Legal exposure
Who's responsible when the AI gets it wrong

The short answer
Usually, you are. Courts so far treat what your AI tells people as what your company told them. And the AI companies' contracts put checking the answers on you: they cover copyright claims, not wrong answers.
Manage it as you would any advice your company gives: limit what it may answer, show where each answer came from, and keep a person on decisions that matter.
Think of a new hire on your front desk
If they tell a customer the wrong refund rule, the customer can hold you to it. "That was only the receptionist" won't get you far. You'd train them, give them the rule book, and tell them when to fetch a manager.
Your AI speaks for your company, just like your staff.
What courts have said so far
In 2022, Air Canada's website chatbot told a customer, Mr. Moffatt, that he could claim a bereavement fare refund after travelling. The airline's own policy page said he couldn't. When he claimed, Air Canada argued it wasn't responsible for its chatbot. A Canadian tribunal disagreed in February 2024:
"It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot."
The airline paid about C$812. The sum was small; the principle wasn't (Moffatt v. Air Canada, 2024 BCCRT 149).
Professionals are held to the same standard. In 2023 a New York court fined two lawyers and their firm $5,000 for filing a brief with cases ChatGPT had invented (Mata v. Avianca). A researcher tracking such cases worldwide had counted more than 2,300 by November 2026, over 900 of them involving lawyers (AI Hallucination Cases database). In 2025, Deloitte refunded the Australian government about A$98,000 of a A$440,000 contract after a report it wrote turned out to cite sources that didn't exist (Information Age, October 2025).
The law is moving the same way. Since January 2026, California doesn't let a company that used AI argue that the AI "autonomously caused the harm" (California Civil Code 1714.46). And from December 2026, the EU's new product liability rules treat software, AI included, as a product, so its makers can be liable for defects without anyone proving fault (Directive (EU) 2024/2853). That covers injury, damaged property and lost personal data, not business losses.
What the AI companies' contracts cover
The big AI companies all promise to defend you if someone claims an answer copied their work. None promises the answers are right.
Microsoft
"Customer is solely responsible for evaluating the accuracy, completeness, and suitability of Output Content."
Copyright: covered. Accuracy: yoursOpenAI
"Customer is solely responsible for all use of the Outputs and for evaluating the accuracy and appropriateness of Output."
Copyright: covered. Accuracy: yoursGoogle Cloud
Its AI services "may provide inaccurate or offensive Generated Output, and are not designed for or intended to meet Customer's regulatory, legal, or other obligations."
Copyright: covered. Accuracy: yoursAnthropic
Customers must tell their users that "factual assertions in Outputs should not be relied upon without independently checking their accuracy."
Copyright: covered. Accuracy: yours
Insurance is shifting too. Some large US insurers have asked regulators to let them exclude AI from business policies, and standard exclusion wording became available in January 2026 (TechCrunch, November 2025; Independent Agent, October 2025). Specialist AI cover now exists, at Lloyd's among others. Ask your broker what your policies say.
How much to let it say on its own
Answers for staff, who check
Employees use it to find things, see the source, and decide what to do.
Best for: a first project. A person stands between the answer and anyone outside.
Gives up: some speed, and it only works if staff really do check.
Answers for customers, kept to what's written
It answers customers only from approved pages, quotes them, and hands over when it can't.
Best for: common questions with clear written answers, such as opening hours, delivery or returns.
Gives up: the harder questions, and it needs your public pages to be right and consistent.
Answers that decide something
It approves a refund, sets a price, or advises on a claim, with a person signing off.
Best for: high volumes of routine decisions, with legal advice before launch.
Gives up: the safety net. Every wrong answer is a commitment, so it needs testing, review and records.
What goes wrong
- It fills a gap with a confident guess. No page covered the question, so it made up a policy.
- Your own pages disagree. An old policy and a new one, both in the collection, and it picks the wrong one.
- Relying on a disclaimer. Air Canada's policy page was right, and it still lost.
- Nobody checks what goes out under your name. A report or a letter cites sources that don't exist.
- Assuming the vendor carries the risk. Their contract covers copyright claims. Wrong answers are yours.
Questions to ask your lawyer, broker and vendor
- For your lawyer: will customers, applicants or the public ever see its answers directly?
- For your lawyer: what could we be held to if it gets a policy, price or promise wrong?
- For your vendor: does every answer show its source, and what does it do when it can't find one?
- For your vendor: what exactly does your contract cover, and up to what amount?
- For your broker: do our policies exclude losses caused by AI, and should we add cover?
Other names you may hear
- Made-up answer, hallucination
- A confident answer that isn't true or isn't in the documents.
- Liability
- Being legally responsible for harm, and having to pay for it.
- Negligent misrepresentation
- Carelessly telling someone something untrue that they rely on. The basis of the Air Canada decision.
- Indemnity, indemnification
- A promise in a contract to cover certain claims against you.
- Copyright commitment
- Microsoft's name for its promise to defend customers against copyright claims over AI output.
- Product liability, strict liability
- Responsibility for harm from a defective product, without anyone having to prove carelessness.
- Limitation of liability, liability cap
- The most a vendor's contract says it will ever pay you.
- Human in the loop
- A person reviews the AI's answer before it is acted on.