AI hallucination insurance: what pays when AI gets it wrong

Is there insurance for AI hallucinations?

Yes, though not under that name. Hallucinations, meaning confident but false output, are named in AI liability policies, in AI endorsements to tech E&O and professional liability, and in performance covers that pay when a model misses agreed accuracy.

12 of the 28 listings in our register name hallucinations or model error. Ollive's planned cover for AI vendors is announced but not yet on sale; the others are confirmed on the issuer's own pages.

Related risks often sit in the same wording: model drift, underperformance and harmful generated content^{1, 2}. A policy that names one usually names the others.

The losses look ordinary once they happen. A support chatbot invents a refund policy, a research tool cites a source that does not exist, or a pricing tool quotes the wrong figure to a customer. The claim that follows is for money lost by relying on the answer.

What does hallucination cover pay for?

It depends on who loses money. Cover for false AI output falls into three types:

• Third-party claims: someone relied on false output and sues. Armilla's policy names model errors including hallucinations, inaccuracy and drift¹. Testudo's policy for US companies deploying generative AI includes a Generative AI Errors insuring agreement, which it describes as hallucination liability³.
• Your own financial loss: the model underperforms and you lose money with no lawsuit. Mosaic Insurance and Munich Re cover losses when a model misses defined performance thresholds, with claims settled on measurable performance data⁴.
• A promise to your customers: Armilla Guaranteed lets an AI vendor stand behind specified accuracy or quality metrics, with a pay-out funded if they are not met. Armilla calls it a contractual guarantee, not an insurance policy¹.

Third-party cover typically pays defense costs and damages. Fixing the model is a separate cost: Embroker's endorsement covers the cost to fix or remove a flawed AI system at a client's request, capped at $150,000⁵.

Which insurers cover AI hallucinations?

These are the listings in our register that name hallucinations or model error. Limits and how to buy are shown where the issuer publishes them.

• AArmillaArmilla Insured$25Mup to, per organizationSurplus lines brokers, or Armilla underwriting after an AI risk assessment
• CChaucer and ArmillaVanguard AI$25M+AI aggregate, alongside $10M cyberSurplus lines brokers
• TTestudoGenerative AI Liability Insurance$10Mlimits from $1M to $10MThrough brokers, on a US excess and surplus lines basis
• MMosaic Insurance and Munich ReAI performance cover15Minitial capacity, EUR, USD or CADThrough brokers
• CCFCAffirmative AI wording across seven productsLimit not publishedBuilt into CFC products
• MMunich ReaiSureLimit not publishedMunich Re directly, or partner programs
• CCounterpartAffirmative AI CoverageLimit not publishedThrough brokers
• HHiscoxTechnology Professional Indemnity with affirmative AI coverLimit not publishedHow to buy not stated
• CCorgiModular AI coverage on tech E&OLimit not publishedOnline, direct
• EEmbrokerArtificial Intelligence Coverage Endorsement$150Kfor algorithm removalOnline, direct
• AAIUCAIUC-1-backed AI agent insuranceLimit not publishedThrough AIUC-1 certification
• OOlliveAI liability for AI vendors, with runtime monitoring (announced)Limit not publishedBook a demo with Ollive

CFC says its affirmative AI wording across seven products addresses hallucination, AI-generated content and model drift². Munich Re's aiSure lists hallucinations among the AI liabilities it covers, and describes the suite as suited to frequent, smaller AI errors that can add up to large aggregate or class-action exposure⁶.

Counterpart added explicit cover for mistakes made by AI tools the insured uses or relies on, aimed at small businesses and professionals such as consultants and financial advisors⁷. Hiscox's UK technology PI wording expressly covers client claims arising from the insured's use of AI, including generative AI⁸.

Corgi lists inaccurate or harmful generated content among its AI modules⁹. AIUC-backed cover for AI agents is described around harm such as fabricated or incorrect information, for example misquoted advice^{10, 11}. Ollive's planned cover lists AI output liability, including harmful outputs and unsafe advice¹².

Why might your current policy not cover a hallucination claim?

Because it may say nothing about AI, or it may exclude it. Both are worth checking before a claim, not after.

Cyber policies are a common gap. They are built around a security event, and a hallucination usually involves none. Vanguard AI was designed with this in mind: its Armilla layer is for losses from model behavior, such as wrong outputs, where no cyber event occurred, while Chaucer's cyber and tech E&O cover breach-driven losses and tech service failures^{16, 17}.

Trade press says ISO published six standard forms in July 2025, with first carrier filings that August, but the sources we read name only three¹⁴. Professional lines are affected too: a May 2025 law-firm alert described a W. R. Berkley 'absolute' AI exclusion in D&O, E&O and fiduciary policies¹⁸.

The exclusions tracker lists the forms and filings we follow, and the coverage checker shows where to look in your own policy.

How much hallucination cover can you buy?

Few of these listings publish a limit, and none publishes a price. These state a figure:

• Armilla: up to $25M per organization¹.
• Vanguard AI, from Chaucer and Armilla: AI aggregate limits of $25M or more, alongside $10M of cyber¹⁶.
• Testudo: limits from $1M to $10M, per its launch post. Atrium's February 2026 release cited up to $9.25M per insured^{19, 20}.
• Mosaic and Munich Re: up to 15 million of initial capacity, in euros, US dollars or Canadian dollars⁴.
• Embroker: $150,000 for algorithm removal, with other AI claims at full policy limits, Embroker says⁵.

CFC, Hiscox, Counterpart, Corgi, Munich Re's aiSure and AIUC state no AI limit on the pages we read, and Ollive's program is still being designed²¹. A published limit is the most a policy will pay, not what a hallucination claim would pay.

What do insurers want to see before covering hallucination risk?

Evidence about the AI system itself. Armilla's policies come with independent AI system certification¹, Mayflower Specialty says it scores risks including hallucination and drift²², and AIUC links insurance terms to an AIUC-1 audit¹⁰.

In practice that means showing how you test outputs before release, how you watch for drift in production, where a person reviews high-stakes answers and what your customer contracts promise about accuracy. A contract that promises accuracy can create liability a policy excludes, so check the contractual liability clause.

AI Insurance Tracker does not sell insurance. A licensed broker can confirm whether a specific policy would respond to a hallucination claim. For agents, see insurance for AI agents; for the wider market, see AI liability insurance.