Like a lot of investors this week, I spent an evening with Anthropic's S-1. I'd opened it for the financials. I stayed for the risk factors, which run 94 pages and read better than many dystopian novels. Aside from the publicity value, I'm not sure that's the effect the securities lawyers were going for.
We've all heard the headline risk: a company asking the public for money while warning, in writing, that its product could pose an existential risk to humanity. On Bloomberg Surveillance this morning, Tom Keene was trying to recall an old line from the critic Fredric Jameson, that it's easier to imagine the end of the world than the end of capitalism. Anthropic managed both at once. It imagined the end of the world in a document built to raise capital.
Once you get past the apocalypse (hey, it's better than the demise of capitalism?), the S-1 turns oddly practical. There are pages on compute suppliers and copyright suits and governance quirks. Somewhere in the middle sits a paragraph about insurance.
I've written before about how actuaries, the people everyone assumes are boring, quietly decide when a technology goes mainstream. Fire underwriters nervous about Edison's wiring gave us Underwriters Laboratories. Cyber insurers got multifactor authentication deployed in about two years after security teams had begged for it for a decade. I've also argued that insurance pricing, more than any law, will eventually move us out of the driver's seat with self-driving vehicles.
The uninsurable part
Here's the gist of the filing: a catastrophic AI incident could cause harm so large that no single insurer, and possibly not the industry as a whole, would have the capacity to underwrite it. Anthropic calls the market for AI-specific coverage nascent, and it's already felt the gap. Its insurance didn't fully cover the $1.5 billion Bartz copyright settlement.
Size, oddly, is the manageable part. Aon's Kevin Kalinich told the Financial Times the industry could handle a single $400 million loss but would struggle with 1,000 or 10,000 correlated claims traced to one AI provider's mistake.
Any engineer will recognize that one. It's common-mode failure. I can install two redundant pumps, but if both hang off the same power supply, I own one pump. Insurance depends on the same kind of independence, so when 50,000 companies route customer service and contract review through the same model, a single flaw can hit all of them on the same afternoon. To an underwriter, that's one policy with 50,000 addresses. We got a preview in July 2024, when a bad CrowdStrike update knocked roughly 8.5 million Windows machines offline. One file. Everywhere.
The throttle
Insurers are doing what they always do with a risk they can't price yet: exclude it. Since January, Verisk/ISO endorsements have let carriers strip generative AI out of commercial general liability policies, and at least six major carriers have filed their own exclusions with state regulators. Berkley's is the bluntest, removing D&O and E&O coverage for any claim tied to anyone's use or development of AI. Anyone's.
The exclusions land on the companies deploying AI, and the vendor contracts lean the same way; Gallagher Re concluded that AI vendor terms leave deployers holding most of the liability. Picture the meeting. Operations wants to roll an AI agent out across customer service, and the CFO learns the general liability and D&O policies now carve that exact activity out. Having sat in the CFO chair, I know which side of that table wins. No statute slows adoption that fast.
The accelerator
In an engine, the throttle and the accelerator are the same valve. Insurance works that way too: the mechanism that blocks one deployment can clear the next.
My favorite example is a startup I hadn't heard of until this week. The Artificial Intelligence Underwriting Company, founded by Rune Kvist (a former Anthropic executive) and Rajiv Dattani, built a standard called AIUC-1 and modeled it on SOC 2. Auditors red-team a specific AI agent, and an agent that passes can be insured. ElevenLabs ran more than 5,000 adversarial simulations before it got certified and covered for its voice agents. In September AIUC raised $40 million to start auditing frontier models.
If that structure sounds familiar, it should. UL, Underwriters Laboratories, began with fire insurers who couldn't price electrical risk, so they paid someone to test it. No legislature created it. More than a century later its mark sits on the power strip under my desk, and I trust it. A certified agent with a policy behind it is something a risk committee can approve. Coverage turns an unknown into a line item.
I think Geoffrey Fehling, a coverage lawyer at Hunton Andrews Kurth, has the trajectory right. He argues that cyber insurance spent years in exclusions before a dedicated market formed, and that AI is on the same road. Cyber gave us MFA as a condition of coverage. I'd bet AI gives us agent certification the same way.
Why underwriters may beat regulators
Government rules get written once, argued over for years, and revisited when the next crisis forces the issue. Underwriters reprice at every renewal, and when they're wrong, they lose their own money. That's a feedback loop measured in months. AIUC says it updates its standard quarterly; I can't think of an AI regulation anywhere that moves at that speed. Regulators need consensus. Insurers need loss ratios.
What insurance can't touch
The S-1 paragraph also marks the limit, and we've been here before. In 1957 a federal study estimated a worst-case reactor accident could cause up to $7 billion in property damage, while private insurers would cover about $60 million. Private capacity covered less than 1% of the estimated worst case. Congress passed the Price-Anderson Act, with the government committing $500 million beyond the private layer. It got reactors built. It also took the insurers' veto off the table.
If AI follows that path, insurers will end up governing everyday risk, the chatbot that invents a discount or the agent that leaks a file, while the catastrophic tail goes to taxpayers or to nobody.
So the actuaries may write better rules for ordinary AI than Washington will, and faster. Who writes them for the tail? When the people paid to price risk say they can't, is that Anthropic's business problem, or the most honest data point we have about the technology? Now there's a plot twist better than anything in those dystopian novels.