Generative AI liability is being re-drawn - here's what insurers need to do now
As banks and enterprises scale generative AI and lawsuits stack up, carriers are redrawing coverage lines. Verisk has endorsed an optional exclusion for generative AI within the ISO General Liability program, which many carriers use for GL forms and rules. Starting in January, expect some carriers to pull AI-related exposures out of GL, while others keep them if they have in-house expertise.
That split will push risk into specialty markets. Demand is real: in a Geneva Association survey of 600 businesses, over 90% said they would value insurance for gen AI risks, and more than two-thirds would pay at least 10% more for it.
What's being excluded - and why
The ISO option lets carriers "laser out generative AI," said Joseph Lam, vice president of general liability at Verisk. For insurers without the expertise (or appetite) to underwrite model risk, hallucinations, algorithmic bias, and IP exposures inside GL, the exclusion creates a clean boundary - and a need for affirmative AI coverage elsewhere.
The loss themes already here
- Algorithmic errors and hallucinations that trigger customer harm and bad advice claims
- Bias and discrimination in underwriting, lending, and employment decisions
- IP, copyright, patent, and trademark disputes tied to training data and outputs
- AI washing - enforcement for false "we use AI" marketing claims
- Broader cyber attack surface and data leakage via gen AI tools
- Misleading content that erodes customer trust and brand value
Real cases are setting expectations. Deloitte Australia refunded the government after an AI-generated report included fake citations and incorrect claims. An adjudicator held Air Canada responsible for its bot's incorrect guidance on bereavement fares and awarded damages. The New York Times sued OpenAI and Microsoft over alleged use of copyrighted articles in training. Wells Fargo faced class actions alleging algorithm-driven discrimination in mortgage and refinance practices.
Regulators are acting on AI washing: the SEC charged two advisors for misleading AI claims, resulting in civil penalties. See the SEC's release for details: SEC charges for "AI washing".
Specialty markets are stepping in
New entrants are building products to fill the gap. Testudo, backed by Lloyd's of London, offers a gen AI liability policy that aligns with the ISO exclusion and covers suits stemming from AI use - from copyright and discrimination to IP, patent, and trademark claims. It also includes personal injury, bodily injury, defamation, libel, slander, and data privacy, with initial limits up to $10 million.
Testudo tracks AI-related litigation globally and uses those cases to inform underwriting and pricing. "Once you have that map of the real-world risk, we can underwrite or decline based on it," said founder George Lewin-Smith.
Armilla AI, also Lloyd's-backed, insures AI liability for model errors, hallucinations, regulatory violations, and data leakage. CEO Karthik Ramakrishnan describes an evolution similar to cyber: moving from questionnaires to evidence-based external evaluation and continuous monitoring to quantify reliability and governance.
Vouch says it covers algorithmic bias, IP infringement, defense costs for AI-specific regulatory investigations, and damages from AI products or services. NSI Insurance Group is preparing offerings as well. If mainline carriers exclude, specialists will keep filling the gap.
What this means for carriers and brokers
Prepare for a two-track market: GL with an AI exclusion, plus affirmative AI policies or endorsements. Buyers want "zero gaps" across cyber, E&O/tech, D&O, media/IP, and now AI. Pricing will hinge on use cases (e.g., lending vs. marketing), model governance, and proof of controls - not just questionnaires.
Underwriting checklist: the signals to price and select risk
- Use-case inventory: Where is gen AI deployed? High-severity areas (credit, underwriting, employment) flagged separately.
- Model lineage: Model types (proprietary vs. third-party), training data provenance, license posture, and output filters.
- Controls: Human-in-the-loop, pre-deployment testing, red-teaming, bias audits, calibration, monitoring, rollback plans.
- Data safeguards: PII handling, prompt injection defenses, output logging, retention, and access controls.
- Legal guardrails: IP indemnities, acceptable-use policies, vendor warranties, and incident notification terms.
- Governance: Model risk committee, approval gates, versioning, audit trails, and board-level reporting.
- Claims readiness: Evidence capture, forensic playbooks, counsel engagement, and PR response for AI incidents.
- Aggregation: Scenario tests for correlated events (framework or patch failures across many insureds) and reinsurance fit.
Program structure for insureds
- Ask directly: Will your GL include the ISO gen AI exclusion? If yes, where will AI risks be affirmatively covered?
- Coordinate towers: Map cyber, tech E&O, media/IP, D&O, and AI policies to avoid gaps or double coverage.
- Tighten contracts: Secure vendor indemnities for training data/IP, usage controls, and audit rights.
- Prove governance: Maintain testing reports, bias assessments, and change logs - underwriters will price off evidence.
- Incident playbooks: Define triage for hallucinations, bias complaints, data leakage, and AI washing allegations.
Watchlist: what could move pricing quickly
- Regulatory actions from the SEC, DOJ, FTC, and EEOC on discrimination, AI washing, and deceptive practices
- Class action outcomes tied to algorithmic bias and consumer harm
- Copyright and training data case law that clarifies liability up and down the model supply chain
- Major model vulnerabilities that trigger correlated losses across many insureds
Skill up your team
Underwriting AI risk now depends on how well you can assess model behavior and governance. If you're building internal expertise across underwriting, claims, and broker teams, these resources can help:
Bottom line: GL exclusions will push gen AI risk into specialty lines. If you can measure control quality and loss potential with evidence - not vibes - you'll write the right risks at the right price and avoid concentrated surprises.
Your membership also unlocks:
