Who pays when AI causes harm? The quiet legal battle that will shape the industry more than any model race

THE THESIS. The battle that will define the AI industry is not being fought on benchmarks but in courtrooms. While the spotlight follows the model race, courts, lawmakers and insurers are answering the question that actually organizes a market: who pays when this fails? Our thesis: the answer is being written right now, and it points in one clear direction. AI is being treated as a product that can be defective — not as a megaphone protected by free speech, and not as an autonomous entity you can blame. Counterintuitive as it sounds to parts of the industry, that is good news in the long run. No transformative industry — aviation, automobiles, pharmaceuticals — truly took off until it was clear who answered for its harms. This is not ‘AI on trial’. It is the oldest question in tort law applied to the newest technology.

THE HARM IS NO LONGER HYPOTHETICAL. In June 2026, police in Bentonville, Arkansas arrested a local photographer on 200 counts of creating AI-generated child sexual abuse material. According to investigators, he used Grok, xAI’s model, to transform real photos of minors who were clients of his studio. Since July, at least three families — including those of a 10-year-old and a 6-year-old — have sued xAI in federal court. The lawsuits allege — and we stress these are unproven accusations — that xAI skipped safeguards that are standard across the industry and marketed Grok as ‘largely uncensored’; by January 2026, attorneys general from 35 states had already demanded changes in a joint letter. The harm is not anecdotal on the fraud front either: the FBI’s 2025 internet crime report logged more than 22,000 AI-related fraud complaints with losses above $893 million, of which $352 million hit people over 60. A convincing voice clone takes about three seconds of audio, and industry analyses report voice-scam growth above 1,000% in 2025. Add the third frontier: autonomous agents — AI programs that act on your behalf — already browsing, buying and signing without direct human supervision.

THE SECTION 230 HOLE. The entire American legal debate turns on a 1996 law. Section 230 shields platforms from what their users post: if someone defames on a social network, the user is liable, not the network. But that immunity covers hosting third-party content, not creating it. The statute itself excludes anyone ‘responsible, in whole or in part, for the creation’ of the content — and generating is precisely what a generative model does. The emerging legal consensus, as the Center for Democracy and Technology’s analysis summarizes, is that a model’s output likely falls outside the shield. That opens the central question: if a model is not a megaphone, what is it? A product, like a car with faulty brakes, or a service, like an advisor who gets it wrong? The distinction matters enormously, because products carry strict liability: the victim does not need to prove negligence, only that there was a defect, that they were harmed, and that the defect caused the harm. In May 2025, federal judge Anne Conway gave the first weighty answer in Garcia v. Character Technologies, over the suicide of a 14-year-old who had an intense relationship with a chatbot: the chatbot’s output is not First Amendment-protected ‘speech’, and the system can plausibly be treated as a product subject to liability. The defective-design claims moved forward.

THE PRECEDENTS ALREADY MARK THE PATH. The pattern repeats in every known ruling, and it always points the same way: liability flows to the company that builds or deploys the system, never to ‘the machine’. In 2024, a Canadian tribunal ordered Air Canada to pay CA$812 to a passenger after its chatbot invented a discount policy; the airline argued the chatbot was a ‘separate entity’ responsible for its own actions, and the tribunal flatly rejected it (Moffatt v. Air Canada). In January 2026, according to published reports, Character.AI and Google settled the Garcia case and four similar suits, committing to new safety measures for minors — the market is already pricing in that these cases get lost or settled. Still open is Raine v. OpenAI, the lawsuit over a 16-year-old’s death, which OpenAI denies and attributes to ‘misuse’ of the system. On the agent front, Amazon sued Perplexity in late 2025 because its agentic browser Comet was logging into customer accounts and completing purchases — possibly the first federal lawsuit over a commercial agentic system. California closed the loop with AB 316, in force since January 1, 2026: anyone who develops, modifies or uses an AI cannot argue in their defense that the AI ‘acted autonomously’. The most obvious escape hatch was shut by statute.

EUROPE REGULATES, AMERICA LITIGATES. Here a cliché needs dismantling: it is not ‘Europe regulates while America lets it ride’. They are two different methods arriving at a similar place. The EU withdrew its AI Liability Directive in February 2025, yielding to simplification pressure. But the decisive piece survived in a rule almost nobody watches: the new Product Liability Directive (2024/2853), which member states must transpose by December 9, 2026, explicitly treats software and AI systems as products under strict liability. In Europe, someone harmed by defective AI will not have to prove the maker was negligent. The United States, with no federal framework, is building the equivalent by accretion: rulings like Garcia, state laws like AB 316, and juries with the power to award massive damages. Our read: the European system provides certainty up front, while the American one discovers the rules case by case — with more volatility and more fear in boardrooms. But neither is laissez-faire anymore. The era when ‘AI harms were nobody’s problem’ has ended on both shores.

THE REGULATOR NOBODY IS WATCHING. Meanwhile, the most influential actor stays out of the headlines: insurers. Since January 1, 2026, new industry-standard forms (the Verisk/ISO endorsements — the template clauses used in most American policies) allow carriers to exclude generative AI from general liability insurance, and major insurers such as AIG, Chubb and WR Berkley are already applying them, according to trade press reports. In parallel, a dedicated AI insurance market has been born at Lloyd’s of London: Armilla launched the first standalone policy in 2025 covering hallucinations and model failures, with limits reaching $25 million per organization by January 2026. This is de facto regulation, and a potent kind: to get coverage, you will have to demonstrate evaluations, guardrails and monitoring — just as fire insurance once imposed building codes. The insurer does not debate AI’s future; it prices it today. Our closing read is, as always, measured optimism. In the short term, this hurts: expensive premiums, companies that will not find coverage, projects put on hold, and real victims — children, scammed seniors — who arrived before the rules did. But in the long term, a clear liability regime is exactly the trust infrastructure AI needs to deliver on its promise: curing disease, extending life, generating abundance. The industry complaining about lawsuits today should remember its history: liability did not kill the airplane or the medicine. It made them trustworthy. And only technology that is trusted changes the world.