.

WHAT THE VOTE MEANS

Insurers are the only participants in this economy whose business model punishes optimism. They cannot ride a bubble. If they misprice risk, they pay the claim. So when the whole industry writes a risk out of the policy at the same moment the capital markets are pouring $750 billion a year into it, one of the two is wrong, and the one with skin in the game has the better record.

The frontier labs are now running the most valuable technology in the world without cover, on investor money, with correlated risk the actuaries say cannot be pooled. If a foundation-model failure ever hits ten thousand customers at once, there will be no insurer to pay. There will be a bailout or there will be nothing. That decision has already been made by omission.

WHAT HONEST WOULD LOOK LIKE

Say it in the filings. Every public company deploying frontier AI in a product should disclose that its AI liability is excluded from general cover and state what self-insurance backs it. Every lab should publish its cover limits. And the government should say, before the catastrophe rather than after, whether it intends to be the insurer of last resort, and on what terms.

CORRECTIONS

September 6, 2026. A night-shift fact-check of every receipt in this piece against primary documents found four errors and a dozen imprecisions in the first version. All are corrected above. The errors: the "black box" line was attributed to the carriers seeking exclusions, when the speaker was an underwriter at Mosaic quoted by the Financial Times, and the words "to underwrite" were never inside the quote; the Lloyd's Market Association line was a trade-site paraphrase printed as a quotation, now replaced with David Powell's actual words; the data-center premium projection was given as $23 billion by 2030 when the Swiss Re report says $24.2 billion (the $23 billion figure is Verisk's, from a different baseline); and the Gallagher survey was credited to an outlet that never covered it. Also tightened: the ISO endorsement is optional and "became available" rather than "took effect"; the 80 percent approval figure is The Information's, from an April 2026 analysis of the Berkshire, Chubb, and Travelers filings, not the November requests; OpenAI's cover was "up to" $300 million and disputed; the Armilla and Testudo limits date to 2026 expansions, not their launches; the "five products" count is one blogger's headline whose own list names three; the correlated-loss reasoning is on record from brokers and academics, not from reinsurers as stated; two Swiss Re figures came from a March report, not July; the hurricane-aggregation and surcharge observation is a broker's from April, not Verisk's from September; and the $700 billion capex figure now matches the $750 billion Swiss Re estimate the piece already cites. We publish our corrections the same way we publish our claims: with the receipts.

HONEST-EDGE APPENDIX

• No major insurer has formally declared AI "uninsurable" as a category. The exclusions, carve-outs, and small limits are the functional equivalent, and we label them as such.
• The 80 percent approval figure comes from The Information's report on a Wolfe Research analysis of filings. We have not seen the underlying analysis; we cite the outlets that carried it.
• Lloyd's affirmative products show that some underwriters believe the risk is priceable at small limits. We report that honestly.
• Actuarial models for AI liability are immature. The frontier paper maps the problem; it does not solve it.
• The correlated-loss argument is on record from brokers (Aon, Gallagher Re) and academics. We found no reinsurer stating it as its own reason for pulling back.

VERIFIED

September 6, 2026: all receipts resolved and checked against primaries. The ISO form text is confirmed from the sibling endorsement CG 40 48 01 26 and Verisk's published descriptions; the CG 40 47 PDF itself is not online. The Financial Times pieces are paywalled and are cited through the outlets that quoted them verbatim. The Gallagher survey page is script-rendered; the one-in-five figure is confirmed from two trade reports. See CORRECTIONS.

Companion to THE CORRECTION and THE MONOCULTURE. Full case file public. (A+I)²