Artificial intelligence (AI) may be new to the insurance market, but uncertainty is an old customer. Insurers have spent centuries figuring out how to price unfamiliar risks, reward precautions, and avoid betting the whole portfolio on the same weak link. AI gives them new facts to assess, not a reason to discard that experience.
In a recent Truth on the Market post, Ian Adams revisited his decade-old warnings against coverage mandates and taxpayer backstops for cyber insurance. The cyber market’s subsequent growth supports his case for patience. Insurers have expanded coverage while refining prices, exclusions, and security requirements. As Adams argued, AI insurance deserves the same opportunity as courts clarify legal duties, reasonable precautions, and the link between misconduct and injury.
That adaptation has, in fact, already begun. For example, directors and officers (D&O) insurance, which covers certain claims against corporate leaders and companies, already addresses some AI-related risks. It also offers a window into how existing coverage can adapt, even though it covers only part of AI’s potential liabilities.
Together, these markets illustrate three classic bits of wisdom about the value of insurance: It can serve to preserve honest experimentation, reward effective precautions, and account for shared failures. Policymakers should give the market room to apply that wisdom before prescribing coverage or handing taxpayers the bill.
The Market Learned on the Job
Cyber insurance’s growth shows that a market can expand as insurers continue to learn. The National Association of Insurance Commissioners (NAIC) reports that domestic cyber premiums grew from $2.75 billion in 2020 to $7.46 billion in 2025. Domestic insurers reported 5.84 million policies, up 33% from 2024.
Global estimates vary considerably. Jamie MacColl, Jason R.C. Nurse, and James Sullivan’s 2021 Royal United Services Institute report cited roughly $5 billion in premiums. The NAIC cites more than $16 billion for 2025; Shivani Zoting’s Precedence Research analysis estimates $26.32 billion and projects $288.42 billion by 2035.
Cyber insurance expanded without the broad purchase mandates or dedicated taxpayer backstop some observers advocated. That history supports giving markets room to adapt, while leaving open how well they cover particular risks.
But cyber insurance’s growth also shows why the terms of coverage matter as much as availability. MacColl, Nurse, and Sullivan found weak risk data, shared-loss concerns, and limited evidence that insurance consistently improved cybersecurity outcomes. Aggeliki Tsohou, Vasiliki Diamantopoulou, Stefanos Gritzalis, and Costas Lambrinoudakis’ 2023 review documented inconsistent underwriting questions and ambiguous coverage. Huntress later reported that 58% of surveyed policyholders had seen cyber coverage narrow, though its survey cannot establish marketwide prevalence.
These shortcomings fit a market still learning from court decisions and claims experience. Clearer liability rules, better loss data, and technical testing can help insurers refine coverage. Alarm over AI is no reason to rush that work, just as it wasn’t for cyber risks. Sound insurance products take time to develop, and poorly designed or mispriced coverage serves neither insurers nor the policyholders who depend on them to pay claims.
AI raises similar questions for existing policies. As Risk & Insurance explains, AI-related claims can arise under general liability, professional liability, cyber, employment, and D&O policies—even when those policies say nothing about AI. That silence can leave insurers and policyholders disagreeing over which losses they agreed to cover.
Some insurers are making those boundaries explicit. Insurance Business reports that the Insurance Services Office, which develops standard policy language for insurers, introduced optional exclusions in January 2026 for losses involving generative AI. Insurers can add these exclusions to commercial general liability policies, which typically cover claims involving bodily injury, property damage, and certain other harms. Each insurer decides whether to adopt them, while specialty AI coverage is also emerging. Businesses therefore need to check both what their existing policies exclude and what protection they can buy elsewhere.
AI Meets the Fine Print
As Priya Cherian Huskins explains for Gallagher, D&O policies protect executives, reimburse companies that cover executives’ eligible losses, and cover certain claims against companies themselves. Public-company coverage generally limits that last category to securities claims. Defense costs, settlements, and judgments can all qualify for coverage.
Kevin LaCroix of The D&O Diary has counted 30 AI-related securities class actions filed thus far 2026, compared with 16 in 2025. Those totals track filings, not active cases or insured losses, and definitions of “AI-related” vary. Economic consulting firm Cornerstone Research and Stanford’s Securities Class Action Clearinghouse counted 15 in the first half of the year, while National Economic Research Associates economists Ivelina Velikova and Svetlana Starykh counted 18.
The technology may be new, but many of the allegations are quite familiar for D&O insurers. Writing in Bloomberg Law, Ross Shikowitz, Adam McCall, and William Massa identify premature commercialization promises, mischaracterized software, and abrupt corporate reinventions among the common causes of action. Disputes involving Apple’s Siri, Upstart, Tempus AI, and Endava concern performance claims, revenue expectations, affiliated transactions, and accounting—all well-trod ground for such claims.
In D’Agostino v. Innodata, an investor accused the data-engineering company and its senior executives of touting proprietary AI while relying heavily on employees doing labor-intensive data entry and labeling. Judge Jamel Semper of the U.S. District Court for the District of New Jersey dismissed the complaint without prejudice in April 2026, finding, among other deficiencies, that the plaintiff had failed to allege facts strongly supporting an inference of fraudulent intent.
The Opendoor litigation produced a $39 million settlement over alleged misrepresentations about its home-pricing algorithm. The company reported that the agreement fell within insurance limits and recorded expected insurance recoveries. That illustrates insurance’s existing role even when fraud is not formally established.
Copyright and privacy disputes can also generate claims against corporate leaders. In the Adobe derivative complaint, a pension-plan shareholder sued on Adobe’s behalf, alleging that its leaders allowed the company to train its SlimLM language models on datasets containing unauthorized copies of copyrighted books and other works. The suit also challenges assurances about lawful sourcing and the commercial safety of Adobe’s AI products.
The shareholder alleges that officers and directors breached their duties through those practices and misleading disclosures, exposing Adobe to litigation and other harm. A derivative suit seeks recovery for the company. Such allegations can generate oversight claims without automatically bringing the underlying copyright liability within D&O coverage.
Policy language, which can vary, also remains decisive. Chubb requires final, nonappealable findings for specified fraud exclusions, so allegations alone need not end defense funding. Other exclusions, notice requirements, deductibles, and limits remain; insurability of fines varies by jurisdiction.
Disappointment Is Not a Cause of Action
The first bit of insurance wisdom for AI markets, well known to D&O specialists, is to distinguish misleading disclosures from business disappointments. In Dura Pharmaceuticals v. Broudo, investors alleged misleading statements about prospects for Food and Drug Administration (FDA) approval of an asthma-treatment device. The Supreme Court unanimously held in 2005 that an inflated purchase price alone did not establish “loss causation”—the connection between alleged fraud and investors’ economic loss.
The same distinction matters for AI investments. Investors may lose money because managers conceal problems, but they may also lose money because competitors improve, financing costs rise, or an experiment simply fails. A securities-fraud claim—the kind D&O insurance generally covers—must connect the alleged misconduct to the investor’s loss.
To point to a recent example: CoreWeave’s weather-delayed data-center construction offers a mundane reminder that even AI needs concrete. Managers must disclose material information as securities law requires, including about construction delays. But a setback and a falling stock price do not, by themselves, establish fraud.
Knowing that a project could fail should not make company leaders liable each and every time that risk materializes. Investors can lose money even when managers disclose risks honestly. Treating those losses as grounds for liability would turn disclosure law into investment insurance, discouraging companies from pursuing promising but uncertain projects.
Cornerstone reports that the 15 AI-related securities class actions filed in the first half of 2026 involved $385 billion in declines in companies’ stock-market value around alleged disclosures. Those cases represented 13% of filings excluding merger-and-acquisition class actions, but 73% of the stock-market losses captured by its disclosure-loss measure. Those declines are not necessarily damages attributable to fraud. Treating them as recoverable losses without establishing that connection could discourage companies from pursuing promising experiments whose risks investors knowingly accept.
Clearer duties and causation rules help insurers assess which failures could produce covered claims, while coverage can give managers confidence to pursue uncertain projects. In a recent working paper, Wenshuai Xu and Jianfei Peng report that firms carrying D&O insurance invest more in AI. Their analysis suggests that greater managerial willingness to bear risk, while fewer financing constraints help explain the association. The association is also stronger among firms with less-overconfident executives, suggesting that coverage can support experimentation by relatively cautious managers.
That finding supports giving the market room to develop, and is consistent with the combination of investment and restraint that advocates of faster insurance expansion say they want. Mandates or underpriced guarantees could undermine that balance by weakening the pricing and coverage decisions through which insurers distinguish prudent experimentation from poorly managed risk.
To be sure, more investment does not necessarily mean better investment, and these findings alone cannot establish what the presence of insurance coverage causes. They nevertheless highlight a benefit worth preserving. Insurance can give managers room to pursue uncertain projects, while liability rules hold them accountable for actionable misconduct. The aim is to make worthwhile risks bearable while keeping dishonesty expensive.
Good Governance Might Earn a Discount
The second bit of insurance wisdom concerns rewarding effective precautions. Insurance broker Aon’s Financial Services Group already advises applicants to expect questions about how they use AI, how heavily their business depends on it, and how it supports their financial projections. Underwriters also want to know who oversees AI risks and checks related disclosures before publication.
A survey by The D&O Diary and Allianz Commercial suggests that insurers take those questions seriously. Among insurer respondents, 91.2% agreed that inadequate transparency about AI deployment should count against applicants in underwriting assessments. Only 59.1% of brokers agreed—a sizable gap between those providing coverage and those arranging it. The responses reveal how participants think transparency should affect coverage decisions, but don’t establish whether insurers changed premiums or whether greater transparency reduced claims.
Turning those expectations into useful assessments requires evidence. Underwriters need to check whether customer contracts support AI-related revenue claims, whether performance tests back up claims about what the technology can do, and whether firms have the legal right to use their training data. They should also ask how employees report failures, who receives those reports, and who can require corrective action. A board-approved policy offers little protection if employees cannot get management to address a problem.
Technical testing has limits, too. RAND researchers Sasha Romanosky and Celine Robinson explain that performance tests work best for defined tasks with clear answers. Test results alone cannot establish how often errors will produce insured losses, especially as models and their uses change. Insurers need to understand both how a system performs and what happens when it fails.
Research on conventional D&O insurance shows both the promise and the limits of evaluating how companies manage risk. In a 2007 University of Chicago Law Review study, Tom Baker and Sean J. Griffith found that underwriters considered corporate governance when assessing applicants and pricing coverage. Their judgments emphasized management’s character and corporate culture alongside formal board structures. Charging better-managed firms less could reward sound governance. The study did not, alas, establish that those price differences actually prompted firms to improve.
Baker and Griffith’s companion Georgetown Law Journal study examined what happened after insurers issued coverage. They found that D&O insurers neither monitored governance during the policy term nor actively managed litigation defense costs after claims arose. The authors attributed the lack of monitoring partly to conflicting incentives—what the law & economics literature would call a “principal-agent problem.” Managers buy coverage that protects them personally, while the corporation—and ultimately its shareholders—pays the premiums. Managers may therefore prefer protection with few strings attached.
AI-related underwriting faces the same incentive problem. As insurers learn more about AI-related exposure, they could offer better terms to firms with tested safeguards and reliable disclosures, while charging more, requiring larger deductibles, or narrowing coverage where precautions fall short. But rewarding firms that already manage risks well won’t necessarily persuade other managers to improve.
Insurance markets have a long record of adjusting prices and terms as claims experience improves their understanding of risk. The same process can shape AI coverage. Competition and better information can lower premiums for well-understood risks, while firms with weaker safeguards face higher prices or tighter terms.
Claims experience should further clarify frequency (how often claims arise) and severity (how much they cost). And even unsuccessful suits can generate substantial defense bills. Larger settlements will reach excess insurers, which cover losses above the primary policy’s limit.
That distinction also helps clarify which AI losses need attention now. In a company’s layers of coverage, one insurer might cover the first $10 million above the amount the company pays itself, while another covers the next $10 million. The first faces smaller claims and routine defense bills. The second faces unusually costly litigation. Catastrophic scenarios attract the headlines, but ordinary claims and legal expenses are the more immediate test as AI coverage develops. Insurers need to price both, and those covering different portions of the losses can reasonably assess the same company differently.
A Thousand Policies Can Share One Problem
The third bit of insurance wisdom comes from managing losses that arrive together. Insurers have long understood that a thousand policies on homes along the same coastline can amount to one large bet on the weather. Concentration means too much exposure to a particular source of loss, while correlation means losses tend to occur together. Either can undermine the protection that a large, varied pool of customers normally provides.
Reinsurance—insurance that insurers buy—is central to managing that exposure. In a reinsurance contract, an insurer transfers part of its potential losses to another company, which can spread them across a broader portfolio of regions and risks. As the Federal Reserve Bank of Chicago explains, reinsurance helps insurers manage concentrations in particular locations or lines of business. It also gives them capacity to offer more coverage than they could prudently support alone.
AI presents a new version of this familiar problem. Companies in different industries and countries may depend on the same model or cloud provider. An insurer’s customer list may appear to be diverse while concealing a common vulnerability. Reinsurance can spread the financial burden, but reinsurers must identify those dependencies across all the insurers they cover. Transferring a risk does not make the underlying connection disappear.
In AI markets, those connections can also extend beyond the underlying technology. Law firm Quinn Emanuel’s analysis of data-center financing describes developers relying on customer commitments to repay construction debt and, in some arrangements, using specialized computing chips as collateral. Construction delays, power shortages, customer defaults, or falling chip values can threaten repayment. Cross-default clauses, which allow a default under one agreement to trigger defaults under others, can spread the distress.
One disruption could therefore produce claims against developers, customers, lenders, and corporate leaders under different policies. Ordinary defaults and business losses would, of course, each still require a separate basis for coverage. Insurers and reinsurers nevertheless need to assess their combined exposure, including claims that look manageable individually but become costly together.
Shared exposure can also emerge in court. RAND researchers Romanosky and Robinson identify legal changes as a potential source of correlated claims when many firms follow similar practices. Their analysis also makes an important distinction about exclusions. Removing a risk from an insurance contract leaves the business to bear it. Better information can improve risk allocation, but it cannot guarantee unlimited insurance capacity.
These sorts of potential limits on private capacity offer the strongest arguments for considering a public backstop. Gregory C. Allen of the Center for Strategic and International Studies argues that restrictions imposed by the insurance market might actually impede AI adoption. He recommends preparing a staged catastrophic backstop alongside better incident data and verification.
To be sure, that concern deserves serious consideration. But any proposal should explain which losses private insurance and reinsurance cannot absorb, why that gap warrants public intervention, and how the arrangement would preserve incentives to reduce losses. A government guarantee priced below the risk it assumes could weaken those incentives before taxpayers ever pay a claim. Coverage mandates could likewise lock in obligations before insurers understand their combined cost.
Public policy can also play a useful role by helping markets to learn without having taxpayers assume those obligations. RAND recommends standardized notices explaining how policies address AI, common categories for reporting incidents and claims, and shared scenarios for testing potential losses. Such information could help insurers and reinsurers identify common vulnerabilities and negotiate coverage with fewer surprises. Standardizing how firms describe coverage need not dictate what they sell.
The Fine Print Still Applies
AI presents some novel and unfamiliar risks, but familiar insurance principles still apply. As it has in other markets and with respect to other kinds of risk, AI insurance can serve to preserve honest experimentation, reward demonstrated precautions, and watch for shared failures. Cyber insurance shows how markets can learn, while D&O coverage shows that adaptation to AI has already begun. Courts can clarify responsibility, insurers can refine coverage, reinsurers can spread risks, and competition can reward firms that manage them well.
Policymakers should strengthen the conditions apt to produce those lessons before prescribing coverage or shifting losses to taxpayers. New technology alone supplies no reason to discard accumulated wisdom about risk, regulation, and policy. Indeed, the reality is quite the opposite: Even AI can learn something from the fine print.
