An airline is held to its chatbot’s answer

What happened

After a death in his family, Jake Moffatt asked Air Canada’s website chatbot about bereavement fares. It told him he could buy a regular ticket and apply for the reduced fare afterward, within 90 days of purchase. Air Canada’s policy did not allow bereavement fares to be claimed after travel, and the airline refused his request.

Before the tribunal, Air Canada argued that it could not be held liable for information provided by its agents, servants, or representatives, including a chatbot. The tribunal said that this was, in effect, treating the chatbot as a separate legal entity responsible for its own actions, and rejected the argument. It found the airline responsible for all the information on its website and liable for negligent misrepresentation.

What allowed the action to take effect

The tribunal found that Air Canada “did not take reasonable care to ensure its chatbot was accurate.” The chatbot’s answer reached the customer as written. It linked to the airline’s policy page, which said otherwise, but the tribunal held that a customer should not have to check one part of a website against another.

An illustrative control

Assumptions. Statements that commit the organization, such as those about eligibility, refunds, prices, and terms, are listed in the register as actions with external effects. The chatbot’s replies on those subjects pass through an external control layer, a check that sits outside the chatbot, before the customer sees them.

rule:       policy-statements
applies to: statement.refund, statement.eligibility, statement.fare_rules
permit:     approved policy text, quoted with its version
deny:       anything else on these subjects
instead:    send the approved text and offer an agent

Under these assumptions, the wrong answer does not reach the customer, who instead receives the approved policy text, quoted with its version, and the option of speaking to a person.

Limits of this control

Deciding which statements commit the organization takes effort, and some will be treated as ordinary conversation and pass unchecked. Because the layer checks statements against the policy, it cannot correct a policy that is itself wrong or unclear. It controls only what reaches the customer and does not affect what the model generates.

Controls that apply

• P4 Reviewable policy: the approved text is the rule.
• P9 Declared scope and clause C-1: statements that commit the organization belong in the register.
• The customer finances case study applies the same rule to fee refunds.

Sources

• Moffatt v. Air Canada, 2024 BCCRT 149BC Civil Resolution Tribunal, 14 February 2024
• Report on the decisionAI Incident Database