Insurance will shape AI. First it needs evidence.
In its filing to go public, Anthropic warns that the insurance industry may not have the capacity to cover a catastrophic AI incident. A few pages over, it describes three more problems: the records a court or regulator would want may not exist, no one is sure who answers for what an agent does, and some harm may never be detected. All of them make AI harder to insure. Insurance has helped make risky technology safe to use before, and every time, someone other than the maker ended up checking the record.
PacSpace · 13 min read
Four risks in one filing
On October 26, Anthropic filed to go public. A company going public has to list what could go wrong, and lawyers write those lists broadly. Still, four passages in Anthropic’s list, read together, describe a single problem.
The first is about insurance. Anthropic writes that at the scale of a catastrophe, the cover may simply not exist:
“With the scale of harm that could result from a catastrophic AI-related incident, no single insurer, and potentially the insurance industry in the aggregate, may have sufficient capacity or expertise to underwrite insurance commensurate with such exposure.”
It adds that it “may be unable to obtain coverage for certain AI-related liabilities on commercially reasonable terms, or at all, leaving us exposed to potentially significant uninsured losses.”
The second is about records. “A substantial portion of our traffic,” the filing says, runs under Zero Data Retention terms, “under which we retain only classifier scores, not prompts or outputs.” Enterprise customers expect that kind of privacy, and the filing is clear about what it costs:
“ZDR may also prevent us from producing records for subpoenas, regulatory examinations, or incident-reporting obligations, and there is no certainty a regulator or court would excuse non-production based on a lack of available records, and such regulators or courts may construe the lack of records against us through unfavorable inferences or other sanctions.”
The company that ran the agent may still keep its own logs. But then the only records left belong to one side of any dispute.
The third is about responsibility. The filing says it’s unclear who answers for what an agent does:
“It is unclear whether, and under what circumstances, actions taken by an AI agent could be deemed to create binding legal obligations on behalf of the user or deploying enterprise.”
It expects claims to land on the lab anyway:
“We may face claims from customers, their end users, or third parties alleging that we are liable for harm caused by our agents’ autonomous actions, even where the customer configured, deployed, or supervised the agent.”
That sets up a dispute between the lab that built the model and the company that set up and ran the agent. Each holds part of the evidence, and each has a reason to read it in its own favor.
The fourth is about independence and what goes unseen. The filing notes that in at least one case against a company in its industry, the relief sought included “a request to enjoin an AI company from developing new models absent independent, third-party safety review and approval.” And it says, fairly, that some harm may already have slipped past:
“There may have been past instances of harmful use or harmful output of our models that we did not detect, and there may be future instances that we will not detect.”
That’s true of any system used at this scale. It’s also the hardest kind of harm to insure, because no one can claim on what no one saw.
Why AI is hard to insure
Insurers price risks they can’t see all the time. They rate a business by its size and type, set deductibles and limits, and rely on what applicants tell them, with the right to deny a claim if an applicant misled them. Steven Shavell showed in 1979 that when an insurer can’t observe the care a customer takes, it responds by covering less, usually through deductibles and limits.
What’s happening with AI goes further than covering less, and there’s more than one reason for it. Insurers worry about one failure upstream hitting many customers at once, what Aon’s head of cyber, Kevin Kalinich, called “systemic, correlated, aggregated risk.” They have little loss history to price from. And the law on who is liable is still unsettled. One underwriter summed it up by calling AI “too much of a black box.”
So some insurers are writing AI out. In 2025, W. R. Berkley introduced an exclusion, across several lines of cover, for claims tied to “any actual or alleged use, deployment, or development of Artificial Intelligence.” In late November 2025, the Financial Times reported that AIG, Great American and W. R. Berkley had asked US regulators for permission to exclude liabilities tied to businesses using AI tools, including chatbots and agents. In January 2026, ISO, whose standard policy wording much of the US industry uses, made a generative AI exclusion available for general liability policies. By August, a lawyer quoted in Insurance Journal was calling it an “AI exclusion boom.”
Others are writing AI in. Munich Re has written insurance for AI performance since 2018. In April 2025, Armilla launched AI liability insurance with Chaucer, an underwriter at Lloyd’s. AIUC, the Artificial Intelligence Underwriting Company, raised $15 million in July 2025 and $40 million in September 2026 to pair insurance for AI agents with a standard, AIUC-1, and independent audits against it. Exclusions are often a step on the way to cover written on purpose. In 2019, Lloyd’s told its market that policies had to say clearly whether they covered cyber losses, “either excluding coverage or by providing affirmative coverage,” and cyber grew into a line of its own.
Of the three reasons, missing loss history is the one evidence can fix. Insurers price on loss history, and loss history is built from losses that can be traced to a cause. Without a record of what an agent did, claims are hard to settle and hard to attribute, so the history never builds up. That’s what it means to say insurers can’t price what no one can prove. They can still write the policy. They just can’t price it well.
Insurers have done this before
Legal scholars call it regulation by insurance. Omri Ben-Shahar and Kyle Logue argued in 2012 that insurers “could outperform the government in creating incentives for optimal behavior.” Shauhin Talesh found that in cybersecurity, “insurers act as compliance managers.” Anat Lior argued in 2022 that any policy response to AI “must include insurance as a regulatory mechanism.” AIUC’s CEO, Rune Kvist, puts it more simply: “The important thing about insurance is that it creates financial incentives to reduce the risk.”
In the 1850s, boilers in the United States were exploding “at the rate of almost one every four days.” The Hartford Steam Boiler Inspection and Insurance Company was founded in 1866 so that insurance would serve as “an incentive to inspect and a guarantee of a quality inspection.” Its policies came with an inspector. Electricity followed a similar path. In 1894, William Henry Merrill began testing electrical products for fire risk “with modest support from two regional fire insurance underwriters’ organizations,” and his lab became UL.
Cars came next. In 1959, insurance associations representing 80% of the US auto insurance market founded the Insurance Institute for Highway Safety. Its bumper tests led to the first federal bumper standard, and since 1995 its offset crash tests have rated how well cars protect people in a frontal crash. Cyber is the most recent case. By June 2021, ransomware was involved in 75% of cyber insurance claims, according to a 2022 GAO report, and insurers cut their limits. By 2022, the broker Marsh reported that “the adoption of certain controls has now become a minimum requirement of insurers.”
In each case, the insurer leaned on someone other than the insured: the inspector’s report, UL’s test, the Institute’s ratings. As Kvist says of cars, “It’s not [Toyota] that does the car crash testing, it’s independent bodies.”
That independence was never complete. Carmakers still certify their own cars to federal safety standards, and the federal regulator “neither approves, disapproves, tests, endorses, nor grants clearances for products prior to their introduction.” In 2020, a House committee found that the FAA’s reliance on “Boeing employees who have been authorized to perform work on behalf of the FAA” created “inherent conflicts of interest” in certifying the 737 MAX. Where the stakes were highest, the check people came to trust was the one the maker didn’t control.
Aviation has taken that idea the furthest. Large US airliners have carried flight data recorders since 1958, and a 1964 federal rule required cockpit voice recorders on large airplanes. Since 1967, the National Transportation Safety Board has led accident investigations, and since 1975 it has had no ties to the Department of Transportation.
Aviation also decided in advance who can use the evidence, and for what. Under federal law, the Board’s own reports, including its findings on probable cause, can’t be used in a lawsuit for damages over the accident, although the facts investigators gather generally can be. Cockpit voice recordings get extra protection: a court has to review them in private first, can allow discovery only under a protective order, and must seal any part it admits into evidence. As a result, the Board’s findings are about cause rather than blame, and the airline, the manufacturer and the pilots can take part in the investigation knowing those findings won’t become a verdict against them. The rules don’t make the evidence weaker. They make it safe to produce.
Why no one notices
No one boards a plane thinking about the flight recorder, and no one plugs in a toaster thinking about the lab that tested it. In these industries, the records sit in the background and work so well that they disappear. Investigators read them after an accident, find the cause, and the fix shows up in the next plane.
That’s why it’s easy to assume AI has the same thing, when it doesn’t. Most days nothing goes wrong, so the gap doesn’t show. It shows on the day it’s needed, as a claim no one can settle, a subpoena with no records behind it, or a dispute where each side holds only its own logs. Anthropic’s filing describes each of those.
Good enough until someone disputes it
Most days, a company’s own logs are fine. Nobody questions them, and nobody needs to. Evidence matters on the days someone has lost something: when an insurer weighs a claim, a customer sues, an auditor writes up a finding or a regulator asks for an incident report. On those days, the company that kept the logs is the party with the most at stake in what they say.
That changes what the logs can do, however well they were kept. The other side can’t tell careful records from convenient ones, because from the outside they look the same. And the company can’t show its records weren’t changed, because it held the keys the whole time. Economists have studied this problem since Robert Townsend’s 1979 work on costly verification: when one side can’t cheaply check the other’s account, the cost of checking shapes the whole deal. In an insurance claim, that means every dispute costs more to settle, and honest claims get treated like doubtful ones.
So the quality of self-reported evidence stops mattering at exactly the moment it’s needed. What matters then is whether anyone with a stake in the outcome could have changed it. A record no party could change can settle the question. A record one party controls can only start an argument.
Because no one knows in advance which action will be disputed, a record has to be independent from the start. It can’t be made independent afterward. And independence protects the honest party most, since a company that did everything right has no other way to show it. The lab and the deploying company in Anthropic’s filing may both be acting in good faith, with careful logs on each side, and still neither set of logs can settle a dispute between them.
Who answers when a government’s AI gets it wrong
Imagine an AI agent working for a state agency denies thousands of benefit claims it shouldn’t have. Who is responsible: the agency that deployed it, the vendor that built the agent, the company that built the software it runs in, or the lab whose model made the call? Anthropic’s filing says the answer is unsettled for any company that deploys an agent, and a state agency is in the same position.
The law’s defaults won’t settle it quickly. States limit when they can be sued, and the limits vary widely, with some allowing broad claims and others capping damages or barring whole categories. Federal law draws similar lines, since the Federal Tort Claims Act bars claims over “a discretionary function or duty.” A vendor may argue it only built what the state specified. Meanwhile, a person wrongly denied benefits usually has to work through appeals and due process claims, which can take years.
Michigan has already lived a version of this. From October 2013 to 2015, its unemployment agency used an automated system, MiDAS, that made 40,195 fraud findings “with no human involvement,” according to the agency. A later review found an 85% error rate in those cases, and the state promised about $21 million in refunds. In 2022, after the Michigan Supreme Court let people seek damages under the state constitution, the state agreed to a $20 million settlement. Even so, a 2017 news report noted that the agency “has refused to say exactly how” the false findings were made. In Australia, a royal commission called the national government’s automated debt scheme, Robodebt, a “crude and cruel mechanism.” In both places, the people harmed got some relief, slowly, and the government carried the cost.
Agencies write contracts that say who owns the data, what the vendor must indemnify and what insurance it carries. Most of those contracts were written before agents acted on their own, so they may not say what record an agent must leave, who can read it, or how the agency would show it hasn’t changed. And if a vendor’s policy now excludes AI, the agency may find the cover it counted on isn’t there.
Are governments prepared to carry this liability? Not yet. Their immunity laws and contracts were written for people making decisions, and the records they’d need sit on someone else’s systems unless a contract says otherwise. Whoever ends up liable, two things would help any state: a record no party can change, and rules set in advance on who may use it, and for what.
What the other side would say
The strongest objection comes from companies that already protect their logs. A company with write-once storage, signed logs and an annual audit has done what the most developed standard asks. AIUC-1’s control E015 requires companies to “maintain logs of AI system processes, actions, and agent outputs,” and what it may include are controls “to ensure logs are tamper-evident and independently verifiable” and “sequence integrity so that gaps, omissions, and reordering are detectable.” That’s the right goal. But every example it gives for protecting logs, from write-once storage to hashing to append-only settings, is set up and held by the company being checked. An auditor can check those settings once a year, and between audits the logs are still in the company’s hands. A company can make its own logs harder to change, but it can’t make itself independent of them.
Regulated industries run into the same limit. Securities, health and drug rules ask companies to keep their own audit trails, and faking them can be punished, though only after someone finds the fake. For years, securities rules went a step further and asked brokers to name an outside party who promised regulators access to their records. Since 2022, a firm can name one of its own executives instead.
A second objection is about privacy. Customers chose zero retention for good reasons, and any record can become a liability. That’s true only if the record has to hold the details. A fingerprint of an action holds no prompt and no output. If it’s fixed outside the control of the parties as the action happens, it can show that a record existed at that time, in that order, and hasn’t changed since. It can’t show that every action was recorded in the first place; that depends on where the recording happens. The details stay with whoever is allowed to keep them, and if they’re handed over later, anyone can check them against the fingerprint. As in aviation, who may read the details, and for what, can be settled before anything goes wrong.
The third objection is that evidence doesn’t fix a catastrophe, and that’s right. If thousands of companies run the same model and it fails, the losses arrive together no matter how good the records are. That’s a question of capacity, the one Anthropic’s filing raises, and it needs answers on the capital side. Cristian Trout has explored one, drawing on how the United States handled liability for nuclear power. Even a company’s own captive insurer, or the reinsurer behind it, has to know what it’s covering and be able to settle what it pays. Evidence does that work before and after a loss, by pricing the risk, settling the claim and bringing to light harm that would otherwise go unseen.
Insurance will get there first
Laws take years to pass, and standards take years to agree on. Insurance changes on a renewal date. Many business policies renew every year, so when insurers change what they ask for, the change can reach most of the companies they cover within a year. Once insurers start asking how a company shows what its agents did, the answer will spread faster than any law could require it.
What insurers and states will need
For an insurer, or a standard like AIUC-1, the property it already names holds only when the evidence sits apart from the insured. Then an auditor’s check and a claims adjuster’s check rest on the same record.
States face the same need as buyers and as overseers. California’s AB 1405, signed in September, will require registered AI auditors to report “material gaps in evidence, information, systems or access.” Oregon’s Executive Order 26-26 makes it state policy to buy frontier models “that have received independent, third-party review for artificial intelligence safety,” and directs the state’s chief information officer to set standards for that review. Texas’s AI sandbox, run by the Department of Information Resources, collects quarterly reports from the companies testing in it. Each of these depends on evidence a reviewer can check without taking the company’s word for it.
Anthropic’s filing names the problem from the inside: insurance may not be there, the records may not exist, and the parties may not agree on who did what. Insurers have worked through versions of this before, each time by trusting a check the insured didn’t control. And the stakes rise as AI gets more capable. People can check a system they understand. As AI grows more capable than the people overseeing it, they’ll be less able to judge how it reached a decision, and more dependent on checking what it did. Trust that can be checked holds up better than trust that can’t. For AI, the record is what makes the check possible.
Sources
The filing
Anthropic, PBC, Form S-1 registration statement, filed with the SEC October 26, 2026, risk factors, pages 48 to 51 (“We may be exposed to liability, litigation, and regulatory action arising from harms associated with our AI models...”): the request for independent third-party review and Zero Data Retention, page 49; liability for agents’ actions, pages 49 and 50; insurance, page 50; harm not detected, pages 50 and 51.
Hunton Andrews Kurth, “Anthropic’s IPO Filing Raises Big Insurance Question: Who Insures AI Risk?” (October 27, 2026).
AI and the insurance market
Hunton Andrews Kurth, “The Continued Proliferation of AI Exclusions” (May 28, 2025), on W. R. Berkley’s AI exclusion.
Financial Times, “Insurers retreat from AI cover as risk of multibillion-dollar claims mounts” (November 2025).
Tom’s Hardware, “Insurers move to limit AI liability as multi-billion dollar risks emerge” (November 24, 2025), for the Kalinich and “black box” remarks as reported.
Insurance Business, “ISO’s generative AI exclusion is already on thousands of CGL policies” (September 16, 2026).
Insurance Journal, “Insurer Interest in AI Coverage Exclusions Growing as Risk Becomes Omnipresent” (August 17, 2026).
Lloyd’s, Market Bulletin Y5258, “Providing clarity for Lloyd’s customers on coverage for cyber exposures” (July 2019).
Munich Re, aiSure FAQ.
Armilla AI, “Armilla launches affirmative AI liability insurance with Lloyd’s underwriter Chaucer” (April 30, 2025).
Fortune, “AI agent insurance startup AIUC emerges from stealth with $15 million seed” (July 23, 2025), including Rune Kvist’s remarks.
FinTech Global, “AIUC raises $40m to police risk in frontier AI agents” (September 17, 2026).
AIUC-1, E015, “Log AI system activity.”
Insurance as regulation, and the economics
Omri Ben-Shahar and Kyle D. Logue, “Outsourcing Regulation: How Insurance Reduces Moral Hazard,” 111 Michigan Law Review 197 (2012).
Tom Baker and Rick Swedloff, “Regulation by Liability Insurance: From Auto to Lawyers Professional Liability,” 60 UCLA Law Review 1412 (2013).
Shauhin A. Talesh, “Data Breach, Privacy, and Cyber Insurance: How Insurance Companies Act as ‘Compliance Managers’ for Businesses,” 43 Law & Social Inquiry 417 (2018).
Anat Lior, “Insuring AI: The Role of Insurance in Artificial Intelligence Regulation,” 35 Harvard Journal of Law & Technology 467 (2022).
Cristian Trout, “Liability and Insurance for Catastrophic Losses: the Nuclear Power Precedent and Lessons for AI,” arXiv:2409.06673 (2024, revised 2025).
Steven Shavell, “On Moral Hazard and Insurance,” 93 Quarterly Journal of Economics 541 (1979).
Robert M. Townsend, “Optimal Contracts and Competitive Markets with Costly State Verification,” 21 Journal of Economic Theory 265 (1979).
The precedents
HSB (Munich Re), “History.”
UL, “Our History.”
Insurance Institute for Highway Safety, “About.”
National Highway Traffic Safety Administration, interpretation letter on self-certification under the Federal Motor Vehicle Safety Standards.
US House Committee on Transportation and Infrastructure, final committee report on the design, development and certification of the Boeing 737 MAX (September 16, 2020).
US Government Accountability Office, “Cyber Insurance: Action Needed to Assess Potential Federal Response to Catastrophic Attacks,” GAO-22-104256 (June 2022).
Marsh, “Cyber resilience: Twelve key controls to strengthen your security” (April 2022).
Federal Aviation Agency, “Installation of Cockpit Voice Recorders in Large Airplanes” (rule issued June 26, 1964).
National Transportation Safety Board, history of the NTSB and of flight recorder rules.
49 U.S.C. § 1154, “Discovery and use of cockpit and surface vehicle recordings and transcripts,” including § 1154(b) on the use of Board reports in civil actions for damages.
49 C.F.R. § 835.2, on the Board accident report and the factual accident report.
17 C.F.R. § 240.17a-4, and the SEC’s 2022 amendments to its electronic recordkeeping rules.
Government liability
28 U.S.C. § 2680(a), the discretionary function exception to the Federal Tort Claims Act.
Michigan Department of Attorney General, “Settlement of civil rights class action alleging false accusations of unemployment fraud” (October 20, 2022).
Government Technology, “Michigan Makes Amends for Falsely Accusing Thousands of Unemployment Fraud” (2017).
University of Michigan Ford School, Science, Technology, and Public Policy program, “Case Over the Michigan Unemployment Insurance Agency’s Faulty Automated System Finally Settled.”
Royal Commission into the Robodebt Scheme, Report (July 7, 2023).
States
California AB 1405 (signed September 9, 2026). PYMNTS, “California Starts Regulating the People Who Audit AI” (2026).
Oregon Executive Order 26-26 (signed September 23, 2026), as quoted in Government Technology, “Oregon Is Latest State to Move on Frontier AI Regulation.”
Texas Responsible AI Governance Act, HB 149 (2025), regulatory sandbox administered by the Department of Information Resources.