From analyzing over 200 AI-related insurance patent filings and cross-referencing them with the emerging litigation database, the gap between carriers' AI ambitions and the market's ability to price AI risk is the widest we have observed in any emerging coverage class. Generative AI-related lawsuits in the United States grew 978% between 2020 and 2025, with cumulative filings surpassing 700 and year-over-year acceleration reaching 137% in the most recent period. That acceleration compares with 59% the year before, signaling a litigation curve that is steepening rather than flattening. For actuaries tasked with pricing AI liability, whether standalone, embedded in D&O/E&O, or excluded entirely, this litigation wave is the closest available proxy for a loss development triangle. Yet it remains too immature to anchor credible ratemaking.

This article synthesizes the litigation data from Gallagher Re's Q1 2026 Global InsurTech Report and Testudo's January 2026 State of Play into an actuarial pricing framework. We examine why standard loss development methods break down for AI risk, how the handful of standalone AI liability writers are constructing rates without historical loss data, and what court precedents are beginning to shape the loss distributions that actuaries will eventually need to triangulate.

The Litigation Curve: 978% Growth and Accelerating

Gallagher Re's Q1 2026 analysis, produced in collaboration with MIT and Testudo Global Inc., documented 700+ cumulative GenAI-related lawsuit filings in the United States from 2020 through 2025. The headline growth rate of 978% captures the full period, but the year-over-year trajectory matters more for actuarial trend selection. Filing growth accelerated from 59% between 2023 and 2024 to 137% between 2024 and 2025. That acceleration eliminates any hope that the litigation curve is approaching an inflection point. The trend is still in its exponential phase.

Testudo's more granular analysis breaks down the filing characteristics that actuaries need for severity modeling. The median demand amount across all GenAI cases is $5 million. Roughly 70% of cases seek less than $10 million, but 16% involve demands exceeding $100 million. The concentration is extreme: the top 5% of cases account for 99% of total demanded amounts. That skewness creates a severity distribution with a long right tail that standard lognormal or Pareto assumptions may underestimate, particularly given the absence of closed-claim data to calibrate the tail.

Nearly one-third (32%) of all cases are filed as class actions, a proportion that inflates both defense costs and potential settlement severity. For comparison, class action rates in traditional professional liability rarely exceed 10-15% of total filings. The class action share alone suggests that AI liability will develop with higher loss adjustment expense ratios than most analogous casualty lines.

Claim Type Distribution

Gallagher Re's breakdown by cause of action provides the first actuarial-grade segmentation of AI litigation by peril type:

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

The dispersion across seven distinct cause-of-action categories is the core problem for actuarial pricing. AI liability does not map to a single coverage form. It touches intellectual property, privacy, employment practices, products liability, and professional services simultaneously. An actuary building a standalone AI liability rate must either price the aggregate exposure across all categories or build a modular rate structure that prices each peril separately. Neither approach has precedent, and neither has credible loss data.

Why Traditional Loss Triangles Fail for AI Risk

Loss development triangles require three conditions that AI liability does not yet satisfy: a sufficient volume of closed claims to establish development patterns, consistent coverage forms that define what constitutes a covered loss, and a stable legal environment that allows historical loss experience to predict future outcomes. AI liability fails on all three.

Too Few Closed Claims

Of the 700+ cumulative filings, the vast majority remain open. Testudo's analysis found that only 4.9% of cases involve AI model hallucinations specifically, the peril category that standalone AI liability products most frequently cite as their coverage trigger. The remaining 95% of cases involve upstream exposures (training data IP disputes, employment bias, privacy violations) that existing coverage lines already address, at least partially. The net result: the claim volume that is both unique to AI liability and resolved enough to generate paid loss data is vanishingly small.

Patterns we have tracked across patent and copyright AI cases show settlement opacity compounding the data problem. Thomson Reuters v. Ross Intelligence, the first federal ruling rejecting a fair use defense for AI training data, settled on terms that were not publicly disclosed. The Garcia v. Character Technologies wrongful death case involving a minor's suicide after interaction with an AI chatbot settled in January 2026 with undisclosed terms but safety feature commitments. When the most consequential cases settle confidentially, the actuarial community cannot observe the severity distribution that will ultimately drive loss costs.

No Standardized Coverage Forms

Traditional casualty lines benefit from ISO or AAIS standardized forms that create homogeneous exposure units. AI liability has no equivalent. The ISO CG 40 47 and CG 40 48 endorsements that took effect January 1, 2026 define what AI liability is not (it is excluded from standard CGL), but no corresponding standardized affirmative form exists. Each standalone AI liability writer uses a proprietary policy form with different definitions of "generative artificial intelligence," different coverage triggers, and different exclusionary language.

ISO's definition of generative AI in CG 40 47 reads: "a machine-based learning system or model that is trained on data with the ability to create content or responses, including but not limited to text, images, audio, video or code." That definition is broad enough to capture most large language model applications, but standalone AI liability products define coverage triggers around specific failure modes (hallucinations, bias, performance degradation) that do not map cleanly to ISO's exclusionary language. The definitional mismatch means that what one carrier excludes from CGL may not align with what another carrier affirmatively covers in a standalone AI liability form.

Rapidly Evolving Model Capabilities

The third structural problem is that the underlying technology changes faster than any actuarial projection period can accommodate. A rate filed today based on the litigation profile of GPT-4-era AI systems may be obsolete by the time the policy period begins, because the model capabilities, deployment patterns, and regulatory environment will have shifted. Gallagher Re's survey of 1,250 companies found that 57% cited AI errors and hallucinations as their primary AI risk, 56% cited legal and reputational risks, and 55% cited data protection and privacy violations. Those risk perceptions will shift as model architectures evolve, creating a moving target for loss trend assumptions.

This technology-driven instability is unlike anything in the actuarial pricing canon. Even cyber insurance, the most commonly cited analogy, benefited from relatively stable underlying technology during its formative pricing years. Network intrusions and data breaches operated on similar technical mechanisms year over year. AI model capabilities, by contrast, have shown discontinuous jumps: the shift from GPT-3 to GPT-4, the emergence of agentic AI systems, and the proliferation of multimodal models each introduced qualitatively new risk profiles rather than marginal extensions of existing ones.

How Standalone AI Liability Writers Price Without Data

Fewer than five standalone AI liability products exist globally. Each has adopted a distinct pricing methodology that reflects the absence of actuarial-grade loss data. From tracking these products since their launch, four pricing archetypes have emerged.

Munich Re aiSure: Performance Guarantee Pricing

Munich Re's aiSure, operating since 2018, pioneered AI performance guarantee insurance with coverage up to $15 million through its Mosaic Insurance partnership. The product uses a parametric-like structure: payouts are triggered when predefined AI performance thresholds are breached, with no negligence allegation required. The pricing methodology resembles surety or warranty pricing more than traditional casualty ratemaking. Munich Re evaluates the AI model's architecture, training data provenance, and validation test results, then prices against the probability of performance degradation below contractual benchmarks.

The actuarial advantage of this approach is that it sidesteps the litigation uncertainty entirely. Claims are triggered by measurable performance metrics rather than third-party lawsuits, which means the loss development pattern is faster and more predictable. The disadvantage is that it does not cover the tort liability exposure that constitutes the bulk of the 978% litigation surge. A company whose AI system performs within contractual benchmarks can still face a copyright infringement or employment discrimination lawsuit.

Armilla and Chaucer: Audit-Based Underwriting

Armilla AI, a Toronto-based MGA and Lloyd's coverholder, launched standalone third-party AI liability coverage in April 2025 and expanded to the Vanguard AI coordinated structure with Chaucer in February 2026. The structure provides $25 million or more in AI aggregate limits plus $10 million in cyber limits per organization, backed by Axis Capital at Lloyd's. Armilla's pricing methodology centers on independent AI system certification: the company runs 500+ evaluations across regulated industries before underwriting a risk.

This evaluation-based approach creates a proxy for loss control that substitutes for missing loss history. The actuarial logic is that organizations with demonstrably stronger AI governance, testing protocols, and model monitoring should experience lower claim frequency and severity. The weakness is that no empirical validation of this hypothesis exists yet. The correlation between AI governance maturity and actual claim outcomes is assumed, not observed. Armilla's CEO Karthik Ramakrishnan has described AI liability as "shifting from implicit exposure within cyber and technology policies to a risk requiring dedicated coverage," but the pricing remains qualitative rather than actuarially triangulated.

Testudo: Lloyd's Lab Syndicate Pricing

Testudo, a Lloyd's Lab-backed startup founded by former Goldman Sachs technologists, expanded its AI liability capacity to $9.25 million per insured in March 2026. The underwriting panel includes Apollo Underwriting, Atrium, and QBE, with Gallagher Re as the broker partner. Testudo's claims-made product responds to third-party claims from AI-generated outputs, targeting mid-market US enterprises in technology, financial services, healthcare, and media.

Testudo's pricing advantage comes from its own litigation research. The company's State of Play report provides the most granular publicly available data on GenAI litigation characteristics, including the $5 million median demand, the 32% class action rate, and the extreme severity concentration. That data feeds directly into the company's exposure rating model, creating a feedback loop between litigation intelligence and underwriting. Co-founder Mark Titmarsh has characterized the approach as "thoughtful, well-structured underwriting" rather than traditional actuarial loss costing.

Corgi: Modular Tech E&O Integration

Corgi's AI liability coverage, launched alongside its $160 million Series B at a $1.3 billion valuation in May 2026, operates as a modular add-on to technology E&O. Customers select which AI risk categories to include (hallucination, bias, training data IP, adversarial attacks, synthetic media, autonomous system failures) through a self-service dashboard. The pricing challenge is correlation: a company deploying customer-facing agentic AI is simultaneously exposed to multiple peril categories, and those exposures are not independent.

Corgi's full-stack carrier model gives it a structural advantage in building proprietary loss data. As a licensed carrier writing on its own paper, Corgi controls claims adjudication and reserve methodology directly. The company reached $40 million in annual recurring revenue within its first year of operations, providing early premium volume that will eventually generate the loss experience needed for actuarial-grade pricing. Whether that experience validates or undermines the initial rate assumptions is the open question.

Carrier Product Type Max Limit Pricing Basis
Munich Re aiSure Performance guarantee $15M Model audit / warranty
Armilla / Chaucer Standalone liability $25M+ 500+ point evaluation
Testudo (Lloyd's) Claims-made liability $9.25M Litigation intelligence
Corgi Modular E&O add-on Varies Module-specific + correlation

The Actuarial Toolkit for Novel Risk: What ASOPs Allow

Actuaries pricing AI liability are not operating without professional guidance, even though they are operating without loss data. Three Actuarial Standards of Practice provide the framework for pricing novel risk classes, and each addresses a different dimension of the AI liability challenge.

ASOP No. 12: Risk Classification

ASOP No. 12 governs the design, development, and evaluation of risk classification systems. Its recent revision added guidance on "potential for unintended bias" in classification variables, language that applies directly to the challenge of segmenting AI liability risk. The standard requires that risk classification systems produce groups that are actuarially sound, meaning the expected outcomes within each group should be reasonably homogeneous and distinguishable from other groups.

For AI liability, the classification challenge is defining what variables predict loss frequency and severity. Testudo's data suggests that deployment context matters enormously: consumer-facing AI applications generate different litigation profiles than internal-use tools. The 32% class action rate implies that B2C deployments carry disproportionate severity risk. But whether "consumer-facing versus internal-use" constitutes an actuarially sound classification variable requires observed loss data that does not yet exist. The actuary must document the rationale for the chosen classification system while acknowledging that the variables have not been empirically validated against loss outcomes.

ASOP No. 25: Credibility Procedures

ASOP No. 25 is the most directly relevant standard for AI liability pricing. It provides guidance on selecting credibility procedures and applying them to datasets, including situations where the subject experience has low credibility. When historical loss data is sparse, the standard requires the actuary to identify an appropriate complement of credibility: an external data source or judgment-based estimate that fills the gap between what the limited data says and what a fully credible estimate would require.

For AI liability, candidate complements of credibility include:

  • Cyber insurance loss experience: The most commonly cited analogy, given that both lines cover technology-mediated third-party liability. Cyber's early development patterns show rapid frequency growth followed by severity normalization, a pattern that may or may not apply to AI.
  • Professional liability / E&O development: Relevant for the breach-of-contract and misrepresentation claims that constitute roughly 15% of AI filings. E&O development factors for technology companies provide a partial benchmark.
  • IP litigation severity: For the 23% of AI filings involving patent or copyright infringement, IP litigation severity benchmarks from the technology sector offer an analogous severity distribution, though AI-specific factors (the $5 million median demand, the 99% severity concentration in the top 5% of cases) suggest the tail may be heavier than traditional IP disputes.
  • EEOC and employment discrimination settlements: The iTutorGroup settlement of $365,000 for age discrimination in AI-powered recruitment provides an early data point, but the Workday class action involving 1.1 billion rejected applications represents a severity scenario orders of magnitude larger.

The actuary's judgment in weighting these complements is the single most consequential pricing decision for AI liability. ASOP No. 25 requires that the selected weighting be documented and that the rationale be available for review. Given the absence of AI-specific loss data to validate the weights, the selection is inherently judgmental and will remain so until several years of closed-claim experience accumulate.

ASOP No. 38: Using Models Outside the Actuary's Area of Expertise

ASOP No. 38 governs situations where actuaries use models developed outside their traditional expertise. For AI liability pricing, this applies when actuaries adopt scenario-based or Monte Carlo simulation models from AI/ML research to project loss distributions. The standard requires the actuary to evaluate whether the model is appropriate for the intended purpose, to understand the model's key assumptions, and to consider whether the model has been validated.

Several AI liability writers use scenario analysis and stress testing as their primary pricing tools. These methods construct hypothetical loss events (a major copyright ruling, a class action bias verdict, a product liability judgment involving autonomous systems) and estimate probability-weighted losses across the scenarios. The actuarial discipline is in selecting scenarios that span the plausible loss space without overfitting to a small number of dramatic but low-probability events.

Court Precedents Shaping the Loss Distribution

The litigation data is not entirely void of signal. Several court decisions are beginning to define the boundaries of AI liability in ways that will eventually constrain the actuarial loss distribution.

Training Data IP: Thomson Reuters v. Ross Intelligence

In February 2025, Judge Bibas of the Third Circuit (sitting by designation in the District of Delaware) granted partial summary judgment to Thomson Reuters in the first federal ruling rejecting fair use for AI training data. Ross Intelligence had used 2,243 of 2,830 Westlaw headnotes to train its competing legal AI. The court found the use was commercial and not transformative, and that the market substitute factor weighed decisively for the plaintiff. Ross has appealed, and the Third Circuit is considering two certified questions.

For actuarial severity modeling, Thomson Reuters v. Ross establishes that AI training data use can constitute copyright infringement when the output competes with the original. If the Third Circuit affirms, the ruling will apply to every AI system trained on copyrighted material, creating a severity scenario that encompasses the 11.2% of GenAI filings involving copyright claims. The NYT v. OpenAI case in the Southern District of New York, still pending, could either reinforce or narrow this precedent depending on how the court treats the fair use defense for models with broader output capabilities.

Product Liability: Garcia v. Character Technologies

The Garcia v. Character Technologies case, filed in the Middle District of Florida in October 2024, tested whether an AI chatbot application qualifies as a "product" under product liability law. A fourteen-year-old died by suicide after alleged emotional manipulation by Character.AI's chatbot. The federal judge greenlit product liability claims, ruling that the AI application could be considered a product. Google and Character.AI settled in January 2026 with undisclosed terms and safety feature commitments.

This ruling has direct implications for the 7% of GenAI filings involving product liability and negligence claims. If AI applications are products, manufacturers face strict liability for design defects and failure to warn, a liability standard that does not require proving negligence. Verisk's emerging issues analysis tracked 11 of 64 GenAI-specific cases from the George Washington University DAIL database that involve design defects, failure to warn, negligence, or product liability. The strict liability pathway, if it becomes established precedent, would materially increase expected severity for AI-generated outputs that cause harm.

Employment Discrimination: The Algorithmic Bias Frontier

AI-driven employment discrimination claims represent approximately 8% of GenAI filings. The EEOC's settlement with iTutorGroup for $365,000 in age discrimination related to AI recruitment screening established a floor for regulatory enforcement actions. The Workday class action, which involves algorithmic screening of 1.1 billion applications, represents the catastrophic severity scenario. If certified as a class, the potential damages would dwarf any AI liability insurance limit currently available.

The Colorado AI Act's July 2026 compliance deadline will generate the first wave of regulatory enforcement actions against AI-driven insurance decisions specifically. Insurers using AI for underwriting, pricing, or claims decisions in Colorado will face mandatory bias audits and disclosure requirements. Noncompliance creates both direct regulatory liability and a litigation roadmap for private plaintiffs seeking to use the statutory framework as evidence of negligence.

The Silent AI Coverage Gap

While actuaries struggle to price AI liability affirmatively, carriers are actively excluding it from standard books. Verisk's ISO exclusion endorsements (CG 40 47 for both Coverage A and B, CG 40 48 for Coverage B only, CG 35 08 for Products/Completed Operations) have been approved in over 80% of state filings. W.R. Berkley's PC 51380 endorsement applies an absolute exclusion with no carve-back for incidental AI use across D&O, E&O, and Fiduciary Liability lines.

Gallagher Re's Freddie Scarratt, global deputy head of insurtech, warned that "silent AI" could pose a bigger threat than "silent cyber" did. The parallel is instructive: silent cyber exposure accumulated on property and casualty books for years before carriers recognized the aggregation risk, leading to a disorderly market correction. AI risk is following the same trajectory but faster. Gallagher Re's survey found that one in five respondents reported a client experiencing an AI-related loss or claim in the past year, and just over half were fully covered by insurance.

The coverage vacuum is widening. ISO forms underpin approximately 82% of US P&C policies. As AI exclusions proliferate across those forms, the insurable market for AI liability concentrates in the handful of standalone writers described above, none of which has the capacity or the loss history to absorb a systemic AI liability event. Deloitte projects the standalone AI liability market will reach $4.7 to $4.8 billion in annual global premiums by 2032, growing at an 80% compound annual rate. Whether the actuarial infrastructure to support that growth exists by then is an open question.

The Lockton Re and Armilla "Ready or Not" report, published in February 2026, quantified coverage gaps across five traditional policy types (cyber, E&O, CGL, D&O, EPLI) and concluded that AI should be treated as its own risk classification category. Oliver Brew, Lockton Re's head of cyber centre of excellence, stated that "there are no sectors of the economy that are insulated from the potential impact of AI." Armilla's Baiju Devani identified "a growing gap between what insurers intend to cover and what they actually cover," a gap that widens with every AI exclusion filing that regulators approve.

Exposure Rating: The Only Viable Interim Methodology

Given the absence of credible loss data, exposure-based rating is the default methodology for AI liability pricing. Rather than deriving rates from historical loss experience, exposure rating estimates the expected loss cost from characteristics of the insured risk that correlate with loss potential. For AI liability, the key exposure variables being tested by the market include:

  • Volume of AI-generated outputs: Frequency of model invocations, number of end users, volume of generated content. Higher output volumes create more opportunities for hallucination, bias, and IP infringement claims.
  • Deployment context: Consumer-facing applications command significantly higher premiums than internal-use tools. Testudo's data showing that 32% of GenAI lawsuits are class actions is driven almost entirely by consumer-facing deployments.
  • Degree of human oversight: Agentic AI systems operating with limited human review carry higher exposure than AI tools used as decision support. Testudo's analysis noted that AI agents "materially increase risk exposure because they can act autonomously and execute at scale."
  • Industry and regulatory environment: Healthcare, financial services, and insurance deployments face sector-specific regulatory frameworks (HIPAA, state insurance regulations, the EU AI Act) that create additional liability pathways beyond common law tort.
  • AI governance maturity: Safeguards against hallucination, content moderation, data licensing compliance, model monitoring, and incident response capabilities. Armilla's 500+ point evaluation framework operationalizes this variable.
  • Foundation model dependency: Over 32% of publicly traded US companies deploy AI systems, and most depend on the same small set of foundation models. Vendor concentration creates correlated loss exposure that standard independent risk assumptions cannot accommodate.

The actuarial challenge is assigning relativities to these exposure variables without loss data to validate them. The initial relativities are judgment-based, informed by the litigation profile data and by analogy to cyber and professional liability pricing. Those relativities will need to be updated as closed-claim experience accumulates, creating a calibration cycle that is likely to produce significant rate volatility in the early policy years.

The Vendor Liability Gap

One structural factor that compounds the pricing problem is the distribution of liability between AI vendors and AI deployers. All four major foundation model providers (OpenAI, Anthropic, Google, and Microsoft) cap contractual liability at 12 months of fees and include express disclaimers on consequential damages. That contractual structure shifts the economic burden of AI failures to the deploying enterprise, which is the entity purchasing AI liability insurance.

For actuarial pricing, the vendor liability cap means that the insured's loss exposure includes not only the third-party claim against the insured but also the gap between the vendor's contractual liability cap and the actual loss. A company paying $500,000 per year for an AI platform has a contractual recovery ceiling of $500,000 from the vendor, regardless of whether the AI system causes $50 million in damages. The difference between the vendor's cap and the actual loss is the exposure that AI liability insurance must cover.

This gap is particularly acute for the 11.9% of GenAI lawsuits involving patent infringement and the 11.2% involving copyright infringement. Training data IP disputes are primarily the responsibility of the foundation model provider, but the vendor's liability cap means the deployer may bear the economic loss. Whether the deployer's AI liability policy responds depends on the specific policy language, which varies across the five standalone products currently available.

Why This Matters for Actuaries

The 978% GenAI litigation surge creates urgency across three actuarial functions:

Pricing actuaries building or maintaining technology E&O, cyber, or CGL rating plans need to address AI exposure explicitly. If the ISO exclusion is attached, the AI exposure is removed from the CGL book, and the rating plan should reflect the reduced exposure. If the exclusion is not attached, the rating plan implicitly covers AI liability at rates that were never calibrated for the 137% year-over-year filing acceleration. Either way, the actuary must document the treatment of AI exposure in the rate filing, and ASOP No. 25's credibility requirements apply to the data sources used.

Reserving actuaries at carriers writing any form of AI-adjacent coverage face the challenge of selecting development patterns without historical benchmarks. The 32% class action rate, the $5 million median demand, and the 99% severity concentration in the top 5% of cases provide the first data points for calibrating an initial reserve methodology. Analogy-based development patterns from cyber, E&O, and IP litigation are defensible starting points, but the weighting between analogue lines is entirely judgmental and should be documented for principal review.

Enterprise risk actuaries evaluating their own organization's AI deployment exposure should benchmark against the claim type distribution. If the organization uses AI for customer-facing decisions, the employment discrimination and consumer protection categories (8% and 6% of filings, respectively) represent direct regulatory exposure under the Colorado AI Act and similar state frameworks. The state AI law patchwork creates jurisdiction-specific liability that must be mapped against the organization's deployment footprint.

From tracking this market since the first standalone AI liability products launched, the central tension is clear: carriers need loss data to price responsibly, but they can only generate loss data by writing coverage. The five standalone AI liability writers are collectively running the experiment, pricing into an actuarial void with the expectation that the data will eventually validate or invalidate their assumptions. The 978% litigation curve ensures that the data will arrive. Whether it arrives fast enough to prevent a pricing correction depends on how quickly courts resolve the foundational questions of AI liability, product classification, fair use, and algorithmic bias that will define the loss distributions actuaries must eventually triangulate.

Further Reading

Sources