The Wall Street Journal reported on October 17, 2026 that, as AI giants rack up potential legal liabilities of tens of billions of dollars amid safety incidents, insurers worry that claims may outpace the limits of existing coverage. The piece (by Jean Eaglesham) sits behind a paywall, so this post does what we do with any hard-to-open report: separate what is claimed, what is documented elsewhere, and what you can check yourself. The short version is that the headline figure is unconfirmed, but the trend behind it, insurers writing AI out of standard policies, is on the record.
If you build with or buy AI, this matters more than most model launches. A risk that cannot be insured is, for many businesses, effectively a risk they cannot take.
TL;DR: claimed vs verified
| Question | Answer | Status |
|---|---|---|
| Do AI labs face liabilities in the tens of billions? | WSJ headline says "potential" liabilities of tens of billions | Reported; the full article is paywalled |
| Is OpenAI's cover about $300 million? | CSIS says "reportedly capped at roughly $300 million" | Reported, not confirmed by OpenAI or insurers |
| Did Anthropic pay $1.5 billion from its own balance sheet? | CSIS says yes and that it is reportedly weighing investor funds for future settlements | Settlement documented in the authors case; funding detail reported |
| Are insurers adding AI exclusions? | ISO generative-AI endorsements took effect January 2026; 60+ carrier groups filed exclusions | Documented by industry sources |
| Have claims actually been denied because of an AI exclusion? | RIMS president says he has not yet heard of one | Not yet observed publicly |
| Will exclusions bite at renewal? | Analysts say it is too early to know before renewals arrive | Open question |
What the WSJ report says, and what we could not open
The only text we could verify directly is the headline and standfirst: insurers fear that claims against AI companies "may outpace limits on existing coverage," and that the industry's rapid growth is the cause. We have not read the body, so we do not repeat any figure from it. If a number in this post is attributed to someone else, it is because their own publication says so, and we link it.
We also could not find a statement from OpenAI, Anthropic, Google or Meta about their insurance programs, which is normal: labs rarely publish policy limits. That gap is itself the story. When the biggest potential defendants will not say how much cover they carry, outside estimates fill the space, and those estimates are shaky.
What is documented: the exclusion wave
Two independent sources describe the same shift.
Global Finance (October 2026 issue). Michel Léonard, chief economist and data scientist at the Insurance Information Institute, pointed to the generative-AI exclusion endorsements for general liability that the Insurance Services Office (a Verisk subsidiary) introduced in January 2026. He said more than 60 property and casualty carrier groups have filed to adopt some form of AI exclusion, while analysts say it is too early to judge the effect before renewals. The magazine reports the broadest "absolute" exclusions in directors and officers, errors and omissions, employment practices and fiduciary lines, and cites a Berkley exclusion that removed D&O coverage touching AI use, deployment or development. Source: Global Finance, "AI Insurance Is a Risk Officer's Nightmare".
CSIS (September 4, 2026). Gregory Allen's report says The Information reported on April 23, 2026 that state insurance commissioners had approved more than 80 percent of carrier requests to exclude AI-related damages from corporate policies. Allen's argument is that insurance has become the de facto regulator of US AI deployment, "regulating by withdrawal." Source: CSIS, "The Insurance Industry's Retreat from AI".
AI liability insurance coverage shown as a shield over a small house
What is reported but unconfirmed: the lab numbers
The same CSIS report says "frontier AI labs cannot fully insure their own litigation exposure," describing OpenAI as "reportedly capped at roughly $300 million in coverage" and Anthropic as having agreed to a $1.5 billion settlement from its balance sheet, while reportedly considering investor funds for future settlements. The word "reportedly" matters. We found no filing or company statement confirming the $300 million limit, and a limit of that size would be tiny next to the exposure described.
For context on what the exposure looks like, CSIS cites litigation data from Testudo Global showing cumulative US generative-AI lawsuits up 978 percent between 2021 and 2025, and lists copyright suits (including by the New York Times and an authors coalition) and a wrongful-death case against OpenAI. Our own AI lawsuit tracker post covers the case count and the copyright rulings in more detail.
Why labs cannot just buy more insurance
Allen's paper leans on the insurance industry's long-used test for insurable risks and argues generative AI currently fails outright on three of its nine criteria. The central failure is information asymmetry: carriers cannot see which models a customer runs, what for, or whether claimed controls exist. That makes it hard to price two classic problems, moral hazard and adverse selection.
Three further factors push limits down:
- Correlated losses. One model bug or one jailbreak can hit thousands of customers at once, the opposite of the independent losses insurers like.
- No loss history. Pricing needs frequency and severity data. AI incidents are new, and public incident databases are thin. Our felony-bench tracker is one attempt to log cases where agents affect third parties.
- Silent cover. Global Finance quotes the pattern from early cyber insurance: risks first sit inside general policies without being priced, losses mount, then insurers exclude. FurtherAI's CEO told the magazine cyber took about 20 years to mature and AI may take 10 to 12.
Where the liability goes when the lab cannot pay
CSIS makes a point that deserves emphasis: vendor liability caps "push residual risk down to enterprise deployers, and from deployers to consumers." If the model provider's terms cap damages and its insurance is limited, the company that wired the model into a customer-facing workflow holds the exposure, and that company's own policy may now exclude it.
This connects to the federal conversation. Our post on the Trump AI task force's liability-not-new-rules approach explains why courts and contracts, not new statutes, are likely to decide who pays when an agent goes wrong. Insurance is the missing piece of that picture.
Real incidents show how quickly claims can arise:
- A lawyer was fined over AI-fabricated citations in the State Farm case, a small sum but a clear professional-liability pattern.
- Police acted on a false tip from a model in the Philadelphia homicide report, the sort of event that raises defamation and negligence questions.
- Banks are already limiting what agents can do with accounts; see how Citi, Bank of America and JPMorgan block AI agents over deposit risk.
What this means for what you build or pay
This is a finance-adjacent story, so here is the practitioner takeaway:
- Ask your broker now. Request written confirmation of whether your D and O, E and O, general liability and cyber renewals contain a generative-AI exclusion, endorsement or extension to an existing cyber exclusion. Global Finance notes such language is not always labeled as AI and may appear as an endorsement rather than on the declarations page.
- Read vendor terms. Look for liability caps, indemnities for IP claims, and whether the provider backs them with real insurance.
- Inventory your AI. Insurers' core problem is not seeing what you run. A current list of models, use cases and owners helps with underwriting and with governance. The risk of staff using unapproved tools is covered in our shadow AI guide.
- Keep humans on high-stakes actions. Agents with narrow permissions are easier to defend and insure. For agent security specifically, AgentBeam, the agent security platform from the explainx.ai team, is built to stop AI agents before they take dangerous actions.
- Watch January 1 renewals. Analysts quoted by Global Finance say the real impact will only be visible when renewals arrive.
How to check the claims yourself
- Read the headline and standfirst of the WSJ article through its aggregator listing, and look for the paywalled original.
- Read the full CSIS report PDF linked from its page for sourced footnotes on the $300 million and $1.5 billion figures.
- Check your state insurance department's filing database for AI exclusion forms from your carrier.
- Look for ISO form names for generative-AI exclusions on your policy documents.
- Compare with our earlier look at a related unverified number in OpenAI's $50 billion run rate, not $70 billion, where we applied the same method.
What to watch next
Three signals will tell us whether the "claims outpace limits" worry is real or overstated. First, whether any carrier publicly denies a claim because of an AI exclusion; so far the head of RIMS told Global Finance he had not heard of one. Second, whether AI labs disclose coverage or reserves in filings as IPO preparations advance. Third, whether regulators act: CSIS recommends a NIST-run AI incident database, a federal-state working group at the insurance commissioners' association, a staged federal backstop for catastrophic correlated losses, and licensed independent verification organizations. None of those exist yet.
There is also a market-side upside. Global Finance quotes FurtherAI's CEO predicting that by 2029 AI liability will be a core product category for leading carriers. Exclusions today may be the first step toward priced, structured cover tomorrow, which would reward companies that can show disciplined AI governance.
For now the honest position is this: the exclusion trend is documented, the lab-level dollar figures are not, and the people most exposed to the gap are the businesses that deploy AI on top of vendors with limited cover. Treat any specific number, including those in this post, as reported until a company or filing confirms it.
Details reflect sources available on October 17, 2026; check back as renewals and filings land.