When an airline's chatbot gave wrong advice, the airline paid

On 11 November 2022 Jake Moffatt's grandmother died in Ontario. The same day, Moffatt went to Air Canada's website to find a flight from Vancouver to Toronto at the bereavement rate, and asked the airline's chatbot. The chatbot said a reduced fare could be claimed after travel. More than a year later a tribunal had to decide who bore the cost of that answer: Moffatt v. Air Canada.

What did the chatbot tell Moffatt, and what happened next?

The chatbot said a bereavement fare could be claimed within 90 days of the ticket being issued, even after travelling, while the airline's own policy page said the opposite. In the screenshot Moffatt supplied, the chatbot tells customers who have already travelled to submit the ticket with a Ticket Refund Application form within 90 days (Civil Resolution Tribunal (British Columbia), 2024, para 15). The words "bereavement fares" were a link to a page headed "Bereavement travel", which said the policy did not apply to requests made after travel was completed (Civil Resolution Tribunal (British Columbia), 2024, para 16 and 17).

Relying on the chatbot, Moffatt booked two one-way flights: Vancouver to Toronto on 12 November for $794.98, and Toronto to Vancouver on 18 November for $845.38 (Civil Resolution Tribunal (British Columbia), 2024, para 18). Moffatt says that on 11 November they also phoned the airline and were told the bereavement fare would be about $380 each way. The tribunal records no evidence that the representative said anything about applying retroactively (Civil Resolution Tribunal (British Columbia), 2024, para 19).

The first application went in on 17 November 2022, inside the 90 days (Civil Resolution Tribunal (British Columbia), 2024, para 20). On 8 February 2023 an Air Canada representative admitted that the chatbot had used "misleading words", pointed to the policy page, and said the airline had noted the issue so it could update the chatbot (Civil Resolution Tribunal (British Columbia), 2024, para 22).

How did Air Canada defend itself, and how did the tribunal answer?

The tribunal rejected three of the airline's arguments. The claim was treated as negligent misrepresentation, a civil wrong with five elements: a duty of care, a false statement, carelessness, reasonable reliance and loss (Civil Resolution Tribunal (British Columbia), 2024, para 25). The tribunal found all five (Civil Resolution Tribunal (British Columbia), 2024, para 26 to 32).

First, the airline said it could not be held liable for information given by an agent, servant or representative, including a chatbot, without explaining why. The tribunal read this as suggesting the chatbot was a separate legal entity responsible for its own actions, and wrote: "This is a remarkable submission." A chatbot is still part of the website, the airline should know it is responsible for everything on that website, and it makes no difference whether the words come from a static page or a chatbot (Civil Resolution Tribunal (British Columbia), 2024, para 27).

Second, the airline said the correct information sat on another page. The tribunal replied that the airline did not explain why the "Bereavement travel" page was more trustworthy than the chatbot, nor why customers should have to double-check one part of its website against another (Civil Resolution Tribunal (British Columbia), 2024, para 28).

Third, the airline pointed to terms in its Domestic Tariff but supplied only its account of what the tariff says. The tribunal called it a sophisticated litigant that should know it is not enough to assert that a contract says something without producing the contract, so no contractual defence had been proved (Civil Resolution Tribunal (British Columbia), 2024, para 31).

How much did Air Canada have to pay, and how was it worked out?

$812.02 within 14 days: $650.88 in damages, $36.14 in pre-judgment interest and $125 in tribunal fees (Civil Resolution Tribunal (British Columbia), 2024, para 44). The decision prints only "$"; because the tribunal sits in British Columbia I read it as Canadian dollars, which the text does not state.

Moffatt claimed $880 (Civil Resolution Tribunal (British Columbia), 2024, para 3). The tribunal recalculated:

The $380 figure rests on Moffatt's account of the phone call. The tribunal accepted it by adverse inference: Air Canada was best placed to show what the fare would have been, produced no evidence, and so would have supplied it had it helped (Civil Resolution Tribunal (British Columbia), 2024, para 35 and 36). Moffatt treated $380 as all-in and worked out $880.36; the tribunal corrected that to $870.36, the maximum (Civil Resolution Tribunal (British Columbia), 2024, para 34). It then added fixed charges and tax, which the original tickets included, bringing damages to $650.88 (Civil Resolution Tribunal (British Columbia), 2024, para 37 to 40). The airline's $200 goodwill coupon was not set off, because Moffatt said they never accepted it and the airline had no evidence otherwise (Civil Resolution Tribunal (British Columbia), 2024, para 41).

What does the decision say about why the chatbot was wrong, and how far does it reach?

On the cause it says nothing. Air Canada gave no information about the nature of its chatbot, so the tribunal describes chatbots only generally: an automated system that answers a person using a website in response to what they type (Civil Resolution Tribunal (British Columbia), 2024, para 14). The decision records one word from the airline on the cause: the employee's email saying the chatbot used "misleading words" (Civil Resolution Tribunal (British Columbia), 2024, para 22).

So I do not know whether this chatbot was generative AI or a scripted responder, or whether it invented its wording, quoted an old text or pulled the wrong document. Three failure modes of generative AI (invented content, agreement with the asker, answers from out-of-date knowledge) is a list for checking AI output, and the decision places this case in none of them. The tribunal's reasoning stops at the answer being wrong and the airline not taking care to make it right (Civil Resolution Tribunal (British Columbia), 2024, para 28).

On reach, this is one tribunal member's decision, Christopher C. Rivers (named at the top), in a small claims dispute before the Civil Resolution Tribunal of British Columbia (Civil Resolution Tribunal (British Columbia), 2024, para 7). I have not checked whether any other court has relied on it, and nothing here makes it binding anywhere.

What should a manager change in the process?

Six things, all my own inference from the decision rather than anything the tribunal wrote.

1. Give each policy one official text, and have the chatbot draw its answers from it. The tribunal refused the argument that the right answer was on another page (Civil Resolution Tribunal (British Columbia), 2024, para 28), so two places saying two things is your risk, not the customer's.
2. Give each group of answers a named owner. The airline's employee said the issue had been noted so the chatbot could be updated (Civil Resolution Tribunal (British Columbia), 2024, para 22); the decision does not say who did it or when.
3. Write down what the bot may not promise: refunds, prices, deadlines, exceptions. The quoted part of the wrong answer in the decision contains one deadline (90 days) and one form (Civil Resolution Tribunal (British Columbia), 2024, para 15). Have the bot quote the policy word for word with a link, and pass exceptions to a person.
4. Keep conversations with timestamps. The chatbot's wording in the decision came from the customer's screenshot (Civil Resolution Tribunal (British Columbia), 2024, para 15 and 21).
5. Hand over to a person when a chat touches money or law, with a script for exactly that question. Here the phone representative quoted about $380 on Moffatt's account, and the decision records no evidence of what they said about retroactive claims (Civil Resolution Tribunal (British Columbia), 2024, para 19).
6. Set a review cadence: whenever the policy changes, and regularly by sampling conversations, ask the bot the money questions again and compare with the official text. How to check in rungs is in the verification ladder post.

What does this look like in a made-up online travel agency?

The example is entirely invented: "Ví Dụ Travel" (Vietnamese for "Example Travel") is an online agency I made up, and every figure is mine.

Ví Dụ Travel takes 1,000 bookings a month at an average of VND 6m a tour. On review, the manager finds three places giving three cancellation deadlines: the policy page says cancel 7 days before departure for a 100% refund less a 5% fee; the promotion page says free cancellation up to 3 days before; the chatbot says a full refund within 48 hours of booking.

Suppose the chatbot's wrong deadline reaches 2% of bookings, 20 customers a month, each off by 30% of the tour price, VND 1.8m. About VND 36m a month, over VND 430m a year, sits in the zone of dispute. Reviewing 20 chats at 10 minutes each takes a little over three hours. I have not checked Vietnamese law, so I do not say the agency would have to honour its chatbot.

Going deeper: what did I read, what did I not, and where does the "three failure modes" frame stand?

I read the full decision, Moffatt v. Air Canada, 2024 BCCRT 149, from the tribunal's own site. I have not read the original chat logs or screenshots, Air Canada's Domestic Tariff (the airline did not supply it either, para 31), any later commentary, or Vietnamese or UK law.

Adding up the decision's figures, I found a mismatch. The two fares in para 18 sum to $1,640.36, while paras 34 and 40 use $1,630.36 (Moffatt's $880.36 matches $1,640.36). In para 38, 380 plus 62.12 plus 22.11 is 464.23, but the decision prints 474.23. Following the itemised figures at both points, the damages would be $670.88, not $650.88. I report the tribunal's order; I do not know which numbers it meant.

The Three failure modes of generative AI frame is this site's grouping; its sources do not call the three modes "the three core failures", and the Air Canada case is not among them.