Gallagher Re's Q1 2026 report, produced with MIT and Testudo, counts more than 700 cumulative US generative AI lawsuits filed between 2020 and 2025, a 978% increase. Year-over-year growth accelerated to 137% in the most recent period from 59% the year before.

The figure that matters for pricing is not the growth rate. It is the shape: Testudo finds the top 5% of cases account for 99% of total demanded amounts.

Key Takeaways

  • 978% growth on 700+ filings with the annual rate accelerating from 59% to 137%, so the curve is steepening rather than approaching an inflection.
  • $5 million median demand with roughly 70% of cases seeking under $10 million and 16% exceeding $100 million, and the top 5% carrying 99% of the total demanded.
  • 32% of cases are class actions, against 10% to 15% in traditional professional liability, which raises defense cost and settlement severity together.
  • 4.9% of cases involve model hallucinations, the peril standalone AI liability products most often name as their trigger, so the coverage-specific claim count is far smaller than 700.
  • Seven distinct causes of action spread across IP, privacy, employment, products and professional services, which means the exposure does not map to one coverage form.

The Filings Do Not Describe One Peril

The growth number describes a category. The composition describes seven, and they belong to different coverage parts.

Claim Category Share of Filings Actuarial Analog
Patent infringement 11.9% IP litigation / tech E&O
Copyright infringement 11.2% Media / advertising liability
Personal injury & privacy violations 10.2% Cyber liability / CGL Coverage B
Breach of contract / misrepresentation ~15% Professional liability / E&O
Employment discrimination ~8% EPLI
Product liability / negligence ~7% Products liability / CGL Coverage A
Consumer protection / UDAP ~6% Regulatory / D&O

That dispersion is the pricing problem stated in one table. An actuary building a standalone AI liability rate must either price the aggregate across all seven categories or construct a modular rate that prices each separately, and neither approach has precedent or credible data behind it.

Severity is concentrated to a degree that compounds it. Against a $5 million median demand, roughly 70% of cases seek under $10 million while 16% exceed $100 million, and the top 5% of cases carry 99% of the total demanded. A distribution that thin in the body and that heavy in the tail is one where standard lognormal or Pareto assumptions are being fitted to almost no observations in the region that determines the answer.

Class actions run at 32% of filings against 10% to 15% for traditional professional liability, which means loss adjustment expense will develop differently from the analogous casualty lines an actuary would reach for.

The Complement of Credibility Is Doing All the Work

With subject experience this thin, the rate is determined almost entirely by whatever external data is chosen to stand in for it. The candidates disagree with each other.

The subject data is thinner than the headline suggests. Of the 700-plus filings, the large majority remain open, and only 4.9% involve model hallucinations, the trigger standalone products most often write to. The remaining share covers training data IP disputes, employment bias and privacy violations that existing lines already touch. The claim volume that is both unique to AI liability and resolved enough to produce paid loss data is very small.

Settlement opacity removes most of what is left. Thomson Reuters v. Ross Intelligence, the first federal ruling rejecting a fair use defense for AI training data, settled on undisclosed terms after the court found Ross's use of 2,243 of 2,830 Westlaw headnotes commercial and not transformative. Garcia v. Character Technologies, which established that an AI application can be treated as a product for liability purposes, also settled on undisclosed terms in January 2026. The two most consequential severity signals available are both unobservable.

So the weighting of the complements is the rate. Cyber loss experience implies rapid frequency growth followed by severity normalization. Technology E&O development speaks to the roughly 15% of filings alleging breach of contract or misrepresentation. IP litigation severity covers the 23% in patent and copyright, though the 99% concentration in the top 5% of cases suggests a heavier tail than ordinary IP disputes carry. Employment discrimination offers a $365,000 iTutorGroup settlement at one end and a Workday class action covering 1.1 billion rejected applications at the other.

Those complements do not converge, and the choice among them is judgment that no data currently constrains.

Every Writer Is Pricing an Unvalidated Hypothesis

Fewer than five standalone AI liability products exist, and each rests on a proxy for loss experience that has not yet been tested against loss experience.

Munich Re's aiSure, operating since 2018 with limits up to $15 million, prices like a warranty rather than a casualty cover: payouts trigger on breach of predefined performance thresholds with no negligence allegation required. That makes development fast and predictable, and it also means the product does not respond to the tort exposure driving the 978% surge. A model performing within its contractual benchmark can still produce a copyright or discrimination suit.

Armilla with Chaucer writes $25 million or more of AI aggregate limit backed at Lloyd's, underwriting off independent certification with more than 500 evaluations across regulated industries. The actuarial premise is that stronger governance, testing and monitoring produce lower frequency and severity. That premise is reasonable and unverified: no observed correlation between AI governance maturity and claim outcomes exists yet.

Testudo, at $9.25 million per insured with Apollo, Atrium and QBE behind it, prices off its own litigation research, which is the most granular public dataset available and is the same 700-case sample everyone else is short of. Corgi sells modular categories inside technology E&O after a $160 million Series B at a $1.3 billion valuation, which puts the correlation problem in the customer's hands: a firm deploying customer-facing agentic AI is exposed to several of the modules at once, and they are not independent.

Each approach substitutes a different proxy for the missing triangle, and none can yet be falsified. The severity concentration is what makes that fragile rather than merely uncertain. When 99% of demanded amounts sit in 5% of cases, the first large resolved verdict does not adjust the estimate at the margin. It becomes the estimate.

Further Reading

Sources