Subscribe
AI Policy
8 min read

Your Insurance Quietly Stopped Covering AI in 2026

Major insurers excluded AI losses from general liability policies in January 2026, ending the era of silent coverage due to surging litigation and unpriced risks.

Graph showing decline in AI insurance coverage availability and rise in corporate AI liability claims for 2026
Graph showing decline in AI insurance coverage availability and rise in corporate AI liability claims for 2026

In 2026, major insurance carriers officially ended "silent AI" coverage by implementing generative-AI exclusion endorsements to commercial general liability policies. This change was driven by a nearly 1000% surge in AI-related lawsuits and the structural inability of insurers to price correlated losses accurately. Companies are now responsible for managing these risks or purchasing specialized affirmative AI liability products.

For most of the last decade, a company that got sued over an AI mistake could assume its existing liability policies would respond. Nobody had agreed to cover AI losses explicitly. Insurers simply had not excluded them. That arrangement, known in the trade as "silent AI" coverage, ended this year, and most of the companies affected have not noticed.

The change is worth understanding, because insurance, not legislation, is becoming the binding constraint on how AI gets deployed.

What actually happened

In January 2026, Insurance Services Office (ISO), the Verisk subsidiary that writes the standardized policy language used across the US admitted market, put three generative-AI exclusion endorsements into effect for commercial general liability. CG 40 47 strips out bodily injury, property damage, and personal and advertising injury arising out of generative AI. CG 40 48 addresses personal and advertising injury claims, and CG 35 08 extends the exclusion logic to products and completed operations coverage, according to reporting compiled by the agency broker AJG and trade press, and summarized in a July client alert by the law firm Shumaker, Loop & Kendrick.

On April 23, The Information reported that state insurance commissioners had approved more than 80 percent of major carriers' requests to add AI exclusions to corporate liability policies, with the highest volumes of approvals in Florida, Connecticut and Maryland. Berkshire Hathaway, Chubb and Travelers were the named carriers; the CSIS analysis that treats the date as a regulatory milestone adds AIG and Tokio Marine to the list of carriers actively filing exclusions. W. Berkley went further, filing what coverage lawyers call an "absolute" AI exclusion across directors and officers, errors and omissions, and fiduciary lines: no carve-backs, applying to any claim touching the actual or alleged use, deployment or development of AI. The Financial Times separately reported that insurers are moving to cap AI-driven losses inside cybersecurity policies as well.

The exclusions began taking effect as early as January, state by state, at renewal.

Why carriers ran

The litigation curve is the proximate cause. US generative-AI-related lawsuits grew 978 percent between 2021 and 2025 according to Gallagher Re data cited by Fast Company, with 137 percent year-over-year growth in 2024 to 2025 in litigation data compiled by Testudo Global and cited by CSIS.

Underneath the curve is a structural problem. A 2025 Geneva Association assessment, summarized in the CSIS paper, finds that generative AI fails outright on three of the nine criteria of the Berliner framework, the industry's four-decade-old test for whether a risk class is insurable at all: maximum possible loss, average loss amount, and information asymmetry. A single defect in a widely deployed foundation model can propagate correlated losses across thousands of unrelated companies running the same model, the kind of accumulation that defeats the diversification logic reinsurance pooling is built on. And severity has no stable distribution: the same failure mode, a system confidently generating false content, has produced claims ranging from trivial to enormous.

When carriers cannot price a risk, they exclude it. The pattern is the cyber insurance playbook of the 1990s replayed on a compressed clock: silent coverage, mass exclusion, then a standalone specialty market.

Illustration: a central machine block radiating concentric rings toward a circle of small desks, one tiny umbrella among them

How broad the exclusions are

The phrase that matters is "arising out of." Under the ISO forms, a carrier does not need the AI system to be the sole cause of a loss; it needs only a close enough connection between the harm and some generative-AI involvement. As the Shumaker alert notes, AI is already embedded in customer service tools, applicant screening, fraud monitoring, marketing platforms, contract review, call summaries and office software. A broad exclusion lets the insurer litigate the mechanism ("did AI contribute?") instead of the harm ("was there an error?"). Coverage lawyers have warned that where exclusions are drafted broadly enough, the practical effect for some policyholders is coverage that is illusory.

Illustration: a torn paper umbrella with rain falling through the hole

What the affirmative AI market actually sells

A small specialty market has formed in the gap, and it is useful to know what these products do and do not cover, because they are not substitutes for a mature casualty line.

  • Munich Re's aiSure, in the market since 2018, ties payout to an agreed model performance metric: coverage responds when a defined deployment falls below a negotiated accuracy or error threshold.
  • Armilla AI, a Lloyd's coverholder backed by Chaucer and Axis Capital, sells affirmative coverage for AI underperformance, hallucinations and regulatory violations, with limits up to $25 million per insured and up to $5 million per model, and requires an independent AI system assessment before binding. In February 2026 it launched "Vanguard AI" with Chaucer, pairing dedicated AI limits with cyber and tech E&O.
  • Testudo Global launched in January 2026 as an MGA partnered with Gallagher Re and MIT, selling claims-made generative-AI liability to mid-market and enterprise deployers, reported at $10 million aggregate per insured as of August 2026, available only through wholesale brokers on a surplus lines basis.
  • The Artificial Intelligence Underwriting Company (AIUC) writes coverage up to $50 million per insured and counts ElevenLabs among its named policyholders, paired with its AIUC-1 certification standard.
  • Corgi, a Y Combinator-backed full-stack insurer, bundles AI exposure into tech E&O for startups; buyers should note it writes primarily on an unrated risk retention group paper, which some contracts will not accept.
  • HSB, a Munich Re subsidiary, began selling AI liability for small businesses in March 2026.

The most instructive product is the one that shows the boundary. AXA XL's CyberRiskConnect generative-AI endorsement, launched in October 2024 and the first significant AI-specific endorsement from a top-tier carrier, covers data poisoning, usage-rights infringement and EU AI Act regulatory violations, but only for companies that develop or deploy their own models. Users of third-party services like ChatGPT are excluded. As CSIS puts it, the customers most likely to need coverage are precisely the ones the endorsement does not cover, because underwriters can verify governance when the insured controls the model, and cannot when the insured is a downstream consumer.

Even the labs cannot buy enough coverage

The retreat reaches the top of the stack. The Financial Times reported in October 2025 that OpenAI has secured cover of up to $300 million for emerging AI risks, brokered through Aon, a cap that looks small against the lawsuits already in flight. Anthropic agreed to pay a $1.5 billion copyright settlement from its balance sheet and is reportedly considering investor funds for future settlements. The net effect, in CSIS's framing, is a regressive distribution: capital-rich labs retain risk on balance sheets and captives at the top, large enterprises buy capped specialty products in the middle, and smaller deployers and consumers carry the residual with no insurance and no contractual indemnity.

Meanwhile the statutory backstops that existed for comparable tail risks have not arrived for AI. The European Union formally withdrew its AI Liability Directive in July 2025, then delayed the AI Act's high-risk rules again via the AI Omnibus that entered force on July 27, 2026. China is going the other way, treating insurance as industrial policy: the state-backed PICC has piloted standalone generative-AI coverage for IP, portrait and reputational infringement claims.

What to do before your next renewal

None of this requires a policy opinion, only a checklist the trade reporting converges on:

  1. Read the endorsement schedule of your GL, cyber, D&O and E&O renewals, not just the declarations page. Ask specifically whether CG 40 47, CG 40 48, CG 35 08 or equivalent carrier-filed language has been attached.
  2. Inventory which production systems involve generative AI, including features inside vendor software, so you know which claims the "arising out of" language could reach.
  3. Review AI vendor contracts and insurance together. Vendor liability caps are commonly low and disclaimers broad; if the vendor caps and the carrier excludes, the gap is yours.
  4. Document governance: approved use cases, testing, human review, monitoring, incident response. The specialty carriers that price AI affirmatively are moving toward controls-based underwriting, the same bargain cyber insurers struck a decade ago.

The cyber precedent says the transition takes years, and the AI specialty market is starting from a fraction of the capacity that cyber eventually had. The uncomfortable summary: in 2026 the insurance market decided AI risk is not yet a priced risk, and handed it back to whoever operates the systems.

Sources

Frequently asked questions

What specific ISO forms exclude AI coverage?

The three endorsements are CG 40 47 (stripping bodily injury, property damage, and personal/advertising injury), CG 40 48 (addressing personal/advertising injury), and CG 35 08 (extending exclusions to products and completed operations).

Why did insurers stop covering AI risks?

Litigation grew 978% between 2021 and 2025, and generative AI fails key insurability criteria like loss predictability and diversification because single model defects can cause correlated losses across many unrelated companies.

Which carriers filed absolute AI exclusions?

W. Berkley filed an "absolute" AI exclusion across directors and officers, errors and omissions, and fiduciary lines with no carve-backs; Berkshire Hathaway, Chubb, Travelers, AIG, and Tokio Marine also actively filed exclusions.

Does AXA XL cover third-party AI users?

No, AXA XL's CyberRiskConnect generative-AI endorsement only covers companies that develop or deploy their own models; users of third-party services like ChatGPT are excluded.

AI PolicyInsuranceEnterprise AI