The Air Canada Chatbot Ruling: What the Tribunal Actually Decided

Moffatt v. Air Canada, 2024 BCCRT 149, is cited as the case that made companies liable for their chatbots. It is a decision of a British Columbia small-claims tribunal awarding CAD 650.88, and what it actually applied was ordinary misrepresentation law with no AI-specific reasoning at all. Both halves of that sentence matter.

The facts the tribunal found

Jake Moffatt’s grandmother died in November 2022. He went to Air Canada’s website and used the support chatbot, which told him he could book a flight at the ordinary fare and apply for a bereavement rate retroactively within ninety days of the ticket being issued. He booked on that basis. Air Canada’s bereavement travel policy page — which the chatbot itself linked to — said the opposite: the reduced fare cannot be claimed after travel has already been completed.

Air Canada refused the refund and offered a policy update instead. The tribunal member, Christopher Rivers, found the chatbot’s statement was inaccurate, that Moffatt had relied on it, and that the reliance was reasonable. The full text is on CanLII — 2024 BCCRT 149, decided 14 February 2024 — and it is short enough to read in ten minutes, which is the right way to check anything written about it, including this page.

The holding, in the tribunal’s own terms

The claim succeeded in negligent misrepresentation. That tort, in Canadian law, requires a duty of care, an untrue or misleading representation, negligence in making it, reasonable reliance, and resulting damage. The tribunal worked through those elements conventionally. Air Canada owed a duty of care to a customer using its own support channel; the representation was inaccurate; the customer relied on it; he paid more as a result.

The award was CAD 650.88 in damages — the difference between the fare paid and the bereavement fare he would have paid — together with pre-judgment interest and the CRT filing fees, bringing the total to a little over CAD 800. That is the entire financial consequence of the case that is described in trade press as a landmark.

The reasoning that gets quoted is the tribunal’s observation that Air Canada did not explain why the page it linked to should be trusted more than the chatbot on the same website. That is a point about a company’s own information architecture, not about machine learning. If a printed brochure and a call-centre agent had contradicted each other, the analysis would have been identical.

The separate-entity argument

The passage that made the case famous is Air Canada’s submission that the chatbot was, in effect, responsible for its own statements. The tribunal rejected it in a sentence, describing the suggestion as remarkable and noting that the chatbot is part of Air Canada’s website; the airline is responsible for all the information on it, however that information is presented.

It is worth being precise about what that rejection is. It is not a ruling on the legal personality of software, and no tribunal needed one to decide this. The airline chose the vendor, deployed the component, and served its output under its own domain. Attribution follows from that, exactly as it would for a badly written FAQ page. The reason the line is quoted so widely is that the argument was unusual, not that the answer was.

What the decision does not establish

This is the part that matters if you are citing it in a memo.

• It is not binding precedent. The Civil Resolution Tribunal is British Columbia’s online small-claims body. Its decisions bind the parties. They do not bind other CRT members, they do not bind the BC courts, and they certainly do not bind anyone outside Canada.
• It decides nothing about generative models specifically. The decision does not record what technology drove the chatbot, and nothing in the reasoning turns on it. A scripted decision-tree bot giving the same wrong answer produces the same result.
• It does not address disclaimers. Air Canada does not appear to have run a serious argument that its terms of service excluded reliance on chatbot output. Whether such a term would work — against a consumer, under consumer-protection law, in a different jurisdiction — is untouched and genuinely unresolved.
• It says nothing about the vendor. Any allocation of risk between Air Canada and whoever supplied the chatbot is a contract question that was not before the tribunal. See AI contract clauses for where that allocation normally lives.

The design consequence

The practical lesson is not “chatbots are risky”. It is that a support assistant which paraphrases policy is generating a representation your organisation will be held to, and paraphrase is precisely what a language model does. An assistant that quotes the policy text and links it, or that hands off rather than improvising when it cannot ground an answer, produces a much smaller surface than one that answers fluently from a vaguely relevant retrieval hit.

The second consequence is evidential. Moffatt had a screenshot. Air Canada, on the face of the decision, did not contest what the bot said. If you cannot reconstruct what your assistant told a specific person on a specific day, you are arguing about your own product from a weaker position than the customer.

For the forward-looking version of the same problem in the EU, the AI Act’s transparency duties require that a person interacting with an AI system be told so — the chatbot disclosure duty covers what that obliges and from when. Disclosure does not, on its own, fix the misrepresentation problem: telling somebody they are talking to a machine does not make the machine’s statement about your refund policy any less yours.