Moffatt v Air Canada: what the $650 ruling settled

TL;DR. In November 2022 a man booking a flight to his grandmother's funeral asked Air Canada's website chatbot about bereavement fares. It told him he could claim the discount within 90 days after flying. The airline's own policy page, which the chatbot linked to, said the opposite. Air Canada refused the refund and argued at tribunal that the chatbot was a separate entity responsible for its own actions. The tribunal rejected that and awarded $650.88, total $812.02 with interest and fees. It is now the most-cited AI liability decision in the world. It is also a small-claims ruling from an informal online tribunal, decided on documents, with no counsel and no contracts in evidence, and not binding on any other court. Both halves matter.

---

Status: established. Decided by the British Columbia Civil Resolution Tribunal, 14 February 2024, published as Moffatt v Air Canada, 2024 BCCRT 149. Tribunal member Christopher C. Rivers. The facts below are from the published decision.

---

On 11 November 2022, Jake Moffatt's grandmother died. He went to Air Canada's website to book a last-minute flight from Vancouver to Toronto for the funeral.

He asked the site's chatbot about bereavement fares. It told him he could book at full price and apply for the reduced rate within 90 days of the flight. It included a link to the airline's bereavement travel page.

That page said the opposite. Air Canada's actual policy did not allow bereavement consideration after travel was completed.

Separately that day, an Air Canada representative told him the bereavement rate would put each leg at roughly $380. He booked the outbound for $794.98 and, five days later and still relying on the same understanding, the return for $845.38.

On 17 November he applied for the partial refund. Air Canada refused, pointing to the policy page, and acknowledged that the chatbot had used misleading wording.

He took it to the tribunal.

What Air Canada argued

The defence is the reason this case is cited, and it is worth stating precisely because it is frequently exaggerated in retelling.

Air Canada submitted that the chatbot was a separate legal entity responsible for its own actions.

The tribunal did not accept it. The reasoning was not about artificial intelligence. It was that Air Canada is responsible for all the information on its website, whether that information appears on a static page or comes from a chatbot, and that a consumer cannot be expected to check one part of a company's website against another part.

That second point does more work than the first, and it is the part with the widest application. The chatbot had linked to the correct page. Air Canada's position was effectively that the link discharged the obligation. The tribunal held it did not.

The test that was applied

Negligent misrepresentation, an ordinary tort with five elements, none of them new. The applicant had to show a duty of care, a representation that was untrue, inaccurate or misleading, that it was made negligently, that he reasonably relied on it, and that the reliance caused damage.

The tribunal found a duty of care arose from the commercial relationship between a service provider and a consumer, that the chatbot's information was inaccurate, that Air Canada did not take reasonable care to ensure its chatbot was accurate, and that Moffatt had relied on it reasonably and lost money as a result.

No new legal doctrine was created. An existing tort was applied to a new kind of statement, and the analysis would have run identically if a human agent had given the same wrong answer.

Damages of $650.88, the difference between what he paid and the bereavement fare. With pre-judgment interest and tribunal fees, $812.02.

What it establishes, stated narrowly

Three things, and each is smaller than the headline version.

An operator is responsible for what its chatbot says on its own website. Established, in this forum, on these facts. The separate-entity argument failed.

A link to correct information does not cure a misleading statement. This is the most transferable holding and the one least often quoted. A system that gives a wrong answer with a correct citation attached has still given a wrong answer.

Reasonable reliance on a customer-facing automated system is available as a claim. A consumer who acts on what a company's automated tool tells them is not thereby the author of their own loss.

And it does not establish that AI systems have or lack legal personality, which is how the case is sometimes reported. The tribunal did not need to decide that, and did not.

What it does not establish, which is more than people assume

This is where most citations of this case go wrong, and the honest treatment requires stating it plainly.

It is not binding on anything. The Civil Resolution Tribunal is a British Columbia small-claims body with a monetary limit of $5,000. Its decisions do not bind other tribunals or any court, in Canada or elsewhere.

There was no hearing. The matter was decided on documentary evidence, with no oral testimony and no cross-examination.

Neither party had counsel. The tribunal is designed for self-represented parties, which is a feature of its accessibility mandate and a limitation on the quality of argument it receives.

The contracts were never filed. A law review analysis noted that the relevant contractual documents did not form part of the evidence, and described the resulting reasoning as disappointing. Contractual defences Air Canada might have raised were therefore never tested, and a differently argued case could produce a different result.

And the sum is $650. The most-cited artificial intelligence liability decision on earth concerns the price difference on two domestic flights.

None of that makes it wrong. All of it constrains what can be built on top of it, and a case cited as settling corporate liability for AI worldwide is carrying considerably more weight than its procedural posture supports.

Why it is cited so heavily anyway

Worth explaining, because the gap between its authority and its influence is itself informative.

It was first, and it was clear. A company argued in a published proceeding that its chatbot was a separate entity, and a decision-maker rejected it in writing. Before this, the question was hypothetical.

The defence was memorable. The separate-entity argument is easy to describe and easy to find unreasonable, which makes the case travel.

And the facts are sympathetic without being extreme. A bereaved man, a modest sum, an airline pointing at its own fine print. Nothing about it requires technical explanation.

A case becomes a landmark by being citable, not by being authoritative, and this one is exceptionally citable. That is a fact about how legal ideas spread rather than a criticism of the decision.

The pattern this case names

Read as an incident rather than a judgment, this is not a hallucination case, and mistaking it for one leads organisations to look for the wrong exposure.

The chatbot did not invent a policy that never existed. It described a bereavement policy in terms that are entirely ordinary in the airline industry. Several carriers do permit retroactive bereavement claims within a window. The answer was plausible because it was true somewhere, and it was wrong because it was not true here.

That failure mode has a name worth using: a contradiction between two things the same organisation publishes.

It is far more common than fabrication and much harder to detect, for three reasons.

It survives a plausibility check. A reviewer reading the chatbot's answer in isolation would find nothing wrong with it. The error is only visible against the policy page.

It survives a source check. The chatbot cited the correct page. Anyone verifying that a source existed would have passed it.

And it is created by ordinary organisational change. A policy is updated, the website page is updated, and the system configured or trained before the change keeps giving the old answer. Nobody made a mistake at any single point.

The test that would have caught it is not a model evaluation. It is a consistency check between what the system says and what the organisation's own documents say, run continuously rather than at launch, and treating any disagreement as a defect regardless of which side is right.

Almost nobody runs one. It is the cheapest available control in this entire series and it requires no machine learning expertise at all: take the questions the system is asked most, take the pages that answer them, and check weekly whether the two still agree.

The operational lesson

Four things follow for anyone running a customer-facing automated system, none of which requires a lawyer to act on.

Your chatbot's answers are your statements. The separate-entity theory has been tested once and failed. Treat output as published policy.

A citation does not discharge the duty. The chatbot linked to the correct page and the tribunal held that insufficient. If a system can produce an answer that contradicts a linked source, the answer is the problem.

The failure mode is a contradiction between two things you publish. This was not a hallucination in the usual sense. The chatbot's statement was a plausible-sounding policy that happened to be wrong and happened to conflict with the page beside it. Any organisation with a policy that changed, and a system trained or configured before it changed, has this exposure.

And the cost is not the damages. $650 is nothing. The tribunal decision, the international coverage, and two years of the company's name being the standard example are the actual cost, and none of it appears on the judgment.

What the record does not show

Applying the standard set for this series, here is what is not established by the available material.

Whether the chatbot was a language model. The decision describes an automated system that responds to prompts. Its architecture is not in evidence, and reporting that calls it an AI chatbot is describing it in 2024 terms rather than from the record.

Why it produced the wrong answer. No technical account exists publicly. Whether it was stale training content, a retrieval error, a configuration written before a policy change, or a hand-authored response is unknown.

What Air Canada changed afterwards. The company removed the chatbot from its website following the decision, which is reported and is not in the ruling. Whether anything else changed internally is not public.

And whether this is representative. One decided case is not a rate. How often customer-facing systems give contradictory answers, and how often anyone pursues it, are both unknown.

The counter-argument

The separate-entity argument was worse in the retelling than in the filing. Air Canada's position is usually described as claiming the chatbot was a legal person. Read narrowly, it was an argument about which party bears responsibility for a third-party tool integrated into a site, which is a question that arises constantly with embedded software and is not absurd. It still failed, and it was a normal commercial argument rather than a bizarre one.

A small-claims decision is the appropriate forum for a small claim. Criticising the ruling for its procedural limits is criticising it for being what it is. The tribunal exists to resolve $650 disputes accessibly, it did so, and the reasoning was adequate for that purpose whatever a law review makes of it.

The outcome was obviously correct. A company published two contradictory statements, a customer relied on one, and the company refused to honour it. No amount of procedural criticism changes that a person was owed money and did not receive it until a tribunal ordered it.

And its influence may be doing useful work regardless of its authority. Organisations that reviewed their chatbot governance because of this case are better off, and whether the decision technically binds them is beside the point. A non-binding case that changes behaviour has more effect than a binding one nobody reads.

The short version

On 11 November 2022 Jake Moffatt asked Air Canada's website chatbot about bereavement fares while booking a flight to his grandmother's funeral. It told him he could claim the discount within 90 days after flying. The airline's own policy page, hyperlinked from the chatbot's answer, said no such claim was possible after travel. He booked at full price, applied for the refund, and was refused.

At tribunal, Air Canada argued the chatbot was a separate entity responsible for its own actions. The argument failed. The tribunal held the airline responsible for all information on its website, from a static page or a chatbot alike, and found that a consumer cannot be expected to check one part of a website against another. Damages of $650.88, total $812.02.

No new doctrine was created. Negligent misrepresentation is an ordinary tort with five ordinary elements, and the analysis would have been identical had a human agent given the same wrong answer.

What it establishes is narrow and useful. An operator answers for its chatbot's statements. A link to the correct information does not cure a misleading one. Reasonable reliance on a customer-facing automated system is a viable claim.

What it does not establish is larger than most citations assume. It binds nothing. There was no hearing, neither party had counsel, and the contracts were never filed, so contractual defences went untested. The most-cited artificial intelligence liability decision in the world concerns the fare difference on two domestic flights and was decided in small claims.

Both halves are the point. A case can be truly important and procedurally slight at once, and citing the first without the second is how a $650 ruling ends up carrying the weight of a doctrine it never created.

Common questions

What was Moffatt v Air Canada about? A man booking a flight to his grandmother's funeral in November 2022 asked Air Canada's website chatbot about bereavement fares. The chatbot told him he could claim the reduced rate within 90 days after flying. The airline's own policy page, which the chatbot linked to, said no claim was possible after travel was completed. He booked at full price, applied for a partial refund, and was refused. The British Columbia Civil Resolution Tribunal found for him in February 2024.

Did Air Canada really argue its chatbot was a separate entity? Yes. Air Canada submitted that the chatbot was a separate legal entity responsible for its own actions. The tribunal rejected it, holding that the airline was responsible for all information on its website whether it appeared on a static page or came from a chatbot. Read narrowly the submission was an argument about responsibility for an integrated third-party tool, which is a common commercial question, though it is usually retold in a stronger form than it was filed.

How much was awarded? $650.88 in damages, representing the difference between the fare paid and the bereavement fare, with a total of $812.02 including pre-judgment interest and tribunal fees. The Civil Resolution Tribunal has a small-claims limit of $5,000.

Is the decision legally binding? No. The Civil Resolution Tribunal is a British Columbia small-claims body and its decisions do not bind other tribunals or any court. The matter was decided on documentary evidence with no oral hearing, neither party was represented by counsel, and the relevant contracts were never filed in evidence, so contractual defences were never tested. It is widely cited and it is not authority.

What legal principle did it establish? No new one. It applied negligent misrepresentation, an ordinary tort requiring a duty of care, an inaccurate representation, negligence, reasonable reliance and resulting damage. The analysis would have been identical if a human agent had given the same wrong answer. What is useful is the application: an operator answers for its chatbot's statements, and providing a link to correct information does not cure a misleading one.

Why is this case cited so often? Because it was first, the defence was memorable, and the facts are sympathetic without needing technical explanation. Before it, corporate liability for chatbot output was hypothetical. A case becomes a landmark by being citable rather than by being authoritative, and this one is exceptionally citable, which is why its influence considerably exceeds its formal authority.

What should companies do differently because of it? Treat chatbot output as published statements of policy. Recognise that linking to the correct page does not discharge the duty if the answer itself is wrong. And look specifically for contradictions between what a system says and what the rest of the site says, since this was not a hallucination in the usual sense but a plausible answer that conflicted with the page beside it. Any organisation whose policy changed after a system was configured has the same exposure.

Do we know why the chatbot gave the wrong answer? No. The decision describes an automated system responding to prompts and says nothing about its architecture. Whether the error came from stale content, a retrieval failure, a configuration predating a policy change, or a hand-written response is not public. Air Canada removed the chatbot from its website after the decision, which is reported rather than part of the ruling.