AI, Explained Plainly

If your chatbot gets a policy wrong, it's still your policy

What happened to Air Canada

On February 14, 2024, a tribunal in British Columbia ordered Air Canada to pay a customer $812.02 because of something its website chatbot said.

In November 2022, after their grandmother died, Jake Moffatt was looking at flights and asked the chatbot on Air Canada's website about bereavement fares, a discounted fare for people traveling because a family member died. The chatbot said Moffatt could book at the regular price and apply for the discount afterward, within 90 days. Moffatt booked two flights and applied well inside that window. Air Canada then said it didn't accept bereavement requests after the fact, which matched a separate page on its own website.

Air Canada argued it couldn't be held liable for what the chatbot said. The tribunal member, Christopher Rivers, wrote that this amounted to treating the chatbot as a separate legal entity responsible for its own actions, and called it "a remarkable submission." He found that the chatbot was part of Air Canada's website, that the airline was responsible for everything on that website, and that it made no difference whether the information came from "a static page or a chatbot." The award was $650.88 in damages, $36.14 in interest and $125 in tribunal fees.

The decision doesn't say what kind of software the chatbot ran on. For the tribunal, that didn't matter.

Why a chatbot says something false with confidence

In September 2025, researchers from OpenAI and Georgia Tech posted a paper that starts with a test. They asked a large language model for one of the authors' birthdays and told it to answer only if it knew. On three tries it gave three different dates, and all three were wrong.

Their explanation has two parts. First, a language model learns from text, and some facts show up in that text only once, the way an ordinary person's birthday might appear in a single obituary. A model can't learn a fact like that reliably, so it produces plausible guesses for it. They estimate that if 20 percent of birthday facts appear exactly once in the training text, a model fresh from its first stage of training will get at least 20 percent of birthdays wrong. They also note that models rarely miss facts that appear constantly, like Einstein's birthday.

Second, models are scored on tests that give a point for a right answer and nothing for "I don't know." Under that scoring, a guess always beats admitting uncertainty, the same way guessing beats leaving a multiple-choice question blank. So the models get trained toward confident guesses. A made-up answer delivered with confidence is what researchers call a hallucination.

Where this stops applying

The Air Canada case is one small-claims decision from one Canadian province. It shows how one tribunal reasoned, not what a court in your state would decide. The 2025 paper is a preprint, meaning it hasn't been through peer review, and it's a mathematical argument about how models are trained, not a test of business chatbots.

What to do with it

If a model picked up your return window or your hours from its training text at all, it picked them up from a handful of pages, which puts them much closer to the obituary birthday than to Einstein's. Before a chatbot answers your customers, ask it the ten questions customers ask you most and compare every answer against your written policy. Moffatt saved a screenshot of the chatbot's answer and sent it to Air Canada, and that screenshot became part of the evidence.

Next step

Checking an AI tool's answer against what you already know is a skill you can practice on your own content. The Audience Psychology Snapshot gives a limited number of free reads: you paste a bio or caption, and it tells you who that language attracts. You're the one who knows your audience, so you can judge whether its read is right. It's at datapsych.net/audience-snapshot.

Sources