Courts in several countries, and some clients, have started sending the bill for AI mistakes to the person who passed them on. The pattern shows up in very different places, and it says something practical about how to use AI at work.
The same rule shows up in four different cases
In the US Tax Court, a lawyer filed a brief citing four cases to support his client's argument. Three of them did not exist, and the judge said they looked like AI inventions. The lawyer was warned and not fined. His client's argument lost anyway, which means the person who first paid for the mistake was the one who never touched the AI.
Deloitte delivered a long report to the Australian government with invented academic sources and a made-up quotation from a court judgment. It later refunded about A$97,000 of a contract worth about A$440,000, less than a quarter of the price. The government did not take the errors up with the AI company. It took them up with the firm that put its name on the report.
The Air Canada ruling follows the same line. The airline's chatbot gave a wrong answer, the airline argued the bot was responsible for itself, and the tribunal said a bot on your website speaks for you.
The fourth case shows the rule works in both directions. When Google's own AI summaries allegedly said false things about a TV producer, a US court let some of his claims go forward in September 2026, and has not yet decided whether Google is liable. A maker that publishes AI's words itself, under its own name, is treated like anyone else who publishes. The line that matters is not who built the model. It is who put the words in front of the other person.
Checking is the part AI cannot do for you
A human mistake usually looks like a mistake: a typo, a gap, a vague sentence. An AI mistake arrives finished, with a confident explanation and references that look real. That is why a quick read does not catch it, and why checking has to mean opening the source and confirming it says what the draft claims.
That costs less than most people expect. Take a consultant who uses AI to draft a 30-page report and gets 40 references back. Opening each one takes about three minutes, so the whole check is about two hours. Drafting that report by hand would have taken a day or two longer. AI still saves most of the time, and the two hours are what separates a report that goes out with invented sources from one that does not.
So the saving is real but smaller than it looks, and the work that remains is the work that carries the responsibility. Reading what AI wrote as if you had to defend every line to the person receiving it is now part of the job, and it is the part that cannot be handed on.
A customer-facing bot needs a written limit
The chatbot case is the riskiest, because nobody reads the answer before the customer sees it. Picture a shop owner with a chatbot on the website that answers questions about returns, delivery times and discounts. If it tells a customer the return window is 60 days when the policy page says 30, the Air Canada ruling suggests the customer can hold the shop to 60. A correct policy page elsewhere on the site does not protect the shop, because the tribunal did not expect customers to cross-check.
The fix is small and can be done in an afternoon. Decide which questions the bot may answer, and have it answer only from the policy page. Send anything involving a refund, an exception or a promise of money to a person. Then ask it the twenty questions customers ask most and compare each answer with the policy. Until that is done, the bot is making promises in the shop's name that nobody has read.