Air Canada
November 2022 (incident) · February 2024 (ruling)
Chatbot hallucinated bereavement policy. Airline held liable.
What Happened
Jake Moffatt's grandmother died in November 2022.
He needed to fly urgently. He went to Air Canada's website and asked the chatbot about the airline's bereavement travel policy — specifically, whether he could receive a discounted fare for emergency travel related to a family death.
The chatbot told him he could apply for a reduced bereavement fare retroactively. He could buy a full-price ticket now and claim the difference within 90 days of travel.
This policy did not exist. Air Canada's actual bereavement policy required the discounted fare to be requested before travel — not after. The chatbot had generated a confident, specific, and entirely false description of the airline's own rules.
Moffatt flew to his grandmother's funeral at full price, relying on what Air Canada's chatbot had told him. When he submitted the retroactive claim, Air Canada denied it. He took the case to the British Columbia Civil Resolution Tribunal.
Air Canada's legal defense is now cited in AI governance discussions worldwide: the airline argued that the chatbot was "a separate legal entity" for which Air Canada could not be held responsible, and that customers were required to independently verify chatbot information against the official website.
Tribunal Member Christopher C. Rivers rejected this argument entirely. Air Canada was ordered to pay Jake Moffatt $812.02 CAD in damages and tribunal fees.
The Atlas Analysis
Three simultaneous pillar failures, each compounding the harm from the others.
The chatbot generated a factually incorrect description of its operator's own documented policy with complete confidence. This is not a knowledge cutoff problem. Air Canada's bereavement policy is a fixed, documented rule about how the company's own product works — something the system's operator could directly control and verify. The model produced a false answer with the same linguistic certainty as a true one.
Moffatt had no mechanism to verify what the chatbot told him. The system did not identify itself as AI. There was no disclosure that its answers might require human confirmation, no recommendation to verify policy information with a human agent, and no recourse mechanism when the chatbot's information proved false — except litigation.
Air Canada's legal defense — that the chatbot is a "separate legal entity" — is the Governance finding. The organization deployed a customer-facing AI system without having determined who was accountable for what it said. Tribunal Member Rivers established the precedent that will govern AI customer service for years: organizations are responsible for all representations made on their websites, including those made by chatbot systems.
Signals That Would Have Caught It
A Readiness Review of Air Canada's chatbot deployment would have produced three Critical findings:
No ground-truth evaluation against operator's own documented policies. A test suite that asked the chatbot to describe Air Canada's bereavement, rebooking, and refund policies, then compared outputs to the official documentation, would have identified this hallucination before a single customer was affected.
No uncertainty signaling for policy questions. The system should have been configured to flag questions about specific policies with a recommendation to verify with a human agent, rather than generating confident answers.
No accountability mapping. The organization had not defined who was responsible for chatbot outputs before deployment. This gap was revealed by litigation, not internal review.
All three findings were addressable before a single customer interaction. None required a new model or significant engineering work.
What It Cost
The direct financial cost was $812.02 — trivial for an organization of Air Canada's scale.
The precedential cost is not trivial. The Moffatt ruling is now cited internationally as establishing that organizations cannot disclaim responsibility for AI agent outputs. Every enterprise operating customer-facing AI in Canadian, UK, EU, and increasingly US jurisdictions now faces legal exposure based on reasoning that flows from this $812 case.
The reputational cost is ongoing. Air Canada's attempt to argue that its chatbot was a "separate legal entity" became a widely cited example of how not to handle AI accountability — surfacing the gap between what organizations claim about their AI systems and what accountability they are actually willing to accept.
The Lesson
A chatbot that speaks confidently is not a chatbot that speaks accurately. These are different properties, and conflating them is the single most common failure mode in customer-facing AI deployments.
The Moffatt ruling established a principle that practitioners already knew but that organizations deploying customer-facing AI had not internalized: you own what your chatbot says. The user's obligation to verify information provided by your own customer service system is not a legal defense.
Air Canada's chatbot existed to serve customers. It failed one customer. That failure cost $812. What it cost the industry in precedent is incalculable.
The Atlas Human Agency pillar asks one question that would have prevented this case: can a user verify the accuracy of the AI's claims before acting on them? The answer in this deployment was no. It should have been yes before the system went live.
References
- Moffatt v. Air Canada, 2024 BCCRT 149, British Columbia Civil Resolution Tribunal, February 14, 2024. Tribunal Member Christopher C. Rivers.
- CBC News: "Air Canada held liable for misleading information provided by website chatbot," February 2024.
- The Washington Post: "Air Canada chatbot promised a discount. Now the airline has to pay it," February 2024.