Cyber and AI Liability Converge Into One Digital Risk Line: Pricing the Unbundled Exposure

On January 1, 2026, Verisk's ISO Core Lines Services released three optional generative AI exclusion endorsements for the commercial general liability form. ISO forms underpin roughly 82% of US property and casualty policies, and carriers are filing the broad version at about three times the rate of the narrow one. State regulators have approved more than 80% of those filings. AI exposure is leaving baseline commercial liability faster than any product is arriving to replace it.

Key Takeaways

• CG 40 47 excludes both Coverage A and Coverage B for claims arising out of generative AI. CG 40 48 excludes Coverage B alone, and CG 35 08 removes Section I products and completed operations coverage.
• "Arising out of" requires only a causal connection, not proximate causation, so a vendor-embedded AI tool peripheral to a loss can still trigger the exclusion.
• Generative AI lawsuits grew 978% between 2021 and 2025 to more than 700 cumulative filings, with year-over-year growth accelerating from 59% to 137%.
• Roughly $4 million is the average AI-related case settlement on Testudo's litigation database, which together with the filing trend is the frequency and severity anchor being used in place of triangles.
• Standalone capacity runs to $9.25 million per insured at Testudo and $25 million per organization at Armilla, against a CGL market measured in the hundreds of billions.

What Left the Policy

The three endorsements differ in reach, and the reach is what the filing mix reveals.

Endorsement	Scope	What It Excludes	What It Preserves
CG 40 47	CGL Coverage Part	Coverage A + B (BI, PD, personal/advertising injury)	Nothing AI-related
CG 40 48	CGL Coverage Part	Coverage B only (personal/advertising injury)	Coverage A (BI, PD)
CG 35 08	Products/Completed Ops	Section I (BI, PD from AI in products)	Premises/operations exposure

All three define generative AI as a machine-based learning system or model trained on data with the ability to create content or responses, including text, images, audio, video or code. That captures large language models, image generators and code assistants, and does not explicitly reach predictive models that classify or score without generating content, though that boundary moves with every model update.

The operative phrase is narrower in words and wider in effect. "Arising out of" requires a causal connection rather than proximate causation. An insured whose vendor-embedded AI contributed to a loss can find the exclusion triggered even where the AI component was peripheral to the claim.

Adoption followed the usual S-curve. Within weeks of the January release, W.R. Berkley, Cincinnati Financial, Frederick Mutual and Philadelphia Insurance filed their own wording, and Berkshire Hathaway, Chubb and Travelers requested AI-related exclusions across their general liability books. Testudo co-founder George Lewin-Smith predicted 95% of carriers would employ the exclusions immediately. Filing data through April 2026 is tracking close to that in the large-account segment, and the preference for CG 40 47 over CG 40 48 at roughly three to one says carriers expect bodily injury and property damage claims from AI systems, not only advertising injury and IP.

Pricing a Line With No Triangles

The exclusion takes effect at the next renewal with no transition, which leaves the exposure somewhere.

It is not with the developers. Standard terms from the foundation model providers cap contractual liability at 12 months of license fees and offer no performance warranties. The enterprise retains liability for a model it did not build, trained on data it did not select, on an architecture it cannot audit.

Nor is it fully with the specialty market yet. Testudo writes $9.25 million per insured after adding Atrium and QBE to its reinsurance panel, Armilla up to $25 million per organization, and Corgi up to $50 million following a $160 million Series B at a $1.3 billion valuation. Munich Re's aiSure has run since 2018 and its HSB subsidiary added SME cover in March 2026. Summed, that is a few billion dollars against a CGL market in the hundreds of billions.

The pricing problem is that none of it has loss triangles. What it has instead is a litigation series. Generative AI lawsuits grew 978% between 2021 and 2025 to more than 700 cumulative filings, with annual growth accelerating from 59% between 2023 and 2024 to 137% between 2024 and 2025, and filings in the first four months of 2025 up 81% on the prior year.

The composition is the part that connects to the form. Patent infringement is 11.9% of cases, copyright 11.2% and personal injury including privacy and data misuse 10.2%. Those are Coverage B claims, which is precisely what both CG 40 47 and CG 40 48 remove. Average settlements run approximately $4 million on Testudo's database, from the pre-precedent phase of the cycle.

That gives a frequency signal and a severity anchor without credible experience, which is what standalone cyber underwriters worked from in 2014 and 2015. The underwriting response is to substitute a proxy: Armilla conditions coverage on continuous model quality assessments and prices off model inventories, monitoring, audit trails and incident response, treating governance maturity as a stand-in for frequency. It is defensible and it is not the same thing as a loss cost.

The Exclusion Is Wider Than the Product Replacing It

The definitional gap between what is being excluded and what is being sold is the structural problem.

ISO's wording is scoped to generative AI. The management liability wording is not. W.R. Berkley's Artificial Intelligence Exclusion (Absolute) for D&O removes any claim based upon, arising out of or attributable to the use, deployment or development of AI, and defines AI as any machine-based system that infers how to generate outputs such as predictions, content, recommendations or decisions. That reaches traditional machine learning in pricing, underwriting and claims triage, not just generative models.

Hamilton Insurance Group has filed a generative AI exclusion for professional liability naming ChatGPT, Bard, Midjourney and DALL-E, and design professional E&O is being targeted with comparable language. AIG and Great American have sought clearance for AI exclusions in management liability.

Stack those and an enterprise deploying AI across customer service, marketing, product design and internal operations can be excluded under CG 40 47 in CGL, under Berkley's absolute wording in D&O, and under Hamilton's carve-out in E&O, with incidental use of AI embedded in enterprise software from Salesforce, Microsoft or Google potentially sufficient to trigger it.

What the specialty market offers against that is generative-AI-shaped cover at single-digit or low-double-digit millions per insured. The exclusions are being written to the broad definition of AI and the replacement capacity is being written to the narrow one, which leaves the predictive-model exposure inside neither.