IOanyT Innovations

Lesson 7 of 10 · 8 min read

Liability is real: Air Canada and beyond

In one paragraph

AI liability starts from a simple principle: a business is responsible for what its AI tells customers, just as it is for anything else on its website. In Moffatt v. Air Canada (2024), a Canadian tribunal rejected the argument that a chatbot was a separate legal entity and held the airline liable for its chatbot's inaccurate answer.

In this lesson

  • Explain the reasoning in Moffatt v. Air Canada
  • Describe what 'reasonable care' means for AI answers
  • List practical controls that keep AI answers within what the business has approved

The video version of this lesson is in production. The full lesson is below.

Lesson 1 introduced the case briefly. This lesson looks at its reasoning, because it’s the clearest statement so far of a principle every business deploying AI should design around.

This lesson is a summary for orientation, not legal advice.

What happened

A customer asked Air Canada’s website chatbot about bereavement fares after a death in his family. The chatbot told him he could book at the normal fare and apply for the bereavement discount afterwards, within 90 days of the ticket being issued. That wasn’t the airline’s policy. When he applied, he was refused.

He took the airline to British Columbia’s Civil Resolution Tribunal. The decision, Moffatt v. Air Canada, 2024 BCCRT 149, came in February 2024.

The airline’s argument, and the answer

Air Canada argued, in effect, that the chatbot was a separate legal entity responsible for its own actions. The tribunal rejected that. Its reasoning, in summary:

  1. The chatbot is part of the website. In the tribunal’s words, “it makes no difference whether the information comes from a static page or a chatbot.”
  2. The business owes a duty of care. Companies must take reasonable care to ensure their representations are accurate and not misleading.
  3. Customers aren’t expected to cross-check. The airline pointed out that the correct policy was on another page. The tribunal didn’t accept that the customer should have checked one part of the website against another.
  4. The airline was liable. It was ordered to pay CA$812.02 in total.

The sum was small. The principle is not: the deployer answers for what its AI says.

What “reasonable care” looks like for AI answers

Courts and regulators will develop this over time, but the practical controls are already clear:

  • Know what your AI can say. Keep an inventory of every AI system that talks to customers and the topics it covers.
  • Ground answers in approved sources. Policy, price and eligibility answers should come from logic or documents your team has approved, not from a model’s general knowledge.
  • Keep commitments in code. Anything that commits the business (a price, a refund, an eligibility decision) should come from deterministic logic (Level 0 or a Level 3 hand-off on the ladder), so the model can phrase it but can’t invent it.
  • Test and monitor. Run a golden dataset of real customer questions (lesson 5) before launch and after every change, and review real conversations regularly.
  • Hand over when unsure. Design clear routes to a person for questions outside the approved scope.

Beyond chatbots

The same logic reaches further than customer chat. Any AI output a business presents as its own (a quote, an assessment, a recommendation, an automated decision letter) is a representation by that business. The safest systems are designed so that what reaches the customer is something the business would have said deliberately.

Try it yourself: the representation audit

Pick one customer-facing AI feature and answer:

  1. What topics can it address? Is any of them a commitment (price, policy, eligibility, timelines)?
  2. For each commitment topic, where does the answer come from: approved logic, an approved document, or the model?
  3. What happens when a question falls outside scope?

Any commitment answered by the model alone is a gap to close first.

What comes next

Liability covers what your AI tells customers. Lesson 8 covers what you claim about your AI, where regulators have already acted.

How an AI answer becomes your liability The legal reasoning in Moffatt v. Air Canada, step by step. 1: It speaks for you — the chatbot is part, of your website. 2: A customer relies — reasonably, on, what it said. 3: It was wrong — the answer did not, match the policy. 4: Duty of care — reasonable care that, answers are accurate. 5: You pay — the deployer, not, the software. IOANYT ACADEMY / LESSON 7 How an AI answer becomes your liability The legal reasoning in Moffatt v. Air Canada, step by step. 01 It speaks for you the chatbot is partof your website 02 A customer relies reasonably, onwhat it said 03 It was wrong the answer did notmatch the policy 04 Duty of care reasonable care thatanswers are accurate 05 You pay the deployer, notthe software the tribunal treated the chatbot as part of the company’s website

Key takeaways

  • The tribunal treated the chatbot as part of Air Canada's website, not as a separate entity.
  • Businesses must take reasonable care that what their AI tells customers is accurate and not misleading.
  • Customers aren't expected to cross-check a chatbot against other pages of the same website.
  • The strongest control is design: commitments come from approved logic, not from a model's improvisation.

Check yourself

Pick an answer, then open the card to compare.

  1. 1. What did Air Canada argue about its chatbot?

    • A.That the chatbot had never answered the customer
    • B.In effect, that the chatbot was a separate legal entity responsible for its own actions
    • C.That the customer had used the wrong website
    Show the answer

    B. In effect, that the chatbot was a separate legal entity responsible for its own actions The tribunal rejected that argument, holding that the chatbot was part of the airline's website.

  2. 2. According to the tribunal, should the customer have cross-checked the chatbot against another page?

    • A.Yes; customers must verify everything
    • B.No; there was no reason for him to know one part of the website was accurate and another not
    • C.Only if the chatbot said so
    Show the answer

    B. No; there was no reason for him to know one part of the website was accurate and another not The tribunal rejected the idea that customers must cross-reference different parts of the same site.

  3. 3. Which control best prevents an Air Canada-style failure?

    • A.A longer disclaimer under the chat window
    • B.Answers about policy, price and eligibility coming from approved logic and sources, not model improvisation
    • C.A more creative model
    Show the answer

    B. Answers about policy, price and eligibility coming from approved logic and sources, not model improvisation Design beats disclaimers: if commitments come from logic the business approved, the model can't invent a policy.

Common questions

Is this decision binding outside Canada?

It's a decision of British Columbia's Civil Resolution Tribunal, a small-claims body, so it isn't binding elsewhere. It's widely cited because its reasoning is simple and likely to be persuasive. Take advice for your own jurisdiction.

Would a disclaimer have helped?

Disclaimers have limits, especially when the system speaks for the business on its own website. Designing so the AI can't make commitments it isn't authorised to make is a far stronger position.

Does this apply to internal AI tools too?

The customer-facing principle is about representations to customers. Internal tools carry their own risks, such as staff acting on wrong answers. The same design principle, keeping commitments in approved logic, helps in both places.