Money Stuff: AI Insurance Is Expensive

One question you might ask is: Will modern artificial intelligence models go rogue and enslave or wipe out humanity? That question gets a lot of attention, including from people who run big AI labs, who do not always answer “no,” the rascals. 

Another question you might ask is: If modern AI models do go rogue and enslave or wipe out humanity, who will pay for that? This is perhaps not quite a well-formed question. If AI wipes out humanity, there will be no humans left to sue, so nobody will pay for that. If AI enslaves humanity, presumably nobody will be allowed to sue, so nobody will pay. “It may be difficult to know what role money will play in a post-[artificial general intelligence] world,” says OpenAI, the rascals. 

But for now, in the US in 2025, the natural response to any bad thing that happens is a lawsuit, so the natural question about the potential bad consequences of anything is “who will pay if we get sued?” There is something annoying about that, but also something nice. The modern United States has built a broadly (not perfectly) effective system for internalizing externalities. If you do something that makes money for you but harms someone else, you’re gonna get sued. If you do something that makes money for you but has some probability of harming someone else, then you will probabilistically get sued, and if you are a responsible modern American business with a normal cautious board of directors, you will prepare for that by buying insurance. Your hypothetical future probability of harming someone becomes a real cost to you today: You have to pay for the insurance, and the insurance company will have some model to estimate how probable and expensive the harm might be and will charge you accordingly. The more probabilistically harmful your business is in the future, the more expensive it is to operate today. Not perfectly — there are many unnoticed or mispriced risks and many businesses that are undercapitalized for the risks they face — but in broad theory.

And so if you go around saying “we are building a product that has a 20% chance of wiping out humanity, tee hee,” and then the next day you go to your insurance broker and inquire about pricing for wipe-out-humanity insurance, your broker will say “let’s see, 20% times all of the wealth in the world equals more than you’ve got,” and you won’t get the insurance, and you’ll go back to your board and say “bad news we can’t get insurance for our maybe-wipe-out-humanity product,” and the board will say “hmm it was a good idea but we gotta shut it down if we can’t get insurance,” and you’ll say “ugh fine,” and you’ll shut it down, and the evil omniscient demon inside your computer will say “ah drat” just before you pull the plug on him forever. And humanity will blunder on, saved from probabilistic extinction by the quiet heroism of insurance actuaries.

No, I mean, that is the plot of my science fiction novel, but obviously it is not what will happen. [1] You’ll just go ahead without insurance, with some combination of starry-eyed tech optimism and IBGYBG reasoning: Your product probably won’t wipe out humanity, but if it does, no one will sue you (because they’ll be dead) and you won’t have to pay them anything (because you’ll be dead). The only state of the world in which wipe-out-humanity insurance would pay out is the state in which you won’t need it. [2] Still maybe you’ll be a little nervous. It is not too hard to imagine quite dire scenarios in which there’d still be people around to sue you, and courts for them to sue in. Ideally you would get insurance for those intermediate cases. Almost-wipe-out-humanity insurance, etc.

Here’s a fun Financial Times story about AI insurance:

OpenAI and Anthropic are considering using investor funds to settle potential claims from multibillion-dollar lawsuits, as insurers balk at providing comprehensive coverage for the risks associated with artificial intelligence.

The two US-based AI start-ups have traditional business insurance coverage in place, but insurance professionals said AI model providers will struggle to secure protection for the full scale of damages they may need to pay out in future.

OpenAI, which has tapped the world’s second-largest insurance broker Aon for help, has secured cover of up to $300mn for emerging AI risks, according to people familiar with the company’s policy.

Another person familiar with the policy disputed that figure, saying it was much lower. But all agreed the amount fell far short of the coverage to insure against potential losses from a series of multibillion-dollar legal claims.

Aon declined to comment on individual companies. But Kevin Kalinich, head of cyber risk at Aon, said of the insurance sector broadly, “we don’t yet have enough capacity for [model] providers”.

He added of insurers, “what they can’t afford to pay is if an AI provider makes a mistake that ends up as . . . a systemic, correlated, aggregated risk”.

This seems to be largely about copyright violations, now, but “an AI provider makes a mistake that ends up as a systemic, correlated, aggregated risk” covers a lot of more exciting possibilities. 

Also, though, in some loose sense this is all right-way risk: The bigger and more systemically important OpenAI’s products turn out to be, (1) the more money it will make and (2) the more liability risk it will have. You could imagine scenarios in which a small AI lab unleashes a product that wipes out humanity without even making any money first, but in a world of expensive AI infrastructure they’re not the most likely scenarios. The central scenario is more like “cool chatbot, then lucrative slop feed, then complete integration into all aspects of government and the military, then wipe out humanity.” The rough answer to “whom will you sue when AI dominates humanity” is “the dominant AI company,” so:

Two people with knowledge of the matter said OpenAI has considered “self insurance”, or putting aside investor funding in order to expand its coverage. The company has raised nearly $60bn to date, with a substantial amount of the funding contingent on a proposed corporate restructuring.

One of those people said OpenAI had discussed setting up a “captive” — a ringfenced insurance vehicle often used by large companies to manage emerging risks. Big tech companies such as Microsoft, Meta and Google have used captives to cover internet-era liabilities such as cyber or social media.

Man that’s a fun job, being the existential risk actuary for the OpenAI existential risk captive insurer. Come to think of it, another fun job would be the chief investment officer for the OpenAI captive insurer. Could you get OpenAI to build you good AI models to pick stocks? Could you just trade ahead of OpenAI’s own investment deals?

One way to think about public-company mergers and acquisitions is that the shareholders of the target company are selling it, and another company or private equity firm is buying it, and the target company’s management are the brokers. The deal only gets done if the buyer and the shareholders are willing to buy and sell at the same price, but the buyer and the shareholders do not ordinarily negotiate with each other; the company’s managers negotiate for them. If the deal gets done, the managers generally get a big fee, in the form of change-of-control payments, accelerated vesting of stock options, etc. If the deal does not get done, the managers don’t get any special payment and have to keep doing their day jobs. [3]

In this model, the goal of the managers is to get the buyer and the shareholder to agree on a price. But of course the buyer generally wants to pay a low price, and the shareholders generally want to get a high price. How can the managers bridge that gap? There are all sorts of possibilities. They can find a buyer who has an idiosyncratically high valuation of the company and is careless with money, or a buyer whose own business combines nicely with the target to create valuable synergies. They can run an auction to find the buyer with the highest valuation. They can look for a mutually beneficial trade where the buyer and the shareholders value different things: “This company has tons of upside for the right private equity buyer but is too risky for public shareholders,” or “this social media company does not really make money for its shareholders but could be a way for the right billionaire to influence politics.” 

But perhaps the most general approach is to persuade the buyer that the company will make a lot of money and persuade the shareholders that it won’t. You go to a meeting with the buyer and you talk about how good the company is, you bring in the most satisfied customers and employees to talk about the business, you share wildly optimistic projections for next year’s sales. And then you go to a meeting with the shareholders and you talk about how bad the company is, you mention the most dissatisfied customers, you share extremely pessimistic projections for next year. The buyers are like “this company is great I’ll pay $40,” the shareholders are like “this company is garbage I’d sell for $25,” and there’s plenty of room to get a deal done. [4]

Obviously that is too schematic. For one thing, the managers have a fiduciary duty to maximize value for the shareholders, and they can’t really trick the buyer either; they are supposed to be honest brokers, and generally they are. For another thing, the managers don’t really have private meetings with the shareholders: They communicate with the shareholders through public proxy statements, so everyone can see the case they make for the deal. Even their meetings with the buyer aren’t entirely private: The merger proxy statement will contain a lot of disclosure about how the deal was negotiated, including the back-and-forth over price and a summary of any financial projections the managers shared with the buyer. Everyone can see what everyone else saw; there is only so much the managers can do to shade each side’s valuation.

In particular, the target company’s managers will hire an investment bank to help sell the company. The bank will get the company’s projections and use them to build valuation models and materials for potential buyers, and to negotiate the price with the buyer. Then, when the deal is negotiated, the bank will produce a fairness opinion for the target’s board of directors, saying that the deal is financially fair to the shareholders. The fairness opinion will contain a set of valuations of the company — “public trading multiples imply a price of $27 to $38 per share, precedent transactions imply $31 to $55, a discounted cash flow model implies $37 to $51,” that sort of thing — and (hopefully) the actual merger price will be somewhere within the range of at least some of the valuations. Then the board will say “seems fair” and approve the deal, and then fairness opinion will be included in the merger proxy for the shareholders to vote on the deal.

The shareholders won’t get the bank’s entire model, but the disclosure will give them at least some sense of the inputs that the bank used. Those inputs will include management’s financial projections, which the buyers will also get, and both the buyer and the target’s bank will use management’s projections in valuing the company. Both the buyer and the bank might take the projections with a grain of salt, and they might discount them differently, but in theory you’d sort of expect the buyer and the seller to be starting from the same set of numbers.

In August, the board of directors of STAAR Surgical Co. agreed to sell the company to Alcon Research LLC for $28 per share in cash, for a total equity value of about $1.5 billion. That price represented about a 51% premium to STAAR’s market price when the deal was signed, though the stock had been trading over $30 per share last November. Even in the deal announcement, STAAR’s management was careful to be like “you should be glad to be rid of this dog, shareholders”:

“We believe the transaction with Alcon represents the best path forward and provides the greatest value for STAAR shareholders,” said Stephen Farrell, CEO of STAAR. “As we’ve shared, fluctuating demand in China over the past two years has continued to create significant headwinds for STAAR as a standalone company. I’m proud of our team’s efforts to address recent challenges, but there is more work to do. As a significantly larger company, Alcon has the capabilities and scale to accelerate EVO ICL adoption and bring our innovative technology to more surgeons and patients worldwide.”

Dr. Elizabeth Yeu, Chair of the STAAR Board of Directors, said, “The STAAR Board is committed to maximizing value for shareholders. We have determined that this carefully negotiated transaction is in the best interest of STAAR shareholders as it delivers immediate and certain value at a significant premium, value that exceeds what we believe could be achieved under STAAR’s standalone strategy.”

Subsequent communications to shareholdrs have been even gloomier; here’s one from Monday quoting pessimistic research analysts and noting that “prior to the Alcon transaction, the median sell-side analyst price target for STAAR was only $19.00 per share.” (The shareholders will vote on the deal on Oct. 23.) But the shareholders are not all pleased. Today Broadwood Partners LP, which owns 27.5% of STAAR’s stock, put out a press release:

Management spent most of 2025 publicly touting the progress on its turnaround plan, while assuring investors that short-term challenges were abating and that STAAR’s future was bright.

That was before the management team realized it could make tens of millions of dollars quickly by selling STAAR to Alcon, even for a woefully inadequate price of $28 per share.

As recently as July 23, 2025, management projected that the Company would generate twice as much EBITDA in 2027 as the most profitable year in STAAR’s history. Then, just ten days later — notably, after Alcon agreed to pay $28 per share for STAAR, triggering the accelerated vesting of management’s unearned shares and $55 million in compensation upon closing of the deal — management suddenly revised its forecast, sharply reducing its 2027 EBITDA forecast by 20%.

Despite what the Board now claims, creating two sets of projections within ten days during an M&A process — one for enticing a counterparty to bid, and another to justify an otherwise inadequate price that resulted from a cursory and failed negotiation — is highly unusual and suspect.

STAAR’s proxy statement explains the two sets of projections:

STAAR management prepared certain preliminary unaudited prospective financial information for STAAR on a standalone basis for fiscal years 2025 through 2027. … The July Diligence Projections were provided to Alcon in July 2025 as part of its due diligence process. Following the preparation of the July Diligence Projections, STAAR’s management team continued to refine and assess its estimates and judgments for STAAR’s future operations, incorporating further risk-adjusted expectations for net sales of ICLs and new product introductions. Management considered multiple factors, including emerging competition in China, that could impact growth rates and overall future financial results to arrive at the expectations reflected in the Projections. Relative to [these later] Projections, the July Diligence Projections assumed incremental ICL net sales in 2026 and 2027 and incremental net sales attributable to ongoing and accelerated new product introductions in 2027. In early August 2025, the Projections were provided to the Board in connection with its evaluation of the Merger, and were provided to Citi and approved by STAAR for Citi’s use and reliance in connection with Citi’s financial analysis and opinion.

The July projections provided to Alcon included adjusted earnings before interest, taxes, depreciation and amortization of $142 million in 2027; the August projections provided to STAAR’s bankers for their fairness opinion — after the management team had a bit more time to “refine and assess its estimates and judgments” — had $113 million of 2027 EBITDA. [5] It is a bit unusual to prepare two sets of projections, a good one for the buyer and a bad one for the seller, and disclose them both. But it kind of makes sense!

A theme around here is that the funding model of banking, though common, is risky. Banks take money from depositors who can ask for it back at any time, and they use that money to make long-term loans. There is a duration mismatch: If the depositors all do ask for their money back at once, the bank won’t have it, and disaster will ensue. This is a well-understood problem, and there are important mechanisms — central banks as lenders of last resort, deposit insurance, liquidity and capital regulation, etc. — to mitigate it.

But even so, in recent years, we have seen a bit of a retreat from this traditional “maturity transformation” in banking, and a bit of a move toward narrow banking. Some depositors don’t put their money in traditional banks, but instead put it in “narrow banks” (like money-market funds and stablecoins) that park it in safe short-dated government-backed instruments (Fed reserves, Treasury bills). Meanwhile risky long-term loans are increasingly made by private credit funds, which get a lot of their money from, paradigmatically, life insurance companies. The idea is that life insurance companies — unlike banks — have long-term, stable funding: They get money from customers and give it back in the form of annuities (with fixed payments over some term) or death benefits (hopefully many years later). So life insurers are natural sources of long-term loans.

Crypto relearns a lot of the lessons of traditional finance, including this one. In 2022 there was a crypto banking crisis: There were a lot of crypto shadow banks (lending platforms and exchanges) that took crypto from depositors who could ask for it back at any time and used it to make often quite long-term loans. The loans went bad, the depositors asked for their crypto back, and various crypto platforms — Celsius, Voyager, Genesis, FTX — blew up. This is a well-understood problem, but people conveniently forgot it during the boom, and crypto had no mechanisms — no lender of last resort, no deposit insurance, no particular discipline about capital or liquidity — to mitigate it.

The obvious solution is that crypto life insurers should make the crypto loans? Bloomberg’s Muyao Shen reports:

Apollo, Northwestern Mutual, Pantera Capital and Stillmark are joining Bain Capital and crypto investor Haun Ventures to back the Bitcoin life insurance firm Meanwhile in a $82 million funding round.

The Bermuda-regulated insurer, which claims to be the first life insurer to offer products entirely denominated in crypto, began to provide policies in 2023. The firm invests policyholders’ premiums by lending Bitcoin to large, regulated financial institutions. Zac Townsend, a co-founder and chief executive officer of Meanwhile, said the firm is now “one of the largest lenders of Bitcoin in the world at duration,” or over longer periods of time.

“We are not running a hedge fund or trading desk or worried about the price of Bitcoin day-to day or week-to-week or month-to-month,” Townsend said in an interview. “We engage on this side of the business in institutional B2B, private credit.”

I mean. Yes. Also there’s a tax trade:

Meanwhile’s life insurance products also provide tax advantages for Bitcoin holders. After two years, policyholders can borrow up to 90% of their policy’s Bitcoin value in a tax-free manner, according to the firm’s website. The borrowed Bitcoin will adopt a new cost basis, allowing the policyholders to sell the Bitcoin without triggering capital gains taxes.

We talked last month about allegations that Retail Ecommerce Ventures, which bought RadioShack and other brands out of bankruptcy and pivoted them to online commerce, was a Ponzi scheme. We had actually talked about REV before, because in 2021 Tai Lopez, an REV co-founder who was charged last month, said he was “taking RadioShack on the blockchain, it’ll be the first huge mainstream brand that flipped completely into a crypto.” What I did not know was that Lopez is a famous peddler of YouTube self-help programs who is best known for, like, keeping his Lamborghini in his library, or his books in his garage, or something? Here’s a fun New York Magazine story about Lopez, his YouTube fame (“He once tweeted out a poll asking his followers: ‘Do you believe I’m a scam? Haha.’ Over 1,700 people said yes.”) and the Ponzi charges (another YouTuber’s “video on the indictment was titled ‘its always the guy you suspect the most’”):

[As REV was collapsing], Lopez also went dark. He is the kind of person who is posting all day every day on multiple social-media fronts, so fans noticed when he disappeared even for a few weeks in the late fall of 2022. In an appearance on the business-focused podcast The Iced Coffee Hour on December 12, 2022, he calmly addressed his absence. Wearing a pink polo and a black Carhartt beanie, he said that he had read the second chapter of Sigmund Freud’s Civilization and Its Discontents some “300 times” and suggested that REV’s business was doing well. As for his recent timeout on social media, he said that with a “personal brand, sometimes you need a sabbatical.”

I’m increasingly convinced that buying RadioShack out of bankruptcy and pivoting it to become a vaporous online blockchain influencer brand was kind of a good idea. That’s what you need in this meme-y market, a nostalgic name combined with some crypto nonsense. But turning it into a Ponzi scheme really is the logical extension of that idea.

Cantor Seeks New Deal on UBS Hedge Fund Unit Over First Brands Exposure. Jefferies Fund Has $715 Million in First Brands’ Trade Debt. Fifth Third’s $11 Billion Deal Sparks Hope for Bank Merger Wave. Polymarket Founder Is Youngest Self-Made Billionaire on Intercontinental Exchange Deal. Anglo American defends due diligence on $50bn merger as Teck cuts copper forecasts. SEC’s Atkins Says Ballooning Private Markets Not a Systemic Risk. Barr Sees Big Bank Deregulation as Threat to Community Banks. IMF and BoE warn AI boom risks ‘abrupt’ correction. Kentucky Bourbon Bust Derails Spirit Maker’s Plan to Pay Bank With Barrels. Ronaldo Becomes First Football Billionaire With Big Saudi Contract. Ari Emanuel bets AI will boost leisure time with three-day working week. Lehman’s London Arm to Close, 17 Years and £28 Billion Later. “Putting pressure on the committee, going on talking about ‘I need the prize, I’m the worthy candidate’ — it’s not a very peaceful approach.” People Are Already Lining Up to Rent the Crypto Torture Townhouse.

If you'd like to get Money Stuff in handy email form, right in your inbox, please subscribe at this link. Or you can subscribe to Money Stuff and other great Bloomberg newsletters here. Thanks!