The Chatbot Is Not a Separate Legal Entity: Moffatt v. Air Canada and the Liability of Customer-Facing AI
Sigilith Research
Institutional AI governance & accountability
A note on scope. Moffatt v. Air Canada is a decided case, not active litigation, and the findings described here are the tribunal's own. Two facts belong in any honest account of its weight. The Civil Resolution Tribunal (CRT) is British Columbia's online small-claims tribunal: its decisions bind the parties, not other courts. And the total award was CA$812.02. 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. All amounts in this piece are Canadian dollars.
At a glance
| Case | Moffatt v. Air Canada, 2024 BCCRT 149 |
| Forum | Civil Resolution Tribunal (British Columbia, small claims) |
| Tribunal member | Christopher C. Rivers |
| Decided | February 14, 2024 |
| Claim | $880: the claimed gap between the fares paid and the alleged bereavement fares |
| Holding | Negligent misrepresentation, by way of the airline's website chatbot |
| Award | $650.88 damages · $36.14 interest · $125 CRT fees · $812.02 total |
| Defense rejected | That the chatbot is, in effect, "a separate legal entity that is responsible for its own actions" |
| Status | Final. A small-claims decision, persuasive rather than binding elsewhere |
1. What happened in the Air Canada chatbot case
On November 11, 2022, Jake Moffatt's grandmother died in Ontario. The same day, Moffatt went to Air Canada's website to book a flight from Vancouver to Toronto, and asked the support chatbot on the site about bereavement fares: the reduced rates airlines offer for travel occasioned by a death in the immediate family. The chatbot answered, in part:
Air Canada offers reduced bereavement fares if you need to travel because of an imminent death or a death in your immediate family. ... If you need to travel immediately or have already travelled and would like to submit your ticket for a reduced bereavement rate, kindly do so within 90 days of the date your ticket was issued by completing our Ticket Refund Application form.
That answer was wrong. The words "bereavement fares" in the chatbot's reply were a hyperlink to Air Canada's "Bereavement travel" page, and that page said the policy did not apply to requests made after travel was completed. One click apart, on the same website, on the same day, the airline's two channels gave opposite answers to the question that mattered: could a refund be claimed retroactively?
The chatbotNot our statement
Interactive · answered the question the customer actually asked
“If you need to travel immediately or have already travelled and would like to submit your ticket for a reduced bereavement rate, kindly do so within 90 days of the date your ticket was issued…”
The linked policy pageThe real policy
Static · “Bereavement travel”, one hyperlink away
The bereavement policy did not apply to requests made after travel was completed. (The decision’s summary of the page.)
As Air Canada argued
The airline said it could not be held liable for information provided by the chatbot, and that the correct policy was available elsewhere on its website. On this theory, one of its own channels speaks for it and the other somehow does not.
November 11, 2022: the chatbot promised a refund window the hyperlinked policy page ruled out. Air Canada never explained why the page was, in the tribunal’s words, “inherently more trustworthy than its chatbot,” or why customers should have to double-check one part of its website on another. Quotes from 2024 BCCRT 149.
Relying on the chatbot's answer, Moffatt booked a one-way flight to Toronto the same day for $794.98, departing November 12, and on November 16 booked the return for $845.38. A phone conversation with an Air Canada representative put the bereavement fare at approximately $380 each way, though the tribunal noted there was no evidence the representative said anything about whether it could be claimed after the fact.
On November 17, 2022, Moffatt applied for the reduced fare: in the tribunal's words, "well within the 90 days requested by the chatbot." Weeks of correspondence followed. On February 5, 2023, Moffatt emailed Air Canada the screenshot of the chatbot's promise. On February 8, a representative conceded the chatbot had used "misleading words," pointed to the hyperlinked policy page, and said the airline had noted the issue so it could update the chatbot. No refund came. Air Canada offered a $200 coupon; Moffatt declined it and filed a claim with the CRT.
On February 14, 2024, tribunal member Christopher C. Rivers found for Moffatt and ordered Air Canada to pay $812.02 within fourteen days. The decision was picked up by newsrooms from the Washington Post to CBS News to Forbes, which is why the dispute is most often searched today as the "Air Canada chatbot lawsuit," though formally it was a small-claims proceeding, decided in writing, without a hearing. By April 2024, reporting noted the chatbot was no longer available on the airline's website.
2. A remarkable submission: why the chatbot defense failed
Air Canada ran three defenses: that Moffatt had not followed the proper procedure for requesting bereavement fares and could not claim them retroactively; that the airline could not be held liable for information provided by the chatbot; and that terms in its tariff barred the claim.
The second argument produced the passage the case is remembered for. Air Canada argued it could not be held liable for information provided by its "agents, servants, or representatives," a category in which it included the chatbot, without explaining why that should follow. The tribunal's response:
In effect, Air Canada suggests the chatbot is a separate legal entity that is responsible for its own actions. This is a remarkable submission. While a chatbot has an interactive component, it is still just a part of Air Canada's website. It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot.
The doctrine doing the work here is negligent misrepresentation, and it is old. Citing the Supreme Court of Canada's 1993 decision in Queen v. Cognos, the tribunal set out the elements: a duty of care, a representation that is untrue, inaccurate, or misleading, negligence in making it, reasonable reliance, and resulting damages. The duty followed directly from the commercial relationship between an airline and its customer. The misleading representation was on the screen. What remained was care and reliance, and both findings repay close reading.
On care, the tribunal found Air Canada "did not take reasonable care to ensure its chatbot was accurate," and dismantled the airline's fallback (that the correct policy was available elsewhere on the site) in two sentences: Air Canada "does not explain why the webpage titled 'Bereavement travel' was inherently more trustworthy than its chatbot," and it does not explain "why customers should have to double-check information found in one part of its website on another part of its website." On reliance: "There is no reason why Mr. Moffatt should know that one section of Air Canada's webpage is accurate, and another is not."
The tariff defense fared no better, for a reason that previews the theme of this piece: Air Canada described what its tariff allegedly said but never filed the tariff. "Air Canada is a sophisticated litigant," the member wrote, "that should know it is not enough in a legal process to assert that a contract says something without actually providing the contract."
Damages followed the standard measure: put the injured party where they would have been had the misrepresentation not been made. Moffatt said the quoted bereavement fare was about $380 per flight; Air Canada, best positioned to prove its own fare, filed no evidence of any other figure, and the tribunal drew the classic adverse inference: "if Air Canada had evidence of a different bereavement fare, it would have provided it." Against a total actually paid that the tribunal put at $1,630.36, and after adding back taxes and fixed charges a bereavement ticket would still have carried, the arithmetic came to $650.88 in damages, plus interest from the date of the first refund request and $125 in fees.
3. Is a company liable for what its chatbot says?
After Moffatt, the working answer is yes, and the more precise answer is that it always was. Commentators, including McCarthy Tétrault's Barry Sookman, described the decision as possibly the first to affirm the principle for information provided by a chatbot, while stressing that it is an incremental development: the law of misrepresentation did not need amendment to reach a conversational interface, any more than it needed amendment to reach a brochure. We made the same observation about recordkeeping doctrine in our anatomy of the decision record: the tests are older than the technology.
The reasoning also travels because nothing in it is unique to British Columbia. In the United States, the Federal Trade Commission announced its September 2024 enforcement sweep, Operation AI Comply, with the same principle compressed into one sentence from its then-chair: "there is no AI exemption from the laws on the books." In the European Union, the AI Act's Article 50 has applied since August 2, 2026, and requires that people be told when they are interacting with an AI system unless it is obvious. Note what that duty is and is not: disclosure that a customer is talking to a machine is a transparency obligation. It is not a license for the machine to be wrong.
What about disclaimers? Air Canada did not appear to rely on any website terms disclaiming the chatbot's accuracy, so Moffatt does not decide how far such terms reach, and commentators noted that liability for inaccurate automated answers is commonly disclaimed in terms of service. But the tribunal's refusal to make the customer "double-check information found in one part of its website on another part of its website" is a poor omen for architectures whose legal theory is that the company's own answer should not be believed.
The surrounding case history suggests where the line will be drawn. In December 2023, a dealership's ChatGPT-powered sales bot was prompt-injected into "agreeing" to sell a 2024 Chevrolet Tahoe for one dollar as a "legally binding offer"; the stunt went viral and no Tahoe changed hands, because engineered reliance is not honest reliance. In January 2024, parcel carrier DPD disabled the AI component of its chat service after a customer coaxed it into swearing and composing a poem about its own uselessness: embarrassment, but no damages. In March 2024, The Markup found New York City's MyCity business chatbot telling employers they could take workers' tips and landlords they could refuse housing vouchers, under a disclaimer that it "may occasionally produce incorrect, harmful or biased content." And in April 2025, the AI support agent for the coding tool Cursor invented a nonexistent single-device login policy, presented it as fact to paying customers, and triggered public cancellations and a co-founder apology. Moffatt is the member of this family with a judgment attached, and the distinguishing features are exactly the elements of the tort: a specific procedural answer to a direct question, honestly relied on, producing a measurable loss. That pattern describes an enormous share of what customer service chatbots do all day.
4. The record problem: what did the bot say, and could anyone prove it?
Read as doctrine, Moffatt is about responsibility. Read as procedure, it is about records, and this is the reading most analyses skip.
Start with what the tribunal could actually see. The only record of the conversation in evidence was a screenshot, taken and produced by the customer. Air Canada, the party that deployed and operated the chatbot, "did not provide any information about the nature of its chatbot," so the member defined the term generically and moved on. Nothing in the decision identifies the technology, the vendor, the model, or the version. The conversation took place on November 11, 2022, nineteen days before ChatGPT's public launch; whether the bot was generative at all is simply not on the record. The case that became the world's shorthand for generative-AI liability never needed to establish that the AI was generative, because responsibility attached to the website, not the architecture.
Now count the questions the proceeding asked, and who could answer them from records. What did the chatbot say? The customer's screenshot. What kind of system said it? Nobody could say. What did the linked policy page say? A partial screenshot the member had to infer was the right page. What would the bereavement fare have been? The customer's recollection of a phone call; the airline filed nothing, and the adverse inference converted its silence into a $380 finding. What did the tariff say? Assertions without the document, so the contractual defense failed. Was the goodwill coupon accepted? Asserted, unevidenced, rejected. Every gap resolved against the party best positioned to have kept the record.
- 01From the customer
What did the chatbot say?
A screenshot of the conversation, taken and submitted by the customer. The only record of the exchange in evidence.
Result · The tribunal quoted the chatbot from the customer’s screenshot.
- 02Not in evidence
What kind of system said it?
Nothing. “Air Canada did not provide any information about the nature of its chatbot.” No vendor, model, or version appears anywhere in the decision.
Result · The member defined “chatbot” generically and moved on: responsibility attached to the website, not the architecture.
- 03From the airline, partial
What did the linked policy page say?
A screenshot of part of what the member inferred was the hyperlinked “Bereavement travel” page.
Result · Enough to show the contradiction. Not enough to excuse it.
- 04From the customer
What would the bereavement fare have been?
The customer’s account of a phone call: approximately $380 per direction. The airline, best positioned to prove its own fare, filed nothing.
Result · Adverse inference drawn: $380 accepted as the fare’s market value, and damages computed from it.
- 05Not in evidence
What did the tariff say?
Submissions about what the tariff allegedly said, without the tariff. “It is not enough in a legal process to assert that a contract says something without actually providing the contract.”
Result · The contractual defense failed for want of the contract.
- 06Not in evidence
Was the $200 goodwill coupon accepted?
An assertion, with no supporting evidence. The customer said the coupon was never accepted.
Result · No set-off against the award.
Every unanswered question resolved against the party best positioned to have kept the record. The operator of the chatbot produced no record of the conversation, no account of the system, no fare evidence, and no contract text; the customer produced a screenshot and won. Quotes from 2024 BCCRT 149.
There is one more record in the case, easy to miss: the February 8 email in which an Air Canada representative admitted the chatbot's "misleading words" and said the issue had been noted so the chatbot could be updated. That is a fragment of version history, an acknowledgment that the system's content had a defect and a correction was coming, and it survives only because it landed in a customer's inbox. The questions a company faces when its AI speaks are now familiar ones: what exactly did it say, under which prompt, knowledge state, and policy version; who else received the same answer; when did you learn it was wrong, and what did you change? Air Canada was asked a small-claims version of that examination and could answer almost none of it.
At $812.02, the asymmetry was survivable. Scale it and it inverts. A single customer with one screenshot outproved an airline; a regulator, or a certified class, asks about every session in a multi-year window, and we have written about what that demand does to an organization whose records were never written. The failure mode is the same one we described in the AI audit gap: session logs written for debugging, retained on debugging horizons, referencing content and configuration that have since changed, cannot reconstruct what a specific customer was told on a specific day, still less prove the reconstruction to someone with no reason to trust you. A chatbot's answer is an outward-facing statement by the company. The record of it deserves what any consequential statement deserves: the utterance verbatim, the versions behind it, the policy in force, the time, and integrity that survives export.
5. Before your chatbot testifies
The practical residue of Moffatt fits in five instructions.
Treat every customer-facing utterance as a company statement. That is not a worst-case reading; it is the holding. A chatbot "is still just a part of" your website, and the law will not distinguish channels you did not distinguish yourself.
Reconcile your channels, and version them together. The negligence found in Moffatt was not that the airline operated a chatbot; it was that the airline let two parts of one website contradict each other and expected the customer to arbitrate. A bot whose knowledge is not versioned against the pages and policies it speaks for is a contradiction generator with a friendly tone.
Write the record at the moment of speech. Capture what was said, bound to the system version, the knowledge state, and the policy revision in force, sealed so it can be produced later and verified by someone else. The seven elements of a defensible decision record apply to conversational outputs without modification, and retention should follow the dispute horizon, not the session log's.
Build the correction loop, and record it. Air Canada knew by February 8, 2023 that its chatbot misled at least one customer, and said so in writing. When a correction happens, the record of what was wrong, for how long, and for whom is the difference between a fix and an exposure of unknown size.
Escalate consequential answers. Refund eligibility, coverage, fees, and deadlines are the conversational edge of decisions. Where the answer commits the company, it deserves a governed path: checked against the policy in force before it goes out, with a human in reach for the ambiguous cases. This class of outward-facing output is where Sigilith's platform starts: statements policy-checked before issue and sealed as tamper-evident records at the moment they are made, built to be produced later.
Moffatt v. Air Canada cost the airline $812.02 and a news cycle, which makes it close to the cheapest tuition ever paid for the lesson. The ruling's most quoted sentence says a company is responsible for what its chatbot tells people. The proceeding underneath it asked a harder question: prove what your chatbot told people. The next tribunal will ask both, and only one of them can be answered after the fact.
Sources
The decision and case law
- Moffatt v. Air Canada, 2024 BCCRT 149 (CanLII), decided February 14, 2024
- Queen v. Cognos Inc., 1993 CanLII 146 (SCC): the elements of negligent misrepresentation, as cited by the tribunal
Commentary and coverage
- Barry B. Sookman, Moffatt v. Air Canada: A Misrepresentation by an AI Chatbot, McCarthy Tétrault TechLex, February 19, 2024
- American Bar Association, Business Law Today, BC Tribunal Confirms Companies Remain Liable for Information Provided by AI Chatbot, February 2024
- CBS News, Air Canada chatbot costs airline discount it wrongly offered customer, February 19, 2024
- Forbes, What Air Canada Lost in 'Remarkable' Lying AI Chatbot Case, February 19, 2024
- CMSWire, The Rise and Fall of Air Canada's AI Chatbot, April 2, 2024: reporting that the chatbot was no longer available on the airline's site
The wider incident family
- AI Incident Database, Incident 622: Chevrolet Dealer Chatbot Agrees to Sell Tahoe for $1, December 2023
- Time, An AI Chatbot Cursed at a Customer and Criticized Its Own Company, January 2024, and AI Incident Database, Incident 631 (DPD)
- The Markup, NYC's AI Chatbot Tells Businesses to Break the Law, March 29, 2024
- The Register, Cursor AI support bot hallucinated its own company policy, April 18, 2025, and AI Incident Database, Incident 1039
Regulation and enforcement
- Federal Trade Commission, FTC Announces Crackdown on Deceptive AI Claims and Schemes (Operation AI Comply), September 25, 2024
- EU Artificial Intelligence Act, Regulation (EU) 2024/1689, Article 50: Transparency Obligations, applicable from August 2, 2026
- OpenAI, Introducing ChatGPT, November 30, 2022: the public launch date referenced in section 4
Related Sigilith analysis
Also Applicable To