Air Canada is liable for its chatbot

What the chatbot said

In 2022 Jake Moffatt's grandmother died and he went to Air Canada's website to book a flight from Vancouver to Toronto. He asked the airline's chatbot how bereavement fares worked. The chatbot told him to book immediately and then apply for the reduced rate afterwards, within 90 days of the ticket being issued, using the ticket refund application form. He kept a screenshot of the answer. Air Canada's actual policy, on a static page the chatbot also linked to, said the opposite. Bereavement rates cannot be claimed after the flight has been booked.

Moffatt followed the chatbot's instructions and was refused. He spent months arguing with the airline, screenshot in hand. The best offer he received was a promise to update the chatbot and a 200 dollar voucher for a future flight. He turned the voucher down and filed a small claims complaint with the British Columbia Civil Resolution Tribunal.

The argument Air Canada actually made

The part of this case that will be quoted for years is the defence. Air Canada argued that it could not be held liable for information provided by one of its agents, servants or representatives, including a chatbot, and in the tribunal's words suggested that the chatbot was a separate legal entity responsible for its own actions. Tribunal member Christopher Rivers called this a remarkable submission. The airline also argued that because the chatbot's reply linked to the bereavement policy page, Moffatt should have gone there and discovered the correct rule for himself.

Rivers did not accept either line. The decision says Air Canada does not explain why it believes the chatbot is not its responsibility, nor why the page titled Bereavement travel was inherently more trustworthy than the chatbot. Moffatt had no reason to think one part of the website would be accurate and another would not, and the airline does not explain why customers should have to double check information found in one part of its website on another part of its website.

The finding was negligent misrepresentation. Air Canada did not take reasonable care to ensure its chatbot was accurate. The line we expect to see in slide decks is this one: it should be obvious to Air Canada that it is responsible for all the information on its website, and it makes no difference whether the information comes from a static page or a chatbot.

What it cost

The money was small. Moffatt's fare was 1,640.36 Canadian dollars and the tribunal ordered a partial refund of 650.88, plus interest and his tribunal fees, for a total of 812.02 Canadian dollars. Air Canada told Ars Technica it would comply and considered the matter closed. When Ars checked the website after the ruling, the chatbot no longer appeared to be available, and the airline did not confirm whether it had been withdrawn.

The chatbot was not a cheap experiment. Air Canada's chief information officer, Mel Crocker, had told the Globe and Mail the previous year that the initial investment in customer service AI was much higher than the cost of continuing to pay staff to handle simple queries, and that the airline believed automation would lower expenses over time. The stated ambition was to automate every service that did not need a human touch. One bereaved passenger and a screenshot were enough to put that trajectory in front of a tribunal.

Why the decision matters more than the amount

Experts quoted by the Vancouver Sun said this appeared to be the first time a Canadian company had tried to argue it was not liable for what its chatbot said. Gabor Lukacs of the Air Passenger Rights group told the BBC the ruling establishes a common sense principle: if you hand part of your business to AI, you are responsible for what it does, and airlines cannot hide behind chatbots. That principle was already how most consumers assumed things worked. The value of the ruling is that a company tested the alternative and lost.

The technical framing matters here. A chatbot answer is a probabilistic sample from a model conditioned on a prompt. From the product side it feels like a suggestion. From the customer's side, and now from the tribunal's side, it is a statement by the company, with the same weight as a policy page. The gap between those two views is where the liability lives. The same experts told the Vancouver Sun that Air Canada might have escaped if the chatbot had warned users that its answers could be inaccurate, which the airline seemingly had not done.

What we would change in a deployment

The narrow lesson is that a disclaimer might have helped. We think the wider lesson is that the disclaimer is a poor substitute for engineering. If a model is allowed to answer questions about refund policy, the policy text should be retrieved and quoted, and the model should be prevented from paraphrasing eligibility rules it did not retrieve. A 90 day window that does not exist in the policy is exactly the kind of plausible detail that generation produces and retrieval does not.

The second change is about evidence. Moffatt won partly because he kept a screenshot. A company that logs every chatbot exchange can audit what its system promised and settle before a tribunal does. A company that does not log them is arguing from memory against a customer with a picture. As more of this text goes in front of customers, we would want product teams to treat each generated sentence as something they may later have to honour, and to build the system as if that were true.