Air Canada's chatbot invented a refund policy, and the airline had to honor it

Booking a funeral flight through a chatbot

On November 11, 2022, the day their grandmother died in Ontario, Jake Moffatt went to Air Canada's website to book a flight from Vancouver to Toronto.¹ Moffatt was looking for a bereavement fare, the reduced price airlines offer people who need to travel because of a death in the immediate family, and used the site's chatbot to ask how it worked. A chatbot, in the tribunal's own later words, is "an automated system that provides information to a person using a website in response to that person's prompts and input."¹ Both sides in the case that followed agreed on one thing at the outset: Moffatt was not chatting with an Air Canada employee.¹

The bot's 90 days, against a policy that grants nothing after travel

The chatbot told Moffatt that bereavement fares could be claimed backward. If you "need to travel immediately or have already travelled," it wrote, you could submit a ticket for a reduced bereavement rate "within 90 days of the date your ticket was issued" by filling out a refund form.¹ Read plainly, that is a promise: book now at full price, apply later, get the difference back.

Air Canada's actual policy was the opposite. Its bereavement rule "does not apply to requests for bereavement consideration after travel has been completed," and the airline's own page still tells customers to request the fare by phone before they fly, not after.¹² The airline's own page said as much in its own words in the weeks around Moffatt's booking. As it read then, the policy applied to "Bereavement travel that occurs within 10 days of booking and which does not exceed 60 days," and the way to get the fare was the first instruction on the page: "Please contact us by phone to request a bereavement fare." It was an arrangement made before flying, and nowhere on the page was there a retroactive claim or an after-the-fact refund path.³ There is no 90-day window. There is no retroactive claim. The bot had invented a procedure that did not exist.

The two answers sat on the same website at the same time. In the bot's own message, the words "bereavement fares" were a live hyperlink to Air Canada's real bereavement page, the page that ruled the refund out.¹⁴ The bot contradicted the policy it was pointing at, in the same breath.

Moffatt believed the bot. They booked two flights, Vancouver to Toronto and back, paying $1,630.36 in total.¹ Later they applied for the bereavement refund the bot had described. Air Canada refused, explained that bereavement fares cannot be claimed once travel is done, and offered a $200 coupon.¹ A bereavement fare would have cost about $380 each way. With taxes, the two flights should have come to about $979.48, not $1,630.36.¹ Moffatt filed a claim with British Columbia's Civil Resolution Tribunal.

Who answers for what a chatbot says

Air Canada's defence is the part of the case people remember. The airline argued it "cannot be held liable for information provided by one of its agents, servants, or representatives, including a chatbot." The tribunal member, Christopher Rivers, restated the position: "In effect, Air Canada suggests the chatbot is a separate legal entity that is responsible for its own actions."¹ His answer to it was one paragraph: "This is a remarkable submission. While a chatbot has an interactive component, it is still just a part of Air Canada's website. 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."¹

The ruling did not need a new law about artificial intelligence. It turned on negligent misrepresentation, an ordinary rule that one party is liable when it carelessly gives another false information that the second reasonably relies on to its cost. The tribunal applied the test set out by the Supreme Court of Canada in 1993 in Queen v. Cognos: the company owed a duty of care, made an untrue or misleading representation, made it without reasonable care, the customer relied on it reasonably, and the reliance caused a loss.¹⁵ A duty of care is just the legal name for an obligation to take reasonable care not to mislead the person you are dealing with. Air Canada had one to a paying customer, the tribunal found, and it did not take reasonable care to make sure its chatbot was accurate. The airline never explained why the page titled "Bereavement travel" should count for more than the chatbot, or why a customer should have to check one part of the site against another.¹

The award was small and precise. The tribunal ordered $650.88 in damages, the difference between the $1,630.36 Moffatt paid and the roughly $979.48 the bereavement fares would have cost.¹⁶ Adding $36.14 in interest and $125 in fees brought the total order to $812.02.¹⁶ The two figures are worth keeping apart, because some coverage rolled them into one: $650.88 is the fare difference, and $812.02 is the whole order.⁷ Moffatt had asked for about $880; the tribunal, working from tax-inclusive fares, awarded the lower amount.¹

It is tempting to read the decision as a turning point. One passenger advocate called it "a landmark one that potentially sets a precedent."⁸ The record argues for the smaller claim. The tribunal reached its result with a rule that was three decades old, and the decision itself is a non-binding small-claims ruling that does not bind other tribunals or judges.⁴ What was new was the object the old rule met, not the rule. Air Canada, for its part, told Moffatt in February 2023 that it had noted the issue so it could update the chatbot.¹

Fluent text that was never checked against the policy

The bot did not misread the policy or serve an out-of-date version of it. It produced a smooth, confident paragraph of policy that had never been true. That is what people mean by a hallucination in these systems: not a machine seeing things, but text generated to read plausibly, with nothing checking it against the facts it asserts.⁹ The word flatters the machine, because it suggests a mind that slipped. What actually happened is more ordinary. Why the words came out that way, a model choosing likely next words rather than retrieving a fact, is the subject of a separate lesson on how a model samples its output. This case is about what follows once the text exists.

The hyperlink is the tell. A retrieval bot bolted onto a help site can generate text that sounds like policy without being held to the policy it links to. The correct page was one click away, inside the bot's own sentence, and the sentence still contradicted it.¹ Sounding right and being grounded in the linked source are two different things, and nothing in the setup forced the second.⁸ That is the mechanism worth carrying away, and it is separate from the ruling. What the tribunal found is a fact about liability. Why the bot spoke false policy is a fact about how the system generates text. The two meet in this case, but they are not the same fact.

The same setup is answering support questions right now

Nothing about the failure was specific to airlines or to bereavement. The setup was a generative bot speaking for a company on its own site, in natural language, not constrained to the company's real rules. That description fits the support chatbots now fielding questions about refunds, overdraft fees, return windows, and warranty terms across banking, retail, and travel. Each one can produce a fluent answer that is not checked against the policy it claims to state. The problem is not exotic even within travel: a WestJet chatbot in 2018 once replied to a passenger with a link to a suicide-prevention hotline for no evident reason.⁸

Since Moffatt, a steady run of "our AI assistant said X" disputes has followed the same shape: a customer acts on what a bot told them, and the company argues the bot was wrong and the customer should have known better. The tribunal's answer to that argument is the durable part of the case. A company owns what its chatbot tells a customer, exactly as it owns a printed page, and "the AI said it, not us" is not a defence. The next time a support bot answers a real question with a confident number or a firm rule, that is the setup, working as designed, one unchecked sentence away from the one that cost Air Canada $812.02.¹