CG 40 47: The ISO Generative AI Exclusion Explained

By Joel R. Singh · Last verified: 2026-07-29 · How we verify

What CG 40 47 is and what it does to your CGL policy

Standard commercial general liability policies were written before generative AI existed as a commercial product. They cover bodily injury, property damage, and a third bucket called personal and advertising injury, which catches things like defamation, copyright infringement in advertising, and invasion of privacy. None of that language was drafted with AI-generated outputs in mind, which left carriers holding unpriced "silent AI" exposure on every CGL they wrote. Verisk, which publishes the ISO standard forms that underpin the majority of US commercial property and casualty policies, filed a suite of AI exclusion endorsements with state regulators in mid-2025. Those forms became available for carrier attachment on CGL renewals beginning January 1, 2026.

CG 40 47 is the broadest of three related endorsements in that filing. It modifies both Coverage A (bodily injury and property damage liability) and Coverage B (personal and advertising injury liability) by adding an exclusion that applies whenever the triggering harm arises from generative AI. The effect is a categorical removal, not a sublimit or a condition. If generative AI is in the causal chain, the exclusion applies regardless of whether you built the AI system yourself, licensed it from a vendor, or used a third-party tool with AI built in.

The endorsement's definition of generative AI, as reported across multiple trade and carrier sources citing the ISO filing, is: 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 every mainstream tool from ChatGPT to AI image generators to code-completion assistants. The full ISO form text is paywalled through Verisk's ISO Electronic Rating Content platform; the definition above is attributed to primary sources including the Big I Virtual University briefing on the filing and the Gallagher advisory on the same subject.

The operative exclusion language

ISO form text is proprietary and not reproduced here verbatim. Based on consistent reporting from the Big I Virtual University, Gallagher, and the Gridex coverage analysis of the filing, the operative exclusion language in CG 40 47 reads substantially as follows:

Two phrases in that language deserve attention. First, "arising out of" is one of the broadest causal standards in insurance law. Courts have consistently held that this phrase requires only a causal connection between the AI and the harm, not direct or proximate causation. A business whose employee used an AI drafting tool to produce marketing copy that turns out to be defamatory does not get to argue that the human reviewed and approved the copy; the AI is in the origin story, and that is enough. Second, "or attributable to" compounds the breadth further, catching scenarios where the AI's contribution is partial or indirect.

The practical consequence is that the exclusion is not limited to AI you built or AI in your product. It reaches AI tools embedded in software your business licenses, AI features your vendors use in services they deliver to you, and AI outputs that pass through multiple hands before causing harm. Businesses that use AI anywhere in a customer-facing workflow carry meaningful exposure under a policy where this endorsement is attached.

Two scenarios where this exclusion bites

Scenario 1: AI-generated marketing content triggers an advertising injury claim. A regional financial services firm uses an AI writing tool to produce a monthly blog series explaining investment products. One post, reviewed but not substantially rewritten by a human editor, makes a claim about a competitor's fee structure that is materially inaccurate. The competitor sues for defamation. Before CG 40 47, that claim would flow squarely into Coverage B as a personal and advertising injury. With CG 40 47 attached, the carrier declines: the content arose from generative AI, and the exclusion applies. The fact that a human reviewed the post is not a defense under "arising out of" language. The firm has no CGL coverage for the defense costs or any judgment, and no standalone AI liability policy.

Scenario 2: AI-assisted clinical decision support produces a patient harm. A healthcare services company licenses an AI tool that synthesizes patient data and generates treatment recommendations for clinicians. A recommendation based on an incomplete data set leads to a medication error and a patient injury. The injured patient's family files a bodily injury claim against the company. The carrier points to CG 40 47: the injury arose from an AI-generated output. The company assumed its CGL would function as a backstop behind its professional liability policy; that assumption was wrong. The CGL exclusion leaves a gap in the coverage tower exactly when a claim is large enough to exhaust the professional liability limit.

Both scenarios share a common feature: the harm traces to AI output, and the policyholder did not know the exclusion had quietly attached at renewal. That is the typical pattern. Brokers often do not flag new endorsements unless the client asks directly, and AI exclusions are new enough that the ask is not yet standard practice.

The companion forms: CG 40 48 and CG 35 08

CG 40 47 was filed alongside two narrower endorsements that carriers can use instead of or in addition to it. CG 40 48 limits the exclusion to Coverage B only, meaning it removes advertising injury and personal injury coverage for generative AI but leaves bodily injury and property damage intact. CG 35 08 targets the products and completed operations hazard specifically: it excludes bodily injury and property damage arising from generative AI that occurs after the work or product has been delivered, which is the relevant risk for companies that ship AI-enabled software, hardware, or completed services.

The choice of which form to attach is a carrier underwriting decision, not a policyholder election. A manufacturer of AI-assisted devices might find CG 35 08 on its policy; a professional services firm heavy in marketing and content might find CG 40 47 or CG 40 48. There is no standard practice yet on which form carriers prefer; the exclusion tracker on this site documents which forms we have verified on specific carrier renewals as that data becomes available.

What fills the gap and which carriers are writing it

CG 40 47 removes coverage that CGL policies were providing silently, but it does not create replacement coverage. Filling the gap requires either an affirmative AI endorsement from your existing CGL carrier (rare and not yet widely offered) or a standalone AI liability policy purchased separately. The carrier comparison table on this site tracks every verified standalone product in the US market. A summary of the relevant options follows.

Relm Insurance writes dedicated AI liability through its PONTAAI product, structured as a difference-in-conditions / excess wrap that responds where primary CGL or tech E&O does not. Relm is a Bermuda-domiciled specialty insurer and one of the earliest entrants in the affirmative AI market. Its products cover media liability for AI-generated content, regulatory liability, and third-party claims from AI outputs.

Testudo launched in January 2026 as a managing general agent specifically positioned to fill the CG 40 47 gap. Its policy covers hallucinations and negligent errors in AI outputs, IP infringement by AI-generated content, unauthorized data disclosure, bodily injury and property damage arising from reliance on AI output, and regulatory proceedings. Capacity at launch was up to $9.25 million^* per occurrence, backed by Apollo, Atrium, and QBE syndicates; in August 2026 we received information that the limit was increased to $10 million per insured. Testudo is one of the few products explicitly designed with CG 40 47 in mind.

Vouch Insurance packages affirmative AI coverage into its technology company product suite alongside tech errors and omissions, cyber, and D&O. Its AI coverage addresses hallucinations, algorithmic bias and discrimination claims, regulatory investigations, and IP disputes arising from AI-generated content. Vouch acquired Corix's AI book and was itself acquired by Hiscox in August 2025.

CFC Underwriting embeds affirmative AI language across its tech E&O, professional liability, and media lines rather than writing AI coverage as a standalone product. For businesses whose primary exposure is advertising injury and content-related claims, CFC's approach may provide the most efficient path to filling the Coverage B gap created by CG 40 47.

HSB (a Munich Re subsidiary) launched an AI liability product in March 2026 designed specifically for small and mid-sized businesses, available in all 50 states through partner carriers. It covers bodily injury, property damage, and advertising injury arising from AI-generated marketing, advertising, blog posts, and social media content. It is the only verified product in this market with an explicit small-business target and broad state availability.

For the full verified dataset including policy limits, admitted versus surplus lines status, revenue floors, and last-verified dates for each product, see the carrier comparison table. For the intersection of professional liability and AI coverage, see Does E&O Cover AI?

Adoption, timeline, and what it means at renewal

Verisk made its multistate filing in mid-2025. State insurance departments began processing approvals in the second half of 2025, and the forms became available for carrier attachment on policies renewing on or after January 1, 2026. As of mid-2026, regulators had approved the forms in the majority of states where they were filed; Florida, Connecticut, and Maryland processed the highest volumes of AI exclusion filings across carriers.

Adoption by any individual carrier is voluntary. CG 40 47 is a standard form, not a mandate. But the trajectory is clear: carriers including W.R. Berkley, Chubb, Travelers, Berkshire Hathaway, Cincinnati Financial, and Philadelphia Insurance have either adopted CG 40 47 directly or filed proprietary AI exclusion language with similar scope. W.R. Berkley's proprietary version, Form PC 51380, is notable for its breadth: it excludes any actual or alleged use of AI across directors and officers, errors and omissions, and fiduciary liability lines, going further than the ISO standard in some respects.

The pattern is the same one that played out with cyber exclusions in the 2010s. Carriers started attaching cyber exclusions to CGL and property policies before most policyholders knew what a cyber exclusion was. Businesses discovered the gap only when they had a claim and the carrier declined. AI exclusions are following the same curve, compressed into a shorter timeline because generative AI moved from novelty to enterprise tool faster than any previous technology.

Buyer and broker checklist

Last verified: 2026-07-29. Source links checked on that date. ISO form text is paywalled; operative language is reported from authoritative secondary sources. Report an error.