All walkthroughs · Paper briefs

When Does Regulation by Insurance Work? The Case of Frontier AI

Cristian Trout (Artificial Intelligence Underwriting Company)
arXiv 2512.06597, version 3 (2 March 2026); first posted on SSRN in October 2025 and on arXiv in December 2025

Paper · Brief and rating · Dashboard entry

About 22 minutes · Download MP3

AI-generated walkthrough with synthesised voices: a host explains the paper and a discussant questions it. It covers claims, methods, key assumptions, citations and use so far, decision relevance, and what an evaluator should check. Every direct quotation was checked word for word against its source. Public sources only. Not an Unjournal evaluation or decision.

Transcript

Host
This is a walkthrough of When Does Regulation by Insurance Work? The Case of Frontier AI, by Cristian Trout. It is a single-author article in law review format, about ninety pages, posted to SSRN in October twenty twenty-five and to arXiv in December, with the version we are using, version three, dated March twenty twenty-six. The author is a research fellow at the Artificial Intelligence Underwriting Company, an AI insurance startup. This is an AI-generated walkthrough, not an Unjournal evaluation or decision.
Discussant
Start with the question. Economists already know insurance creates moral hazard. What is the paper adding?
Host
It takes a long-running debate in law and economics, the so-called regulation thesis, and tries to give it structure. The thesis is that insurers can act as private regulators: they price risk, monitor, impose safety conditions, fund safety research, and sometimes lobby for stricter rules. Proponents collect success stories, such as auto insurers funding crash testing. Skeptics collect failures, such as cyber insurers whose lawyers help clients limit liability rather than harm. Trout calls the outcome he cares about a net regulatory effect: insurance uptake moves care levels closer to the efficient level than the baseline of liability and markets without insurance, even after counting moral hazard. His contribution is a two-step framework. Step one asks which distortions exist in the uninsured baseline. That sets the potential for a net regulatory effect. Step two asks whether conditions let insurers realize that potential, which mostly comes down to whether they can contain moral hazard. He then applies it to frontier AI developers and ends with a policy proposal.
Discussant
And the headline policy claim?
Host
That the potential is real for catastrophic risks from frontier AI, but it will not be realized by default. He argues for a narrow, deep and restrictive insurance mandate: only the handful of developers at the frontier, covering only named catastrophic risks, with large required limits, no exclusions for things like chemical, biological, radiological and nuclear harms or critical infrastructure failure, and no use of pure captives, meaning a firm's own subsidiary insurer. He expects such a mandate to produce a single specialized insurer, ideally a mutual owned by the AI developers, or failing that a joint underwriting association, which is a regulator-sanctioned pool of insurers. He contrasts this with a broad, shallow and permissive mandate covering many firms and routine harms at low limits, which he thinks would do little.
Discussant
You called it a framework. Is there a formal model?
Host
No. Part One is labelled a model, but it is conceptual. Two figures show care levels under an idealized baseline, then background distortions push care away from the efficient level, moral hazard pushes it down further, and insurers' loss prevention pushes it back. The net effect is the sum. The central logical point is stated plainly. The author writes: "a net regulatory effect only occurs when insurers' incentives are better aligned with welfare maximization than those of other economic actors." So if there are no background distortions, insurance can only make care worse, apart from risk-spreading gains he sets aside. The paper is explicit that it does not quantify. In his words: "Ultimately, the working social engineer must rely on their judgment."
Discussant
So the work is in the catalogue of distortions. What is in it?
Host
Six families. First, judgment-proofness: firms that cannot pay a large judgment under-invest in care, and mandatory insurance with risk-priced premiums draws expected losses forward. He cites Boomhower's work on oil and gas drillers as evidence that both the problem and the mandate remedy are real. Second, competition. Here he reviews industrial organization models of product safety. For product accidents, where victims are customers, consumers partly price safety into willingness to pay, so liability and therefore insurance have less to do. For non-product accidents, where victims are third parties, there is more room, especially when safety research is a fixed, upfront cost with large spillovers, or when firms see lower returns on safety than on other investments. Third, firm survival motives, drawing on Evans and Gilpatric: when a firm's main insolvency risk comes from poor profits rather than liability, it cuts care below even the judgment-proof prediction. Fourth, charity hazard: the expectation of a government bailout or disaster relief. Fifth, behavioural biases around low-probability, high-consequence events: probability weighting near zero, optimism, overconfidence, availability. Sixth, managerial short-termism.
Discussant
Why would insurers correct the competition distortions? Premiums that are actuarially fair leave the firm's trade-offs unchanged.
Host
He concedes exactly that. If insurers only adjusted premiums, insurance would not touch first-mover distortions. His argument is that insurers also invest directly: risk models, private safety codes, safety research, lobbying. And their objective differs from the firms'. The author writes: "Insurers don't care which policyholder wins in a competition." A concentrated insurer covering all the competitors internalizes more of the spillovers from safety research. But he limits the claim. He says: "insurance uptake here certainly can't guarantee efficient outcomes, but can help guarantee a safety floor, especially for catastrophic risks." That floor-not-ceiling language recurs through the paper.
Discussant
Now the application. Are frontier AI firms judgment-proof?
Host
He works with a simple benchmark. The largest commercial liability programmes reach the low tens of billions of dollars, so he asks whether a firm would be insolvent after roughly twenty billion dollars of liability, an amount insurers could plausibly cover. Using mid twenty twenty-five figures, Alphabet, Microsoft, Meta and Amazon are clearly not judgment-proof at that scale. Smaller startups, such as Mistral or Cohere, plausibly are. OpenAI and Anthropic are in between: they lack the cash but might raise it, unless investors flee. His conclusion, which he says surprised him, is that "judgment-proofness is not where the most potential for a regulatory effect lies." This partly reverses his own earlier work and Gabriel Weil's, which treated judgment-proofness as the main distortion. He offers an intuition pump: would the chief executives of OpenAI or xAI behave differently if forced to hold tens of billions in cash against accidents? He thinks not, so something else is driving behaviour.
Discussant
Which distortions does he think dominate?
Host
Four. First, the race for market share. Valuations at many multiples of revenue and leaders' own statements framing artificial general intelligence, meaning systems matching humans across most cognitive work, in existential terms, suggest strong first-mover incentives. Safety work that improves the product, such as reducing hallucinations, where a language model states fabricated claims as fact, is rewarded by the market. Work on rare catastrophic risks is not. The author writes: "Precisely because tail risks are rare and highly uncertain, investments in mitigating them do not clearly translate into a substantial or reliable competitive advantage." As evidence he cites press reports of rushed dangerous-capability evaluations, OpenAI disbanding its superalignment team, and the Grace and co-authors survey of machine learning researchers, where about seventy percent said safety research should be prioritized more or much more.
Discussant
That survey asks researchers what they would like prioritized. It does not measure under-investment against an efficient benchmark.
Host
Agreed, and he uses it as suggestive. The second distortion is the spillover and fixed-cost character of safety research. He argues that fields such as AI control, which studies how to use capable but untrusted models safely, dangerous-capability evaluation, and mechanistic interpretability, which tries to reverse-engineer what a neural network is computing internally, are young enough to be primary research with large spillovers. A single mandated insurer could internalize those spillovers, acting as a forced joint venture on safety research. Third, shared reputation. His analogy is Three Mile Island. After that accident, US nuclear operators realized a careless peer damaged them all, formed the Institute of Nuclear Power Operations to inspect and rate plants, and tied it to their mutual insurer, Nuclear Electric Insurance Limited, which requires membership. He adds the unilateralist's curse: with more well-meaning but uncoordinated actors, the chance that one of them releases something harmful rises. His conclusion: "The insurer becomes the centralized body for industry self-regulation." Fourth, psychology: a young industry of venture-backed founders selected for risk tolerance and overconfidence, and a winner's curse in which the firm racing fastest is likely the one that most underestimates the risk. Insurers, by contrast, are old repeat players with solvency regulation that forces them to plan for one-in-two-hundred-year events.
Discussant
What about product accidents, like bad advice from a chatbot?
Host
He sees little case for intervention there. Existing technology errors and omissions, general liability and cyber policies probably respond to most current harms, a market of specialist AI insurers and vendor indemnities is emerging, and market discipline works better. He even says this weakens the case for some legislative proposals, naming California's SB eight thirteen. The catastrophic layer is different: existing policies exclude war-like acts, nuclear harms and critical infrastructure failure, and per-claim limits are low, so the economy is largely uninsured against AI catastrophes.
Discussant
Part Three is about whether the potential is realized. What determines that?
Host
Four factors. Policyholder type: insurers do far more loss control for a few large commercial clients than for many consumers. Risk type: here the key point is the gap between harm and liability. Insurers can lower losses by reducing harm or by legal manoeuvring, and they will choose the cheaper. He argues that clear, strict and channelled liability shrinks the room for manoeuvring. Under negligence, he says, a monopoly insurer wants policyholders to meet industry practice but has little reason to raise it. In his words, "insurers will help raise the floor, but not necessarily the bar." Long-tail risks, where harm shows up years later, break experience rating. Catastrophic, concentrated, high-limit exposures, by contrast, push insurers to invest in causal risk models up front, since they cannot wait and see. Insurer type: mutuals and niche insurers do best, and he cites the mutual-versus-stock literature, including Hansmann. Market structure: concentration lets an insurer capture the returns from loss prevention, and a mandate removes the fear of being too successful and losing the client.
Discussant
Can insurers even price a risk with no loss history?
Host
Appendix A addresses that with four case studies: cyber, pandemics, nuclear and terrorism. The pattern he reads is that insurers price known unknowns reasonably well, such as cyber accumulation, which stayed profitable through NotPetya and CrowdStrike, and excess mortality for life insurers during COVID. They fail with unknown unknowns: business interruption claims during lockdowns, and terrorism before September eleventh, where insurers were left "failing ruinously with unknown unknowns that fall outside their consideration set entirely." His most striking data point is nuclear: a nineteen eighty-one study inferred from premiums that insurers put serious core damage at about one in four hundred reactor-years, against one in twenty thousand in the government's Rasmussen report, and a later Nuclear Regulatory Commission analysis found the insurers closer. From this he recommends mandated horizon-scanning, modelled on provisions of the Terrorism Risk Insurance Act, to turn unknown unknowns into known ones.
Discussant
Walk through the mandate numbers.
Host
The weak version requires limits around one hundred million dollars, allows exclusions, accepts any insurer, and permits captives. He predicts competing excess and surplus insurers, the specialist market for unusual risks, would underprice to collect data, catastrophic risks would be excluded, and the largest firms would self-insure through captives that change nothing. The strong version requires limits of around two billion dollars or more for named catastrophic risks, no carve-outs, no captives, and a strong financial strength rating. Based on conversations with insurers, he expects ordinary supply to dry up beyond about five hundred million dollars. So policymakers need a carrot, encouraging the labs to form a mutual with lower premiums, and a stick, threatening to force insurers into a joint underwriting association, as happened with American Nuclear Insurers. He writes that ignoring the supply problem would be self-defeating: "Completely halting AI development for lack of insurance availability is a pyrrhic victory in the quest to deliver safe and secure AI."
Discussant
How has the paper been used and cited?
Host
Modestly so far. OpenAlex lists three records for it, the arXiv version, the SSRN version and a duplicate, and none shows a citing work in its index as of early October twenty twenty-six. The Altmetric score recorded on the AI-governance page is one. The main public outlet is the author's own Lawfare essay of December seventeenth, twenty twenty-five, titled Healthy Insurance Markets Will Be Critical for AI Governance, which links the SSRN paper and argues for transparency rules, clearer liability, mutualization and government backstops. The most substantial follow-up we found is again by the author: a July twenty twenty-six report on arXiv, Underwriting the Agent Economy: The Blueprint for an AI Insurance Stack, which Trout leads with co-authors from Stanford, RAND and the insurer Generali, and which cites this paper repeatedly for the nuclear pricing evidence, mutualization and liability channelling. We did not find policy documents, regulator filings or press coverage citing it. That is a search result, and coverage in law reviews and policy memos often lags; it should be read as uncertainty about uptake, not proof of none.
Discussant
Then why might it matter for decisions?
Host
Three audiences. Legislators and regulators considering insurance or financial-responsibility requirements for AI developers get a specific design argument: target catastrophic risk, require high limits, ban captives, set rating standards, and think about the insurer's market structure. Funders working on AI governance face a choice about whether to back market-based mechanisms such as insurance, audits and certification, and this paper is one of the more developed arguments for insurance among them. And labs and insurers deciding whether to build an industry mutual have a design sketch. For an Unjournal evaluation the case is that the argument is mostly qualitative, decision-relevant, written by someone with a commercial stake in the sector, and has had almost no independent scrutiny.
Discussant
Let me take the critiques. First, the conflict of interest. The author works at an AI insurance company that the paper names among the emerging providers.
Host
The paper discloses the affiliation, states that the company did not fund the research, names the Centre for the Governance of AI and the Cambridge Boston Alignment Initiative as supporters, and discloses that Claude was used for research and drafting. The direction of the interest is not obvious. The recommendation is a catastrophic-only mandate served by a mutual or a pool, and the paper is skeptical of intervention in the routine product-liability market where startups like his employer currently sell. An evaluator should still check whether alternatives, such as minimum asset requirements, direct licensing or public reinsurance, get a fair hearing. The paper's stated reason for preferring insurance to other market mechanisms is brief. The author writes: "alternatives seem to be trying to reinvent the wheel; insurance is a mature industry with an ecosystem of tested actors, and whose regulation is more or less a solved problem." That is asserted, not argued, and he says a full comparison with command-and-control regulation is outside scope.
Discussant
Second, internal consistency. He dismisses judgment-proofness because twenty billion dollars is small relative to big tech balance sheets. But the mandate he proposes is around two billion. If tens of billions in cash would not change a chief executive's behaviour, why would premiums on a two-billion-dollar policy?
Host
That is the tension an evaluator should press. His answer has to rest on the non-price channels: coverage conditions, inspections, a centralized research effort, horizon-scanning and peer pressure inside a mutual. Those are the channels the nuclear analogy relies on. But the paper gives no estimate of how large those effects are, and in the nuclear case the self-regulation sat under a regulator with licensing power, which he acknowledges was an important ingredient.
Discussant
Third, capture. A monopoly mutual owned by the frontier labs, shielded from antitrust on safety matters, sounds like a vehicle for collusion or for setting a low standard everyone can meet.
Host
He partly anticipates it. He notes that under negligence a monopoly insurer may resist higher standards, which is one reason he favours strict liability, and he relies on rating requirements and solvency regulation to resist pressure to underprice. But the risk of a mutual whose owners are the risk-creators, setting the standard of care in a young industry, needs more analysis than it gets. So does the barrier to entry: high mandated premiums fall hardest on new entrants, although he suggests scaling coverage with activity.
Discussant
Fourth, does liability even reach the harms he cares about?
Host
That is a precondition he states but largely sets aside. Insurance only transmits liability. Many of the AI catastrophes he lists involve criminal misuse, causal chains across several actors, or systemic effects, where his own footnotes say attribution would be hard and where insurers face third-party moral hazard. Truly existential losses are uninsurable by definition. An evaluator should ask how much of the catastrophic risk distribution sits in the region a two-billion-dollar liability policy can actually price.
Discussant
And the evidence on insurers' track record?
Host
Several checks. Whether the four case studies were chosen to fit the known-versus-unknown-unknowns story. Whether cyber profitability reflects pricing skill or extensive exclusions and sublimits, which the paper itself notes. Whether the counterfactual safety effect of American Nuclear Insurers and the nuclear mutual has been identified, rather than inferred from correlation between insurer ratings and regulator assessments. And whether the industrial organization results on spillovers and fixed costs, mostly from duopoly models of product safety and pollution, transfer to AI safety research. The financial snapshot of the labs uses mid twenty twenty-five figures, already out of date.
Discussant
Last question. What would a useful evaluation deliver?
Host
A check of whether the two-step framework adds discipline beyond the existing regulation-thesis literature, especially Ben-Shahar and Logue and Abraham and Schwarcz; a look at the mandate design under realistic liability rules; whether a small formal model could pin down when a monopoly insurer raises care; and how far the AI-specific claims rest on anecdote.

Sources