When an AI Tool Harms You, Who’s Actually Liable?

Picture the ordinary version of this. You ask an AI assistant a question that actually matters — about a refund policy, a dose, a tax rule, a contract clause — and it answers with the fluent confidence these tools always have. You act on it. It was wrong. Now you are out of pocket, or worse. Whose problem is that?

The answer-first version is unsatisfying but honest: it depends on where you live and who you sue, and almost every party involved has arranged things so the answer isn’t “us.” There is, in most places, no single law that says “here is who pays when an AI gets it wrong.” Instead there is a scramble to fit a new kind of tool into old boxes — contract, consumer protection, negligence, product liability — while the companies that build these systems write terms designed to keep the box firmly shut.

That is worth understanding before you need it, because the gap between how these products are marketed — capable, authoritative, ready for real work — and how their contracts describe them — experimental, unwarranted, use at your own risk — is where liability quietly lives.

“The chatbot did it” is not a defence

Start with the case everyone cites, because it is refreshingly concrete. In Moffatt v. Air Canada, decided by British Columbia’s Civil Resolution Tribunal in early 2024, a grieving customer asked the airline’s website chatbot about bereavement fares. The bot told him he could book now and claim the discount retroactively within 90 days. That was wrong; the airline’s actual policy, sitting on a different page, said no such thing. When he tried to claim, Air Canada refused.

The airline’s defence became briefly famous: it argued, in effect, that the chatbot was “a separate legal entity that is responsible for its own actions.” The tribunal did not buy it. Air Canada was responsible for all the information on its website, it held, whether that information came from a static page or a chatbot, and it owed the customer a duty to take reasonable care that its representations were accurate. The company was ordered to honour the fare and pay damages. The principle is simple and portable: if your business puts an AI in front of customers, what the AI says is what your business said.

When the product itself is the harm

Moffatt is a small-money consumer case. The harder frontier is product liability, and here the landmark is Garcia v. Character Technologies in the United States. The suit was filed in October 2024 by a mother after the death of her fourteen-year-old son, and it named the chatbot maker, its founders and Google. We report it soberly and only for what the court actually did, because the underlying facts are a tragedy, not a talking point.

What the court did was significant. In an early 2025 ruling, a federal judge in the Middle District of Florida denied the companies’ motion to dismiss, rejected a First Amendment argument that a chatbot’s outputs were protected speech, and — crucially — treated the app as a “product” for the purposes of product-liability law, allowing wrongful-death, negligence and product-liability claims to proceed. That was a decision about whether the case could go forward, not a final verdict that the company was liable, and the matter was later resolved through a settlement. But the doctrinal door it nudged open is the one that matters for everyone else: if a court is willing to call an AI system a product, then the whole apparatus of product liability — defective design, failure to warn — comes with it.

The terms you clicked past

Now the part almost nobody reads, which is precisely where the companies do their work. Open the terms of service for a major AI provider and you will find a familiar wall. The service is provided “as is.” Warranties — including the implied ones that the thing is fit for its purpose or of satisfactory quality — are disclaimed to the maximum the law allows. Liability for indirect, incidental or consequential damages is excluded outright. And total liability is capped at a number that is, from the user’s side, close to symbolic.

Anthropic’s consumer terms, for instance, cap its total liability at the greater of the amount you paid in the six months before the claim and $100, while excluding indirect and consequential damages. OpenAI’s provide the service “as is,” disclaim warranties and exclude the same broad categories of damages. This is not unusual language for software, and that is the point: the industry has borrowed the liability posture of a free web app and applied it to tools it simultaneously markets as good enough to replace human workers and to advise you on things that matter.

Whether those clauses actually hold is a separate question, and a jurisdictional one. In many consumer-protection regimes — the UK and EU among them — a business cannot simply contract its way out of liability for its own negligence or for failing to provide a service with reasonable care, and terms that try can be struck down as unfair. A liability cap that looks ironclad in the document can be a good deal softer in front of a consumer tribunal. But you should assume the company will lead with the cap, and that testing it costs time, money and nerve most people do not have.

There is a second layer of friction most users never see until they need to: the same terms typically route disputes into individual arbitration and waive the right to join a class action. That combination is doing quiet work. It means a harm that is small for any one person — a wrong answer that cost you an afternoon, a modest sum, a missed deadline — can rarely be pooled into the kind of case that is worth a lawyer’s time. The economics are designed so that most grievances are individually too small to pursue and collectively too fragmented to assemble. The cap tells you what you could recover; the arbitration clause quietly makes sure you probably won’t try.

The harms that never reach a courtroom

Moffatt and Garcia are the cases that made headlines, but they are the exceptions that prove the rule: most AI harm is too ordinary and too small to litigate, which is exactly why the terms are written the way they are. Think of the everyday versions. A coding agent, run unattended, deletes or mangles work you can’t easily reconstruct. A hallucinated citation or fact makes it into something you file or send, and the embarrassment — or the professional sanction — is yours, not the model’s. An AI summary of a policy is confidently wrong and you act on it. A chatbot says something defamatory about a named person, and the question of who published it is genuinely unsettled.

None of these arrives with a clean defendant and a big number attached, so almost none of them ever becomes a case. They are absorbed, quietly, by the person who used the tool — which is the same pattern we keep documenting, whether the subject is who owns what an AI helps you make or what the rules actually protect you from. The liability isn’t so much assigned as it is left where it falls, which happens to be on you.

Europe pulled up a ladder it had lowered

For a moment it looked as if Europe would write the missing rule. The proposed AI Liability Directive was meant to do something specifically useful for ordinary claimants: ease the burden of proving that an opaque AI system caused their harm, because “the model did something I can’t inspect” is a miserable thing to have to prove. In 2025 the European Commission withdrew the proposal, citing a lack of agreement and a broader push to simplify digital rules. The withdrawal was formalised later that year.

What is left is not nothing, but it is a patchwork. AI harm in the EU now falls to national civil law — which means outcomes can differ from one member state to the next — plus the revised Product Liability Directive, which does one genuinely important thing: it explicitly treats software, including AI systems, as a “product,” and applies strict liability, with its rules taking effect from 9 December 2026. Strict liability matters because it means a claimant need not prove the maker was careless, only that the product was defective and caused harm. It is the most consequential pro-consumer development in this area, and it arrives quietly, on a date most users will never notice.

What actually decides whether you can recover

Pull the threads together and a rough checklist emerges. Whether you have a real claim, rather than a grievance, tends to turn on:

• Was there a relationship the law recognises? A paying customer of a business that used AI to serve you (as in Moffatt) is on far firmer ground than someone who got a bad answer from a free chatbot they used casually.
• Does consumer-protection law apply? In many countries it overrides the fine print, so a “we’re not liable for anything” clause may not survive contact with a tribunal.
• Can the AI be framed as a “product”? If so, product-liability rules — and, in the EU from December 2026, strict liability — may attach, sidestepping the need to prove fault.
• Did the company make specific promises? A concrete representation about accuracy or safety is easier to hold them to than a vague marketing vibe.
• Where are you, and where are they? Jurisdiction, arbitration clauses and which country’s consumer law applies can decide the case before the merits are even reached.

The fair case for the other side

Steel-man the companies, because their position is not pure evasion. Large language models are probabilistic; they will sometimes be confidently wrong no matter how much work goes in, and that unreliability is not fully fixable today. Disclaimers and caps are standard across software precisely because a tool used in a million unforeseeable ways cannot underwrite every outcome. And there is a real risk that clumsy, over-broad strict liability could chill genuinely useful products or push them out of smaller markets. Regulators withdrawing a directive because they could not make it work is, at least, more honest than passing something incoherent.

All true. And all beside the narrower point, which is about symmetry. A company cannot spend one budget telling you a tool is authoritative enough to lean on and another budget telling a court it is an experimental toy you use entirely at your own risk. When the marketing and the terms disagree this sharply, the person left holding the difference should not be the user by default.

What to do while the law catches up

Practical, boring, effective. For anything that carries real stakes — money, health, legal exposure — treat an AI answer as a lead to verify against a primary source, not as advice you can bank. Keep records: the prompt, the answer, the date, a screenshot. If a business’s own chatbot gives you a commitment, that commitment may well bind the business, as Moffatt shows — so save it. Know that in many places consumer law is on your side more than the terms suggest, and that a confident “we’re not liable” is an opening position, not a verdict. And watch the direction of travel: for all the disclaimers, courts and legislators are slowly deciding that when these systems cause harm, “the AI did it” is not the end of the conversation. It is the start of one. The gap between the two answers — the confident tool in the advert and the unwarranted experiment in the contract — is closing, slowly, and mostly not because the companies chose to close it. It is closing because a tribunal in Vancouver, a judge in Florida and a directive in Brussels each declined to accept that the most convenient reading of who pays is also the correct one.