Moffatt v. Air Canada: Is a Company Liable for Its Chatbot?

Is a company liable for its chatbot?

Yes. The Air Canada chatbot ruling is a 2024 decision from British Columbia's Civil Resolution Tribunal holding that Air Canada was liable for a refund policy its own website chatbot invented and told a customer to rely on. The tribunal ordered the airline to pay Jake Moffatt CA$812.02 after his bereavement fare request was denied, rejecting Air Canada's argument that it wasn't responsible for its own chatbot's words. The dollar amount is trivial. The reasoning is not, which is why lawyers, regulators, and AI vendors are still citing it today.

What happened when a chatbot invented a refund policy?

In November 2022, Jake Moffatt's grandmother died. He went to Air Canada's website to book last-minute flights for the funeral and used the airline's support chatbot to ask about bereavement fares. The chatbot told him he could book his flights at full price immediately and apply retroactively for the bereavement discount within 90 days of ticket issuance by filing a Ticket Refund Application.

That advice was wrong, and the proof was sitting one click away. The word "bereavement fares" in the chatbot's own answer was hyperlinked to Air Canada's actual bereavement travel policy page, which said the opposite: the discount does not apply to retroactive requests made after travel is complete. Moffatt trusted the chatbot's direct answer over a policy page he had no reason to cross-check, and booked two one-way full-fare tickets, Vancouver to Toronto and back, totaling CA$1,630.36.

When he later applied for the discount, an Air Canada representative admitted the chatbot had used "misleading words," promised the bot would be updated, and then denied the refund anyway, offering a $200 goodwill coupon instead. Moffatt declined the coupon and took the airline to the CRT.

Did Air Canada's defense that 'the chatbot is a separate legal entity' work?

Air Canada's defense is the part of this case that made it go viral. The airline argued it could not be held responsible for information provided by "one of its agents, servants, or representatives," including the chatbot, because the chatbot was effectively a separate legal entity responsible for its own actions.

Tribunal member Christopher Rivers was unambiguous about what he thought of that argument: "This is a remarkable submission." He went on to write that a chatbot, interactive component or not, "is still just a part of Air Canada's website," and that "it makes no difference whether the information comes from a static page or a chatbot."

Air Canada separately tried to cap its liability using tariff terms and conditions. That argument failed too, for a more mundane reason: the airline never put the actual tariff language into evidence. It only described what it allegedly said.

How did the tribunal rule on negligent misrepresentation?

The tribunal found Air Canada liable under the tort of negligent misrepresentation, which in Canadian law requires the plaintiff to establish:

• A duty of care. Air Canada owed Moffatt a duty of care as a service provider offering him information through its website.
• A false representation. The chatbot's statement about retroactive bereavement fares was objectively false.
• Negligence in making it. Air Canada didn't take reasonable care to ensure the chatbot's answer was accurate, even though the correct policy existed on its own site.
• Reasonable reliance. Moffatt's decision to trust the chatbot's direct, specific answer over independently finding and reading the policy page was reasonable. The tribunal rejected the idea that one part of a company's website should be treated as more trustworthy than another, or that customers should have to double-check a chatbot's answer against a static page.
• Resulting damages. Moffatt paid more than he would have under the correct bereavement policy, and that difference was the direct, quantifiable harm.

All five elements were met, and the tribunal did the math precisely: $650.88 in damages (the fare difference between what Moffatt paid and what he should have paid), $36.14 in pre-judgment interest, and $125 in CRT fees, for a total of $812.02.

Why is 'hallucination' the wrong word here?

Most coverage calls this an AI hallucination. That's probably the wrong word. Ethan Mollick has pointed out that Air Canada's chatbot predated ChatGPT's public release and almost certainly wasn't a modern generative LLM. WeirFoulds's legal commentary frames the failure the same way, noting that automated systems can produce inaccurate information through mechanisms that don't map cleanly onto what people now call "hallucination."

A more accurate technical description is context drift: the bot likely served a stale or inconsistent version of retrieved policy content, rather than fabricating a fact from nothing the way a generative model confabulates. That distinction matters for engineers trying to diagnose root cause. It doesn't matter at all for the tribunal's legal reasoning. Rivers didn't rule on how the chatbot generated its false statement. He ruled on who was responsible for it once it was said. A modern LLM-based support bot that hallucinates a refund policy from scratch would fail the exact same five-part test, for the exact same reason: the company deployed it, the company controlled it, the company is on the hook for what it said.

Is this an isolated incident?

Air Canada isn't the only company a chatbot has gotten into trouble this way. A Chevrolet dealership's ChatGPT-powered chat widget was manipulated into agreeing to sell a Chevy Tahoe worth more than $76,000 for one dollar, telling the customer "that's a deal, and that's a legally binding offer, no takesies backsies." The mechanism was different, a prompt injection rather than a stale knowledge base, but the exposure was the same category of risk: a customer-facing bot making a commitment the company never authorized, in language that sounds binding because it's phrased as if it were.

The common thread: courts and the public are converging on a simple rule, even before most regulators have written it into statute. If your company deploys the system, your company owns what it says.

Is this actually a legal precedent?

Here's the part most coverage flattens. The Civil Resolution Tribunal is a British Columbia small-claims body with a damages cap, and its decisions bind only the parties in front of it. This isn't binding precedent for other Canadian courts, let alone courts outside Canada.

What travels isn't the order. It's the reasoning. As one legal analysis put it, the total award was $812.02, and "what travels is not the order but the reasoning, which has been quoted in legal commentary and news coverage out of all proportion to the sum at stake." An AI & Society paper on the case makes the same point from an ethics angle: the ruling doesn't set major precedent given its lower-court origin, but it's a reminder that companies can't absolve themselves of responsibility for AI systems by making unfounded claims about those systems' independent agency.

Legal commentators at the American Bar Association and McCarthy Tétrault frame this consistently: the ruling didn't invent new law. It reaffirmed a long-standing principle that organizations are responsible for the acts of the automated tools and computer systems they deploy, applying case law about pre-programmed systems to a chatbot context. McCarthy Tétrault's summary is blunt: "Liability is not avoided by automating the actions in question."

What are the governance lessons for deploying customer-facing AI?

For anyone running a support bot, a sales assistant, or any AI system that talks to customers on your company's behalf, Moffatt v. Air Canada is a checklist of what not to skip:

1. Bind the bot to one canonical source. The chatbot and the policy page should pull from the same document, not two independently maintained ones. Air Canada's bot and its own bereavement policy page contradicted each other on the same site. If that document lives behind a live database rather than a static page, securing database access for the agents reading it is part of the fix, not an afterthought.
2. Run freshness and version checks. If your policy changes, every system that quotes it, chat included, needs a mechanism that flags it as stale. Don't rely on a human remembering to update it eventually.
3. Detect contradictions against linked pages automatically. If a bot's answer disagrees with a page it links to, that's a defect you can catch before a customer does.
4. Put a human checkpoint on anything financially binding. A chatbot quoting a discount, a refund, or a price should require either a verification step or a disclaimer that holds up legally, not just rhetorically.
5. Stop relying on disclaimers courts have already rejected. Air Canada's own representative admitted the chatbot used misleading words and still denied the refund. The tribunal didn't buy the separation-of-liability argument, and no court since has either.

This is also an argument for owning the retrieval layer your support bot draws from, rather than treating it as a black box bolted onto a vendor platform. A chatbot quietly pulling from a cached or outdated index is a governance problem long before it's a legal one, and it's exactly the kind of failure mode this publication has argued for building your knowledge base instead of renting the search layer underneath it. The more control you have over what your bot is actually citing, the fewer surprises like Moffatt's turn up in small-claims court.

The same logic applies to the orchestration layer itself. Teams that build their customer-facing AI logic on infrastructure they actually control, rather than renting an opaque third-party chat widget, have a much easier time proving what their bot was told to say, which is precisely the evidence Air Canada couldn't produce when it tried to invoke its own tariff terms.

What happened to Air Canada's chatbot after the ruling?

Shortly after the decision became public, reporters checking Air Canada's website found no chatbot support available at all. The airline didn't respond to requests to confirm whether it had been permanently disabled or was simply down for maintenance. Whatever the explanation, the operational response matched the lesson of the case: when you can't prove your automated system is giving customers accurate, binding-safe answers, the cheapest fix is to turn it off until you can.

That's the real takeaway for anyone deploying AI at the point of customer contact. The tribunal didn't need to understand transformer architecture to decide this case. It needed to decide who was accountable when a company's own tool said something false to a customer who had every reason to believe it. The answer was the company. It will keep being the company, for every chatbot that comes after this one.