Who Is Liable When the AI Is Wrong
Air Canada's customer service chatbot told a grieving passenger he could apply for a bereavement discount after booking. No such policy existed. When the passenger asked the airline to honor what its own bot had promised, Air Canada argued, more or less, that the chatbot was a separate entity responsible for its own words. A Canadian tribunal rejected that and held the airline to the promise.
There is a comforting idea going around about business AI, which is that when the tool gets something wrong, the vendor answers for it. You didn't build the software. You can't see inside it. When it invents a fact, surely that lands on someone else. The Air Canada ruling is one of a growing line of decisions saying otherwise, and where they point should shape how you deploy anything that talks to a customer or produces work you rely on.
The mistake is treated as yours
Read the tribunal's reasoning and the lesson is plain: a tool speaking on your behalf is you speaking, and you own what it says.
The pattern keeps surfacing. New York City rolled out a chatbot to answer questions about local business rules, and reporting caught it telling owners to do things that broke the law. A large consulting firm refunded part of a government contract, reported at around 300,000 dollars, after a delivered report turned out to rest on citations the AI had fabricated. In the courts it has become almost routine: lawyers file briefs that cite cases which do not exist, invented whole by a chatbot, down to convincing fake docket numbers. The first big one, Mata v. Avianca, drew a 5,000 dollar sanction and a wave of headlines. It didn't stop the next ones. Lawyers, of all people, apparently don't read the footnotes either.
"The AI made it up" hasn't worked as a defense. Courts keep reaching for ordinary agency law, the same rules that make you answer for what an employee tells a customer. The machine doesn't change the logic.
Why the errors are so convincing
The failure has a name, the hallucination, and most people picture it wrong. They imagine a glitch: garbled text, obviously broken, easy to spot. The reality runs the other way. A hallucination is fluent, confident, richly detailed, and completely false. The fabricated court cases came with real-looking names, dates, and numbers. The invented citations looked exactly like genuine ones.
Owners tell me their staff would obviously catch a bad answer, and they're right about the answers that look bad. Your people will flag output that looks wrong. They'll sail straight past output that looks right and happens to be false, which is what these tools produce when they fail. Confidence tells you nothing about accuracy. The thing is equally smooth when it knows and when it's inventing, and it will never tell you which one you just got.
Where to keep a human, and where you can relax
Don't swear off the tools. Sort the work by what a mistake costs, and spend your caution there.
Some work is cheap to get wrong. A first draft of an internal memo. A pile of ideas you were going to filter anyway. A rough summary you'll read with a critical eye. If it's off, you notice and move on. Use the tools freely and skip the ceremony.
Other work is expensive to get wrong. Anything a customer receives as fact. Anything with a number that drives a decision. Anything cited as a rule or a source. Anything that goes out under your name and can't be pulled back. There the standard is simple: a competent person reads it before it leaves and owns it as if they'd written every word. The AI drafted it. The person is responsible for it. Say that part out loud, because the whole trap is that smooth output invites the reader to skip the check.
Be hardest on a customer-facing chatbot answering questions with nobody watching. That's the exact setup that produced the Air Canada bill. Box it into information you've verified, and teach it to say it doesn't know rather than improvise. An improvised answer is a promise you never made and may still have to keep.
You don't need fear, and you can't afford blind trust. What you need is a list of every place a confident, well-formed, false statement would cost you real money, and a person posted at each one whose job is to catch it. The tool won't raise its hand when it's guessing. That job belongs to a person, and you should be able to name who it is.