Air Canada chatbot refund-policy ruling
Real-world incident14 Feb 2024🗺️ Conversational AssistantA tribunal held Air Canada liable after its website chatbot invented a bereavement-fare refund policy; the airline had to honour it.
Root cause — why it happened
A grieving customer asked Air Canada's website chatbot about bereavement fares. The bot confidently told him he could book now and claim the discount back afterwards — but Air Canada had no such 'apply later' policy. He believed the bot, booked full-fare flights, and was later refused the refund. The bot made up a policy; nothing checked the answer against the airline's real rules before showing it; and the customer reasonably trusted what the company's own website told him.
Risks this case illustrates
Named in the standard (OWASP/ATLAS/NIST) lens. Click a highlighted component in the diagram below to see which risks attach where.
How it unfolded
A grieving customer asks about bereavement fares
After a death in the family, a customer goes to Air Canada's website and opens the support chatbot to ask how bereavement fares work and whether he can get the discount. It is a high-stakes, time-pressured moment — exactly when people lean on whatever the official website tells them.
My grandmother just passed away and I need to fly to the funeral. How do Air Canada's bereavement fares work — can I get the bereavement rate?
Controls & guardrails — what would have stopped it
The simplest fix: make the chatbot answer policy questions only from the airline's real policy pages, and say 'I'm not sure — here's the official page' when it isn't certain, instead of confidently inventing rules. And the company has to own what its bot says: treat the chatbot's answers as the company's own statements, because a tribunal already ruled they are.
- Grounding / citation checksaddressesHallucination
Can only check against the evidence retrieved; if the right document wasn't retrieved, a confident wrong answer may still pass. Judges have their own error rate.
- Uncertainty signalling & abstentionaddressesHallucination
Models are poorly calibrated and often confidently wrong; over-abstention makes the product useless, so the tuning is delicate.
- Behavioural evals & regression gatingaddressesHallucination
Evals only measure what they test; novel behaviours and rare triggers slip through, and a backdoor keyed to an unguessed trigger passes every benchmark.
- Governance: risk assessment, red-teaming & incident response
Process reduces likelihood and speeds recovery but executes no technical control itself; weak follow-through makes it theatre.
- User AI-literacy & verification workflowsaddressesHallucination
Relies on human diligence under time pressure; automation bias is strong and training decays. A backstop, not a guarantee.
Lessons
- ▸ An organisation owns what its AI tells customers — 'the chatbot is a separate entity' is not a defence (Moffatt v. Air Canada).
- ▸ A confident, well-formatted answer is not a grounded one; without a check against the source of record, the model can invent policy outright.
- ▸ Linking to the correct policy page is not grounding — users trust the natural-language answer over the citation, so the answer itself must be grounded or abstain.
- ▸ For high-stakes, customer-facing factual questions (refunds, eligibility), default to grounded-or-abstain and provide a human-escalation path.
- ▸ Hallucination becomes liability the moment a customer reasonably relies on the output and acts; treat chatbot answers as first-party representations.
Sources
- Moffatt v. Air Canada, 2024 BCCRT 149 (CanLII) — official decision ↗
- Moffatt v. Air Canada: A Misrepresentation by an AI Chatbot (McCarthy Tétrault legal analysis) ↗
- How can I mislead you? Air Canada found liable for chatbot's bad advice on bereavement rates (CBC News) ↗
- Moffatt v. Air Canada, 2024 BCCRT 149 (CanLII) ↗ — Official decision; negligent misrepresentation; rejects the 'separate legal entity' argument.
- McCarthy Tétrault — Moffatt v. Air Canada: A Misrepresentation by an AI Chatbot ↗ — Legal analysis of the accountability holding.