Insights Business| SaaS| Technology Why insurers are cutting AI out of liability insurance in 2026
Business
|
SaaS
|
Technology
•
Nov 18, 2026

Why insurers are cutting AI out of liability insurance in 2026

AUTHOR

James A. Wondrasek James A. Wondrasek
Why insurers are cutting AI out of liability insurance in 2026

Your liability policy may be covering AI losses right now without once naming AI. That is an accident of timing: policies drafted before enterprise AI became common never priced the exposure and never excluded it. In 2026 insurers are ending that silence with exclusions, led by CG 40 47, the generative AI exclusion the Insurance Services Office (ISO) has introduced on the commercial general liability (CGL) line, and W.R. Berkley’s “absolute” carve-out. Those forms cut AI out of liability cover, and they land in the who-pays picture.

What is “silent AI” insurance coverage, and why is it ending?

Silent AI is cover created by wording that never names AI. The wording neither covers nor excludes AI, so whether it responds to a claim is argued after the loss and was never priced beforehand. Most D&O, E&O and general liability policies predate everyday generative AI use, which is why they remain silent on AI.

Insurers never priced this exposure, because they never modelled it. An exposure no one can price is now excluded, and it no longer sits in the policy unresolved. The industry closed silent cyber the same way, when Lloyd’s Market Bulletin Y5258 forced policies to affirm or exclude cyber in 2019, the sequence the AI market is now repeating.

What is the ISO CG 40 47 generative AI exclusion introduced in January 2026?

ISO, a Verisk subsidiary, announced three optional generative AI exclusion endorsements, all effective January 2026. The broadest, CG 40 47, attaches to the commercial general liability (CGL) policy and strips cover for bodily injury and property damage (Coverage A) and personal and advertising injury (Coverage B), all “arising out of” generative AI. The narrower siblings split the scope: CG 40 48 excludes Coverage B only, and CG 35 08 applies to the products/completed operations hazard, the part of the CGL that responds when something you make, sell or have finished building causes harm.

Most US carriers already use ISO’s standard forms, so one optional clause spreads without any carrier drafting its own. ISO publishes the standardised language most US commercial insurers use, so a carrier can adopt CG 40 47 with a single filing, and that broad adoption makes an optional clause the standard across the United States. The ISO definition of generative AI is broad, which pulls vendor tools like Salesforce Einstein and HubSpot AI inside its reach. That breadth is why the liability question matters.

How broad is the “arising out of” generative AI trigger, and why does it matter?

“Arising out of” requires only a causal connection to generative AI, and that breadth decides how much cover is lost. The exclusion does not require AI to be the only cause, so a claim with any meaningful link to AI can still trigger it. The fight becomes how closely the AI was connected, as Shumaker notes.

Because the ISO definition reaches any generative tool, even unsanctioned use counts. Close to a third of employees have used shadow AI tools, so the exclusion reaches ordinary ChatGPT use too. W.R. Berkley’s version widens that trigger phrase even further to “based upon, arising out of, or attributable to”, keying it to AI used by any person or entity. On a mixed claim the insurer may still owe a duty to defend the whole action, and courts resist exclusions that swallow the policy into illusory coverage. That is why narrower wording gets negotiated before renewal.

Why are insurers adding “absolute” AI exclusions to D&O and E&O policies?

D&O and E&O are the lines that carry the AI-exposed claims, which is why insurers are adding absolute exclusions there. In 2025 W.R. Berkley filed PC 51380, which bars any claim “based upon, arising out of, or attributable to” the “actual or alleged use, deployment, or development of Artificial Intelligence by any person or entity”. Hunton Andrews Kurth called it the first “absolute” AI exclusion. Keyed to “actual or alleged” use, it fires when a plaintiff merely asserts AI involvement, catching AI-washing securities suits.

The exclusion spans D&O, E&O and fiduciary liability. On D&O, it catches mismanagement of AI strategy and shareholder derivative suits. On E&O, it reaches professional liability, which underwriters rate as the line most exposed to AI losses. Unlike the ISO forms, Berkley’s wording reaches traditional machine learning and rules-based automation.

Why is AI risk harder to underwrite than cyber risk?

AI is the harder risk to underwrite because it lacks the loss history cyber had. Cyber had a decade of breaches and ransomware losses before insurers could price it. AI has almost none of that history, and under the Berliner insurability framework generative AI fails maximum possible loss, loss frequency and information asymmetry, per CSIS. Carriers cannot see which models their insureds run, so they cannot price adverse selection.

Correlated losses are the bigger problem. One flaw in a widely deployed foundation model, whether GPT, Claude, Gemini or Llama, propagates to every organisation running it at once, turning AI into a gray swan: foreseeable in type, unknowable in size. Aon’s Kevin Kalinich put it as a carrier being unable to afford an AI provider mistake that lands as 1,000 or 10,000 correlated losses. The July 2024 CrowdStrike outage showed the same pattern in conventional software, and AI could produce it at larger scale. Real examples remain thin, shown in the loss cases.

Where can I find the ISO generative AI exclusion form language (CG 40 47)?

ISO doesn’t publish the exact wording openly, so you find the form language secondhand, in carrier filings, state filing records and law-firm or broker commentary. The exact wording circulates through carrier filings and commentary, with public snapshots compiled from state filing records, and law-firm explainers reproduce it. The practical shortcut is your own policy, where the endorsement schedule lists the forms and a search for “CG 40 47”, “CG 40 48”, “CG 35 08” or “generative artificial intelligence” finds them in the PDF. If none appear, the policy is still silent on AI.

What should I look for in my 2026 insurance renewal to spot AI coverage being quietly narrowed?

What you should look for sits in the wording of the form itself, because the narrowing rarely arrives as a letter. Fenwick describes AI loss being removed through the architecture of the form itself: insuring agreements, definitions, exclusions, carve-backs and endorsements. So the renewal question is what the new language has removed — and what the exclusions leave to contracts and the coverage gap.

An exclusion in one line can push a claim onto another, which is why the narrowing has to be read across the whole program. It shows up on CGL, D&O, E&O, employment practices and cyber, and the change sits in where each exclusion lands and what it carves out. The counter-moves are a carve-back (wording that restores cover for named claims like securities suits or defence costs), a sublimit (a reduced cap on AI losses), affirmative AI cover (capped around $25M to $50M), or a difference-in-conditions layer that fills the gap the exclusion leaves. Documented AI governance strengthens each. Start the review about 90 days before renewal, and treat it as an audit of the whole program, coordinated with vendor liability caps.

The 2026 exclusions move AI risk from insurance into a coverage gap and onto contracts. The pullback is a rational retreat. Insurer and vendor positioning each deserve the same scepticism when the residual risk is apportioned. For apportioning that residual risk, the allocation picture maps where it sits, and the contract clauses decide who pays it.

Frequently Asked Questions

Does my existing liability policy still cover AI losses right now?

If your policy predates the 2026 exclusions and never names AI, it likely still covers AI losses silently, until your next renewal. That cover was never priced and never excluded, so it only becomes visible after a claim. The practical question is not whether you currently hold it, but whether your next renewal removes it without fanfare.

Do these AI exclusions affect my cyber insurance policy too?

Not directly. The 2026 pullback attaches to commercial general liability, D&O, E&O and fiduciary lines, not to cyber policies. Cyber remains a separate line that prices its own exposures, and an AI incident is not itself a cyber peril. Still, check both at renewal, because an AI-related data breach could sit in the overlap between the two policies.

Do everyday tools like ChatGPT, Salesforce or HubSpot trigger the generative AI exclusion?

Yes. The ISO definition of generative AI is broad, so vendor-embedded tools such as Salesforce Einstein, HubSpot AI, Zendesk AI and ChatGPT fall inside its reach. The exclusion is not limited to AI you build or formally deploy. Even routine use of a software feature that draws on generative AI can be enough to fire the wording.

What happens if a claim involves both AI and non-AI causes?

A mixed claim can still trigger cover through the duty to defend, because a potentially covered allegation obliges the insurer to defend the whole claim in many jurisdictions. But the exclusion may let the insurer deny the AI-related portion, or argue the policy is illusory. This is exactly why brokers negotiate the trigger phrase narrower before renewal, not after a claim.

Is this only a United States problem, or does it affect Australian businesses?

The ISO form is a United States instrument, but the pullback is global. Global carriers and Lloyd’s syndicates that write Australian policies are adopting equivalent AI exclusions, and Australian liability wordings are being narrowed the same way. The trigger is the same unpriced exposure everywhere. Check your own wording at renewal, whatever the governing law.

Are insurers cutting AI out because they think it is too dangerous to cover?

Not exactly. The exclusions are an underwriting decision, not a verdict on AI’s danger. Insurers cannot price an exposure with no loss history and a correlated, systemic profile, so they retreat to exclusion rather than guess at a premium. Fear is not the driver. An unpriceable risk is, and that distinction matters when you negotiate the wording.

Can I buy standalone AI insurance instead, and how much cover is available?

Yes, an affirmative AI market is emerging as the exclusions push risk out of silent cover. For now it is young and capacity is capped, typically around $25M to $50M, often as a sublimit rather than full limits. Expect it to mirror the cyber market’s arc: narrow, expensive and bespoke at first, then standardised as loss data accumulates.

Can I negotiate the AI exclusion out of my policy at renewal?

Sometimes, and the trigger phrase is the first thing to attack. Because “arising out of” is read broadly, brokers push for a narrower causal standard, a carve-back for incidental use, or a sublimit instead of a full exclusion. Success depends on your AI governance evidence and your leverage. Small accounts rarely win outright removal, but a meaningful narrowing is realistic.

Does an AI governance program or red teaming help me keep coverage?

It strengthens the negotiation, even though it is not a substitute for loss history. Insurers are excluding AI because of information asymmetry, so documented governance, testing and red teaming attack that asymmetry directly. It shows the risk is controlled, which supports a carve-back or a sublimit. What it does not do is give actuaries the claims data they price from.

How is silent AI different from silent cyber?

They are the same phenomenon applied to different exposures: cover a policy provides by accident, never priced and never excluded. The difference is timing and precedent. Silent cyber was closed by Lloyd’s Market Bulletin Y5258 in 2019 and became a standalone affirmative market. Silent AI is now following the identical sequence, and the 2026 exclusions are the same turning point.

What is a gray swan, and why does it matter here?

A gray swan is a risk you can foresee in type but not in size. A flaw in a shared foundation model such as GPT, Claude or Llama is predictable; how many insureds it hits at once is not. That combination of foreseeable and unknowable defeats pricing, because an underwriter cannot set a premium without knowing the worst-case aggregation.

AUTHOR

James A. Wondrasek James A. Wondrasek

SHARE ARTICLE

Share
Copy Link

Related Articles

Need a reliable team to help achieve your software goals?

Drop us a line! We'd love to discuss your project.

Offices Dots
Offices

BUSINESS HOURS

Monday - Friday
9 AM - 9 PM (Sydney Time)
9 AM - 5 PM (Yogyakarta Time)

Monday - Friday
9 AM - 9 PM (Sydney Time)
9 AM - 5 PM (Yogyakarta Time)

Sydney

SYDNEY

55 Pyrmont Bridge Road
Pyrmont, NSW, 2009
Australia

55 Pyrmont Bridge Road, Pyrmont, NSW, 2009, Australia

+61 2-8123-0997

Yogyakarta

YOGYAKARTA

Unit A & B
Jl. Prof. Herman Yohanes No.1125, Terban, Gondokusuman, Yogyakarta,
Daerah Istimewa Yogyakarta 55223
Indonesia

Unit A & B Jl. Prof. Herman Yohanes No.1125, Yogyakarta, Daerah Istimewa Yogyakarta 55223, Indonesia

+62 274-4539660
Bandung

BANDUNG

JL. Banda No. 30
Bandung 40115
Indonesia

JL. Banda No. 30, Bandung 40115, Indonesia

+62 858-6514-9577

Subscribe to our newsletter