On 11 November 2022 the automated assistant on Air Canada’s website told Jake Moffatt he could claim a bereavement fare within 90 days of ticket issue. He couldn’t, and the official page linked inside that same answer said the opposite. British Columbia’s Civil Resolution Tribunal awarded CAD 812.02 for negligent misrepresentation, rejecting the airline’s suggestion that the bot answers for itself.
On 11 November 2022, the day his grandmother died, Jake Moffatt opened Air Canada’s website and typed a question into a chat window. What answered him was an automated conversational assistant: software that parses a question written in plain language and returns an answer drawn from the airline’s commercial information — fare policies, refund procedures, travel rules. It doesn’t sell tickets and it doesn’t open case files. It produces text, and the customer reads that text as if an employee had said it. That is precisely the point on which, fifteen months later, the case turned.
The answer Moffatt got concerned bereavement fares, the reduced price Air Canada applies to passengers travelling because of a death in the family. The bot told him that someone who has already travelled can request the reduction within 90 days of the date the ticket was issued, by filling in the refund form. In that same answer, the words “bereavement fares” were a hyperlink to the airline’s official policy page. That page said the opposite: the policy does not apply to requests submitted after travel. Two parts of the same website, two incompatible statements. How sophisticated the artificial intelligence behind the chat window actually was is a separate question — we’ll come back to it, and the answer is less obvious than the usual retelling — because the problem the tribunal had to solve is older than machine learning.
What happened, in order
Moffatt bought the flights: Vancouver–Toronto for CAD 794.98, Toronto–Vancouver for CAD 845.38. Total: CAD 1,630.36. On 17 November 2022 he submitted his first partial refund request. Months of correspondence followed, until on 8 February 2023 an Air Canada employee put in writing that the chatbot had used “misleading words”. A striking admission — which did not translate into a refund. The airline offered a CAD 200 coupon as a goodwill gesture. Moffatt declined it and took the case to British Columbia’s Civil Resolution Tribunal.
It’s worth explaining what that body is, because international coverage almost always called it a “court” without qualification. The CRT is a Canadian administrative tribunal that operates largely online, with jurisdiction over small claims up to CAD 5,000. In this category a party may file a notice of objection within 28 days: the decision then ceases to bind, and the matter can be relitigated in Provincial Court. It is not a judgment of an ordinary court, and it is not binding precedent in the sense common lawyers mean.
The decision — Moffatt v. Air Canada, 2024 BCCRT 149, file SC-2023-005609 — came on 14 February 2024, signed by tribunal member Christopher C. Rivers. The outcome: CAD 650.88 in damages, CAD 36.14 in interest, CAD 125 in fees. Total CAD 812.02, payable within fourteen days. The CAD 200 coupon was not deducted, because Moffatt had never accepted it.
The line that travelled around the world
In its defence, Air Canada argued it could not be held responsible for information supplied by its “agents”, the chatbot included. Rivers summarised the airline’s position as suggesting that the chatbot is “a separate legal entity that is responsible for its own actions”, and called it a remarkable submission. The reasoning that follows is disarmingly plain:
The chatbot is still just one part of Air Canada’s website. It makes no difference whether the information comes from a static page or from the bot: a customer cannot be expected to check one part of a website against another.
The legal basis was not any AI-specific statute — Canada had none applicable — but negligent misrepresentation, with the five requirements set out by the Supreme Court of Canada in Queen v. Cognos Inc. (1993). The measure of damages followed Ban v. Keleher (2017). The CAD 380 per-leg reference fare behind the award did not come from the bot: Moffatt had been quoted it by a telephone agent, and the tribunal adopted it as market value by adverse inference, since Air Canada produced no contrary evidence. The airline also raised a contractual defence based on its Domestic Tariff, but never filed the tariff text. The defence collapsed.
Here is the first failure, and it is a legal one before it is a technological one. A party that turns up without the contract its defence rests on, and without evidence of market price, lost for reasons that have nothing to do with software.
Three things the case does not say, and that get said anyway
- The bot did not invent a price. The bereavement discount was real. What the system invented was a procedure: the possibility of claiming it retroactively. That distinction matters, because an error about the conditions of access to a service is far harder to catch with automated checks than an error about a number.
- Nobody ordered Air Canada to honour what the bot promised. The CRT awarded differential damages for misrepresentation, which is a different thing: it did not convert the chatbot’s answer into a contractual obligation.
- We don’t know whether it was a generative model. The decision states expressly that Air Canada provided no information about the nature of its chatbot. An airline spokesperson later told the legal press that the tool had been developed before generative AI capabilities arrived. The word “hallucination”, which appears in nearly every retelling, is a journalistic label, not a finding on the record.
That last point is the interesting one, and it inverts the moral almost everyone drew. If the system was not a generative language model, then the fault isn’t the statistical unpredictability of newer models: it’s a content error in an automated answering system, something companies were perfectly capable of shipping well before 2022. Moffatt does not prove that generative artificial intelligence is dangerous in customer service. It proves that a company can put a system online that talks to customers without any process checking whether what it says matches its own published policies. The link to the correct page, sitting inside the wrong answer, is the proof that both contents lived in the same perimeter and never spoke to each other.
Why it gets cited, and how
The ruling drew attention out of all proportion to its legal weight: the Washington Post covered it on 18 February 2024, the BBC on 23 February. It has since become the near-obligatory reference point in the literature on liability for conversational system outputs — it appears, among other places, in a case comment published in the UBC Law Review in 2025. Other versions circulate, from the chatbot’s supposed removal from the website to the project’s alleged costs; verifiable sources do not support them.
The reason CAD 812 weighs more than it’s worth is that the argument Rivers rejected was not some distracted lawyer’s improvisation. It was, stated out loud, the question every company deploying an automated assistant carries around implicitly: if the system gets it wrong, who answers? The CRT’s reply is that the question is badly framed, because it assumes a separation that does not exist. A conversational model is not an interlocutor; it is a channel, like an HTML page or a call centre. Whoever switches it on is publishing, and answers for what they publish.
One operational consequence rarely makes it into the debate. A static page, when policy changes, gets rewritten once. An automated answering system generates an indefinite number of different phrasings of the same policy, any one of which can drift from the original at a single point — a “within 90 days” inserted where none existed. The cost of that difference is not the cost of the model: it’s the cost of the editorial oversight needed to keep it aligned, and in no customer-service automation budget does it appear as a separate line item.