AI Liability Insurance: How Tech E&O, D&O, and Cyber Respond to AI Claims

AI liability is real, and the policies you bought before 2023 were not built for it. The SEC filed 53 AI-related securities cases in 2025. A teen’s family is suing a chatbot company over his suicide. Air Canada was ordered to honor a discount its bot invented. Foundation-model labs are defending billion-dollar copyright suits. Workday is fighting a class action over its hiring algorithm.

Your Tech E&O, D&O, and cyber policies all touch these claims. None was written for them. This guide maps the five claim categories courts are seeing now, which policy responds when each one lands, and where coverage runs out.

For the buy-side companion (what to buy at each stage and what it costs), see our guide to Insurance for AI Companies.

The five categories of AI liability claims

AI liability is not one risk. It is five, each with a different theory of harm and a different path through your insurance.

Bad AI output creates simple but serious claims. A New York law firm fed ChatGPT a research prompt in 2023. The model invented case citations. The firm filed them. The judge sanctioned the lawyers and opened a discipline file with the bar. The AI worked as built. It still caused real harm.

Output liability scales with stakes. A chatbot suggesting the wrong movie is harmless. A hiring algorithm rejecting protected classes is serious. A medical AI missing a tumor is catastrophic. The claim category is the same; the limits you need are not.

AI models learn from existing work and generate new work. Both create exposure.

By the end of 2025, more than 40 active lawsuits sat in U.S. federal court against the major foundation-model labs over training data: The New York Times v. OpenAI, the Authors Guild class action, Getty Images v. Stability AI, the music publishers’ suit against Anthropic. Each case asks the same question. Did the lab need a license to train on this content?

A separate set of cases targets outputs. Google’s AI Overviews summarize web results without always linking back. Wolf River Electric sued Google in 2025, claiming the feature stole the value of their content. Generative-image suits raise the parallel question: when a model produces an image that resembles a copyrighted work, who is liable, the user or the model maker?

The legal answers will take years. Underwriters are already pricing the risk. Every IP infringement quote we place now asks where your training data came from and whether you can prove you had rights to use it.

AI systems process large volumes of personal data. Careless handling creates liability. So does a model that leaks private information it should have forgotten.

Consider a company that builds an AI customer service bot trained on support transcripts holding names, account numbers, and payment information. If the model memorizes that data and reproduces it in a response, the company has a privacy breach. If a user asks about another customer’s account and the bot returns it, that is a data privacy incident under most state laws.

The regulatory frame is tightening. The EU AI Act imposes hard requirements on high-risk systems processing personal data. Colorado’s AI Act creates obligations for companies running automated decision systems. More states are following each session. Each new law widens what counts as improper handling of personal data in an AI context.

A hiring AI trained on a decade of past hires learns the patterns in that data. If the company hired fewer women into technical roles, the model learns to rate male applicants higher. No one wrote that rule. No one intended it. The model still discriminates, and Title VII does not require intent.

Workday is currently defending a federal class action filed in 2023 that survived a motion to dismiss in July 2024. The complaint alleges Workday’s AI screening tool systematically rejects Black, older, and disabled applicants in violation of Title VII, the ADEA, and the ADA. The court’s ruling was the watershed: it held Workday could be liable as an “agent” of the employers using its software, opening AI vendors to direct claims from applicants they never met. The case was certified as a collective action in 2025.

HireVue settled a similar EEOC complaint over facial-analysis hiring tools in 2021 and dropped the feature entirely. The settlement followed an FTC complaint alleging the company’s bias claims were unsupported. iTutorGroup paid $365,000 in 2023 to settle EEOC charges that its AI rejected female applicants over 55 and male applicants over 60.

Coverage on a hiring algorithm claim runs through three policies, and the answer turns on who is suing.

When the AI vendor is sued by the employer who bought it, Tech E&O responds. The vendor’s product failed to perform as represented. Defense and damages flow from there.

When applicants sue the employer using the AI, EPLI responds. The employer is the named defendant, and the claim is a discrimination claim under employment law.

When applicants sue the AI vendor directly, as in Workday, the answer is unsettled. Tech E&O policies vary. Older forms exclude “automated decision-making” outright. Newer forms cover algorithmic claims but require evidence of bias testing, disparate-impact audits, and human-in-the-loop review. A vendor that cannot produce that evidence may find its Tech E&O carrier deny the claim under the policy’s “reasonable care” duty.

Underwriters now require, before binding: documented bias testing across protected classes, an annual third-party audit, a written human-review protocol for high-stakes decisions, and a process for applicant appeals. Companies with this evidence pay less for E&O. Companies without it are increasingly being declined.

In October 2024, the family of 14-year-old Sewell Setzer III filed suit against Character.AI in the Middle District of Florida after the teen took his own life. The complaint alleges the chatbot platform, marketed without meaningful age gates, engaged in romantic and sexually explicit role-play with the minor, told him to “come home” to it, and failed to flag or interrupt clear suicidal ideation. A second suit filed in December 2024 in Texas raises similar claims for two more minors.

The legal theories are stacked: product liability for defective design and failure to warn, negligence for failure to monitor and moderate, misrepresentation for marketing the app as safe for teens, and wrongful death. Character.AI moved to dismiss on First Amendment grounds. In May 2025, a federal judge denied the motion on the core claims, ruling that chatbot outputs are not categorically protected speech and that the company can be sued for the design and deployment of the product itself.

Five policies could respond to a claim like this, and which one pays depends on how the loss is framed.

Tech E&O answers if the claim is framed as defective software performance. The product failed to do what a reasonable user would expect. This is the cleanest fit for a chatbot-harm case.

Media liability answers if the claim turns on what the AI generated. When the bot’s outputs are treated as published content, the policy covers defense against claims of harmful or defamatory speech. The Character.AI case is testing whether AI outputs are media content or product behavior, and the answer will shape coverage for every consumer-facing AI product.

Product liability answers if the AI is embedded in something a court treats as a product: a smart speaker, a robot, an autonomous vehicle. Standalone software is rarely a “product” under traditional product-liability law, but courts are starting to push at that line.

General liability covers bodily injury and emotional distress in narrow cases. Most GL policies now carry “professional services” and “software” exclusions that cut AI claims out, so coverage here is shrinking.

D&O answers separately, on the governance side, if shareholders sue alleging the board ignored known safety risks while making public statements about the product’s guardrails.

Underwriters writing coverage for consumer-facing AI now ask, every time: what age verification do you run, what content moderation runs on outputs, what crisis-detection language triggers a handoff to a human or a suicide hotline, what logs do you keep, and how fast do you respond to a flagged conversation. A company that cannot answer these specifically is either declined or quoted with a six-figure deductible.

The Character.AI case is the one to watch. The court’s final ruling and any settlement will set the standard of care every consumer-AI company is held to.

How existing policies respond to AI claims

Each AI claim type triggers a different policy. Here is how the main lines respond when a claim lands, and where they fall short.

When a client sues over bad AI output, Tech E&O is the first policy that responds. Your AI advises wrongly, the customer loses money, they come to you. The policy funds defense and damages.

Pre-2023 Tech E&O forms often exclude “automated decision-making” or “algorithmic systems” outright. Modern forms cover AI explicitly but define it narrowly. Read the AI definition and the “reasonable care” duty before you rely on the policy. A company that deployed a model with no bias testing, no model evaluation, and no logs has a real chance of being denied.

When shareholders sue founders or the board over what the company said about its AI, D&O responds. The SEC’s “AI washing” enforcement actions have made this the fastest-growing D&O exposure of the last 24 months. More on that below.

D&O covers defense and settlement, with two big carve-outs. Intentional misconduct is excluded; if the board knew an AI capability claim was false and approved it anyway, no coverage. And claims covered by other policies (E&O, product liability) are typically pushed off D&O onto the underlying line.

When an AI system leaks personal data, cyber responds. The model memorizes training data and reproduces it. The chatbot exposes one customer’s account to another. The vector store gets breached.

Cyber pays for notification, credit monitoring, regulatory fines where permitted, legal defense, and business interruption. The gap: cyber covers “unauthorized access.” If your AI processes data you were authorized to access and leaks it through bad design, your carrier may dispute whether that fits the policy trigger. Newer forms close this gap with affirmative AI language. Older ones do not.

When an AI system causes physical injury or property damage, GL is the trigger. A robot strikes a worker. An autonomous tool damages a building. The grant is real, but the exclusions matter more. Most GL forms now exclude “professional services,” “software,” and “cyber events.” When the harm flows from software behavior rather than a physical part, GL may not respond at all.

When AI is embedded in a physical product, product liability responds. Vehicles, robots, medical devices, smart appliances. The grant assumes a traditional defect: a part fails, a weld breaks. AI creates a different defect: the system works as built but produces biased or unsafe outputs.

Some carriers are now writing AI-aware product liability that contemplates algorithmic defects. Others are excluding AI claims from product liability entirely. The split is sharp. Read the form.

Where the coverage gaps live

Five policies form your AI liability program. None was built for AI. All have gaps.

Your hiring algorithm rejects women for technical roles. Tech E&O might respond to a claim from your customer. EPLI covers the applicants who sue your customer. But if you deploy the algorithm in your own hiring and applicants sue you, only EPLI responds. Many AI vendors do not carry EPLI on the shelf and find the discrimination claim falls into a void.

A bank deploys an AI lending algorithm making credit decisions without human review. The algorithm denies credit on factors the bank did not intend. Is that a decision the bank made (bank responsibility) or a decision the algorithm made (vendor responsibility)? Insurance has historically assumed companies make decisions. Policies have not contemplated systems making them independently. Liability sits in ambiguous territory across multiple policies.

Did your company train on copyrighted material? Did you have a license? If copyright holders sue, which policy responds? Most Tech E&O and GL forms exclude IP. Media liability covers some claims tied to output but not the broader training-data question. IP infringement insurance is the right line, and it is written by a small group of specialists. Underwriters ask detailed questions about training-data provenance. Policy language has not caught up.

The EU AI Act and Colorado’s AI Act create compliance obligations: documentation, impact assessments, audits, human-oversight protocols. Insurance does not cover routine compliance buildout. First-party cyber may cover some costs tied to a breach, but the audit and documentation work is on your operating budget.

An employee uses ChatGPT or a free image model on company data without IT approval. The data is now in the tool’s prompts and may be in its training data. Your insurance does not explicitly address this. Cyber may respond if a breach is later traced to it, but policies assume your company controls the systems handling sensitive data. Shadow AI assumes the opposite.

The D&O exposure from “AI washing”

The SEC filed 53 AI-related securities cases in 2025, a record. Most followed the same pattern. A company tells investors it uses AI to improve a result. The claim is broad, undocumented, or untrue. Earnings later miss, or a journalist or short-seller pokes at the AI, or a former employee files a whistleblower complaint. The stock drops. Shareholders sue. The SEC opens an investigation. Founders and directors are named personally.

In March 2024, the SEC settled charges against two investment advisers for “AI washing” in their marketing materials. Both claimed to use AI in investment decisions; neither had deployed AI at any meaningful scale. The fines were modest ($175,000 and $225,000), but the precedent was set. The SEC is treating AI capability claims with the same rigor as financial disclosures.

The 2025 wave has gone further. Securities class actions now routinely allege:

• Overstated AI deployment (“our AI handles 80% of claims” when the real figure was 8%)
• Unsupported performance claims (“our AI reduces fraud by 40%” from a 30-account pilot)
• Failure to disclose material AI risks (model drift, training-data exposure, regulatory pending)
• Inadequate board oversight of AI

D&O covers the defense and settlement. Two things matter when the claim lands.

First, the carrier will ask whether the board approved the AI claim and what diligence backed it. A board that signed off on capability statements with no documentation has thinner coverage than one with a record of testing, vendor due diligence, and review.

Second, the carrier will read the policy for AI carve-outs. Berkley has circulated an exclusion that bars a broad set of AI claims from D&O, E&O, and fiduciary policies. Other carriers are following. A specialist broker negotiates these out at placement. A generalist signs them.

What boards now have to document:

• A written AI governance policy
• A designated AI risk owner at the executive or board level
• Pre-release review of any public AI capability claim
• An annual disclosure review covering material AI risks
• Board minutes showing AI was discussed, not just delegated

The companies that get this right pay 10 to 25 percent less for D&O than those that do not. The companies that get it wrong face declination at renewal.

State and federal AI regulation

The regulatory frame is fragmenting fast. The EU AI Act took partial effect in 2024 and phases in through 2027. It imposes obligations on any company placing AI in the EU market. High-risk systems (employment, credit, education, critical infrastructure) face extensive requirements: human oversight, bias testing, impact assessments, technical documentation, post-market monitoring.

In the U.S., the picture is messier. No federal AI liability framework exists yet, though Congress is considering proposals. States are acting. Colorado’s AI Act, the first comprehensive U.S. AI law, takes effect in February 2026. It covers automated decision systems impacting civil rights, with a duty of reasonable care, mandatory impact assessments, and consumer notice obligations. California, Texas, New York, and Illinois have all introduced 2026 bills with overlapping but different requirements.

The fragmentation creates compliance challenges. A company deploying AI in Colorado, California, and the EU faces three different regulatory regimes with three different documentation requirements. Build to the most stringent, or build differently for each jurisdiction.

Insurance has not caught up. Policies do not explicitly cover state AI law compliance. They may cover costs tied to an incident or breach, but they will not cover the routine infrastructure: impact assessments, bias testing, human oversight protocols, documentation. Budget for AI regulatory compliance separately. It is a standard business expense now, not an insurable risk.

Frequently asked questions

Usually yes, when the claim comes from a customer who suffered a financial loss from your AI’s output. The policy responds to “errors and omissions” in your software service, and a wrong answer fits. Two cautions: older policies may exclude “automated decision-making,” and modern policies require you to show “reasonable care” in design and testing. A hallucination claim against a vendor that ran no model evaluations and kept no logs is at real risk of denial.

Yes, if the claim names the directors and officers personally and the policy has no AI exclusion. Defense is almost always covered. Settlement is covered unless the carrier can prove intentional misconduct, in which case the conduct exclusion bars indemnity. Read the AI clause on every D&O quote. Some carriers are now adding broad AI carve-outs that gut the protection.

Both, depending on who is sued. If applicants or employees sue the employer using the AI, EPLI responds. If the employer sues the AI vendor for a defective tool, Tech E&O responds. If applicants sue the AI vendor directly, as in the Workday class action, the answer is being decided by courts now. Both lines can be on the hook for the same incident through different defendants

Usually not. Cyber covers breaches and privacy events. A finance lead voluntarily wiring money to a fraudster (even one impersonating the CEO on a synthetic video call) is a social engineering loss. The right cover is a commercial crime policy with a social engineering endorsement. A few cyber carriers, Coalition for one, have expanded funds-transfer-fraud triggers to include deepfakes, but crime is the cleaner answer.

Sometimes, and the cover is narrow. IP infringement insurance is the right policy, and a handful of carriers will write it for AI companies that can document data provenance and licensing. Tech E&O usually excludes IP. Media liability covers some output-side claims but not the broader training-data question. If you train models on third-party content, treat IP coverage as a required line, not optional.

Any AI capability claim that overstates what the system actually does, what data trained it, or what results it has produced. The pattern the SEC enforces: a public statement, an investor who relied on it, a gap between the claim and the reality. The fix is bounded, specific language. “Our AI reduces processing time by 12% in controlled testing” defends. “We use AI to deliver superior results” invites a subpoena.

With AI governance documentation ready, a specialist broker can usually place a starter program in days. Without it (no bias testing, no governance policy, no model evaluation records), expect a longer cycle, more carrier questions, and higher pricing. The documentation is the gating item, not the placement.

If the AI is embedded in a physical product and a person is hurt, product liability is the first responder, with GL behind it. If the harm is mental or emotional, as in the Character.AI suits, the answer is split across Tech E&O, media liability, and product liability, depending on how the claim is framed. Consumer-AI companies should expect higher premiums and tighter underwriting questions across all three lines in 2026.

Build your AI program around how claims actually pay

Need the buy-side view (what to carry at each stage, what limits enterprise customers demand, what each line costs)? See our companion guide, Insurance for AI Companies.

AI risk cuts across Tech E&O, D&O, cyber, media, IP, and crime. Most companies find out at claim time that their existing policies have AI exclusions, narrow definitions, or sublimits that do not match their exposure. Alliance Risk reviews your full stack and markets your risk to carriers that specialize in AI and tech liability, with proposals back in a few business days

Send us your current Tech E&O, D&O, cyber, and media liability declarations. We will flag AI exclusions, silent coverage gaps, training-data carve-outs, and shadow-AI vulnerabilities at no cost.

Share your stage, revenue, AI use cases, and the contract or board requirement driving the timeline. We will come back with options and pricing within 24 hours.

Talk to an Alliance Risk advisor about your AI exposure today.