Air Canada’s Chatbot & Misinformation
What Happed?
- A traveller used an airline’s official website chatbot to ask about bereavement-fare rules.
- The chatbot incorrectly stated that the traveller could apply for a bereavement discount after travelling.
- The traveller followed this advice, completed the trip, and later applied for the refund.
- The airline refused, saying the policy clearly required the discount to be requested before travel.
The Dispute
- The airline argued the chatbot was “separate” and that the company wasn’t responsible for what it said.
- The tribunal disagreed, ruling that the airline was fully responsible for all information published on its website, including the chatbot.
- The traveller was awarded compensation.
Why This Case Matters:
This incident became widely reported because it highlights a growing reality:
- If your chatbot provides incorrect information, your organisation is still accountable.
The case shows that:
- AI explanations are treated the same as written website content.
- Businesses must verify that their chatbot aligns with official policy.
- The chatbot made the mistake” is not an acceptable defence.
Mixes Styles, Tones, or Instructions
- Conflicting instructions confuse the model and cause unstable behaviour.
Doesn’t Say What To Do When the Answer Isn’t Available
- Most hallucinations come from this one mistake.
- If the prompt doesn’t allow “I don’t know,” the bot will make things up.
Relies on the Model’s Default Behaviour
- The default behaviour is not tailored to your organisation.
- Without a clear prompt, the bot behaves like a general-purpose AI—not a council, tourism, or support assistant.
What Went Wrong?
This failure illustrates several common chatbot design issues:
No Boundaries
The chatbot confidently gave an answer even when it should have:
- asked for clarification
- provided a disclaimer
- or redirected to a human agent
Incorrect or Unverified Knowledge
- The response contradicted the airline’s official policy.
Incorrect or Unverified Knowledge
There was no system in place to:
- review chatbot answers
- validate policy-related replies
- update or test the bot regularly
Incorrect or Unverified Knowledge
There was no system in place to:
- review chatbot answers
- validate policy-related replies
- update or test the bot regularly
Poor Prompt Design
The chatbot wasn’t instructed to:
- avoid giving definitive answers on complex policy matters
- handle uncertainty with caution
- escalate sensitive queries
- check official sources before responding
