Air Canada chatbot gives incorrect bereavement-fare advice

On November 11, 2022, after their grandmother died, Jake Moffatt used Air Canada’s website chatbot to ask about reduced bereavement fares. It said a passenger could travel and then apply for a reduced rate within 90 days of the ticket’s issue. The linked policy page said bereavement consideration could not be requested after travel was completed. Moffatt relied on the chatbot, bought flights and later sought a partial refund. On February 14, 2024, British Columbia’s Civil Resolution Tribunal found Air Canada liable for negligent misrepresentation: failing to take reasonable care over misleading information on which a customer reasonably relied (§§13–18, 24–32).^[1]

The booking and refund dispute

Moffatt booked Vancouver–Toronto travel on November 11 for departure the next day, then the return journey on November 16 for departure on November 18. They said an airline representative had estimated a bereavement fare of about CAD 380 each way by telephone; the decision records no evidence that this representative addressed retroactive applications. Moffatt first applied for the reduced fare on November 17, within the chatbot’s stated window (§§18–20). The chronology dates the initial advice and booking, rather than the later ruling.^[1]

Emails continued through December 2022 and February 2023. On February 5, Moffatt sent the chatbot screenshot and confirmed submitting the refund form and death certificate. On February 8, an Air Canada representative acknowledged misleading wording, pointed to the linked policy and said the issue had been noted so the chatbot could be updated. The parties did not resolve the dispute (§§20–23). This was a proposed correction, not evidence that an update worked.^[1]

Responsibility and the award

The tribunal considered documentary evidence and written submissions, without an oral hearing. Moffatt represented themself; an employee represented Air Canada. The airline argued that retroactive requests were ineligible and that it could not be liable for its chatbot’s information. Member Christopher C. Rivers rejected that position: the chatbot was part of the airline’s website. The description of a separate legal entity was Rivers’s characterization of the airline’s argument, not a finding that the chatbot had legal personhood. Linking to correct information did not explain why customers should trust one section of the website over another (§§5, 8, 27–29).^[1]

Rivers accepted that Moffatt would not have flown last-minute at full price and had reasonably relied on the advice. Air Canada’s contractual defence also failed because it had not supplied the relevant tariff provisions. This was a failure to prove that defence, rather than a ruling that contractual terms can never matter (§§30–32).^[1]

Moffatt initially claimed CAD 880. The award was smaller because Rivers treated CAD 380 each way as the base fare and added taxes and fixed charges, producing a hypothetical total of CAD 979.48. Using the CAD 1,630.36 paid total stated in the damages section, he awarded CAD 650.88. The reasons contain an unresolved CAD 10 inconsistency: the two ticket amounts in §18 add to CAD 1,640.36, while §§34 and 40 use CAD 1,630.36. These are the decision’s figures, not a recalculated award. A proposed CAD 200 coupon did not reduce damages because Moffatt said they had declined it and the airline supplied no contrary evidence (§§3, 33–41).^[1]

The order required payment within 14 days: CAD 650.88 damages, CAD 36.14 pre-judgment interest and CAD 125 tribunal fees, totaling CAD 812.02, with applicable post-judgment interest (§§42–47). On February 15, The Canadian Press reported Air Canada’s statement that it would comply and regarded the matter as closed. That statement establishes its announced response, not confirmation of payment.^[1][2]

What the incident establishes

The record establishes conflicting automated advice, customer reliance and an institutional responsibility finding. Air Canada supplied no information about the chatbot’s nature (§14), so the decision does not identify a language model or explain how the incorrect answer arose.^[1] Hallucination and the hallucination survey offer vocabulary for inaccurate or unsupported generated text; they do not reconstruct this system. Neither approval-seeking behavior nor a hidden objective follows from an incorrect answer.

The incidents guide compares the evidence available across failures. Here the practical distinction is between providing a correct policy somewhere on a website and ensuring that the advice a customer acts on is accurate.