Air Canada’s Chatbot Case: When a Company Was Held Responsible for Hallucinated Policy
The Air Canada chatbot dispute showed that companies remain responsible for automated information on their websites: inconsistent policy answers can create real liability even when a separate page contains the correct rule.
A customer relied on an automated answer about bereavement fares
In November 2022 Jake Moffatt used the chatbot on Air Canada’s website while arranging travel after a family death. The chatbot said a reduced bereavement fare could be requested retroactively within a stated period after ticket issuance. Moffatt bought regular-price tickets and later sought the difference. Air Canada then rejected the request because its actual policy required the bereavement process to be handled differently. The dispute was small in dollar value but important in systems terms: two parts of the same website communicated contradictory rules, and the customer acted on the automated one.[1]
The failure was policy inconsistency, not merely bad wording
A support bot is part of the transaction architecture. If its answers can change customer behavior, its knowledge source needs the same change control and accuracy guarantees as the official policy pages it summarizes.
The chatbot linked to a page that contained the correct policy
The bot’s response reportedly included a link to Air Canada’s bereavement-travel webpage, where the written policy contradicted the chatbot’s statement. Air Canada argued in part that the customer should have followed the authoritative page. The tribunal rejected the idea that a consumer should have to determine which part of one company’s website is trustworthy. That is a critical design lesson. Links do not cure contradictory automation. If a bot generates or retrieves policy text, it should use the same canonical source as the static page and ideally quote or cite that source directly so inconsistencies are detectable before the customer acts.
The tribunal treated the chatbot as part of Air Canada’s website
In Moffatt v. Air Canada, the British Columbia Civil Resolution Tribunal found the airline responsible for the information provided by its chatbot. The decision rejected Air Canada’s argument that it could not be liable for information supplied by an automated agent, describing the chatbot as simply one part of the airline’s website. The tribunal awarded damages based on negligent misrepresentation. The legal principle maps neatly to engineering ownership: deploying automation does not transfer responsibility to the software. The organization still owns content quality, monitoring, correction, and the downstream consequences of inaccurate guidance.[2]
Automation does not create an accountability gap
A company can outsource implementation, use a vendor model, or deploy a chatbot, but the user still experiences one service. Internal component boundaries do not erase external responsibility.
The case is often called an AI hallucination, but the technical record is narrower
The roadmap title uses the modern shorthand ‘hallucinated policy,’ yet the tribunal decision does not establish that the bot was a large language model or that its error arose from generative-model hallucination in the current technical sense. It establishes that an automated chatbot displayed inaccurate policy information. That distinction matters for remediation. If the bot used scripted responses, the fix may be content governance and synchronization. If it used generative retrieval, grounding and generation controls would also matter. Failure analysis should start from the observed system, not from a fashionable label.
Customer-service automation needs a single source of policy truth
Airline policies change across routes, fare classes, jurisdictions, and time. A chatbot that maintains its own copy of policy logic can drift from official pages unless updates are synchronized. A safer architecture stores policy in a canonical structured system and lets every channel—web pages, mobile apps, call-center tools, and bots—retrieve from the same versioned source. Responses should include effective dates and links to the exact governing rule. When ambiguity remains, the bot should escalate rather than improvise. This is ordinary configuration management applied to conversational interfaces.
Testing has to include contradiction detection across channels
Traditional chatbot tests often ask whether common questions receive plausible answers. Policy bots need stronger assertions: does the answer match the tariff, contract, or official policy page word for word on critical conditions? Can an automated checker detect when two channels disagree? Are edge cases such as retroactive requests, exceptions, deadlines, and eligibility rules covered? The Air Canada dispute could have been caught by a regression test that compared the bereavement chatbot response with the canonical policy after every content change. High-consequence answers should be testable as rules, not judged only by conversational quality.[3]
Plausibility is the wrong acceptance criterion for policy
A customer does not need a natural-sounding answer; they need the rule that will actually be honored. Accuracy and consistency should dominate style for transactional questions.
Legal commentary treated the decision as a warning for deployed chatbots
Canadian and international legal analyses highlighted the case because the tribunal treated automated website content like any other representation made by the company. The amount awarded was modest, but the precedent-like lesson for operations is larger: a bot can create reliance. Organizations should log answers, preserve the policy version used, provide correction workflows, and monitor complaints for systematic errors. Those controls help both users and the organization reconstruct what happened when an automated statement is challenged.[4]
The lasting lesson is that automated answers must be governable business records
A support chatbot should not be treated as a playful layer floating above official systems. When it answers questions about refunds, eligibility, deadlines, prices, benefits, or contracts, it is publishing operational information on behalf of the organization. That requires ownership, source control, testing, logging, escalation, and rapid correction. The Air Canada case became famous because the company tried to distance itself from its chatbot, but the engineering lesson is even simpler: every automated answer needs an accountable source and an accountable owner. If a business cannot prove what rule its bot used and whether that rule matches the service it will actually provide, the bot is not ready for consequential policy questions.[5]
Policy answers need traceable provenance
Every consequential chatbot response should be traceable to a versioned policy source so operators can reproduce what the user saw, correct stale logic, and detect contradictions across channels.
Works Cited
- 01
- 02CanLII Blog — Summary and key tribunal passage blog.canlii.org
- 03
- 04
- 05CIGI — Data governance and the Air Canada chatbot example cigionline.org
CodeHistory is a living archive. Citations document the evidence used for this edition; later evidence may refine the account.