In November 2022, Jake Moffatt's grandmother died. He went to Air Canada's website to book a flight to the funeral, and while he was there he asked the chatbot about bereavement fares.
The bot told him to book his ticket and apply for the reduced fare within ninety days. That wasn't true. Air Canada's bereavement policy doesn't apply to travel you've already taken. When Moffatt submitted his claim, the airline refused it.
Here's the detail worth noticing. In the bot's answer, the words "bereavement fares" were a hyperlink. It pointed at Air Canada's own "Bereavement travel" page, which set out the policy correctly.
The bot cited the right source, but told him the opposite of what was said.
Chatbots get things wrong. That's a known property of the technology, not a scandal. The story here is what Air Canada reached for when it happened, and why that defence will fail you too.
The truth was one click away
Air Canada refused the claim, then defended the refusal. Its own representative had already admitted in writing that the chatbot used "misleading words", and pointed Moffatt back at the link as if that settled it.
Moffatt took it to the British Columbia Civil Resolution Tribunal, an online small claims body that decides on written submissions. In February 2024 he won $650.88 plus interest and fees. Not a trial, not a landmark award. The reasoning is what matters.
Air Canada argued that the correct information was on its website, and Moffatt should have found it. Tribunal member Christopher Rivers wasn't persuaded:
"It does not explain why the webpage titled 'Bereavement travel' was inherently more trustworthy than its chatbot. It also does not explain why customers should have to double-check information found in one part of its website on another part of its website."
Read that twice, because it's the whole argument in two sentences. Air Canada was asking a customer to know, without being told, that the static page outranks the conversational one. Nobody does that. There's no signal available to a customer that would even let them.
Air Canada also argued, and this is the part that made the headlines, that it couldn't be held liable because the chatbot was a separate legal entity responsible for its own actions. Rivers called that "a remarkable submission" (tribunal-speak for you cannot be serious). A chatbot is still just part of the website, he wrote. It makes no difference whether the information comes from a static page or a bot.
A link only works when the interface fails
The defence Air Canada ran is the one almost everybody is planning to run. Reference your sources. Cite the page. Add a footer telling people to verify important information. It appears in responsible-AI guidance as a control, and on paper it looks like one.
The Air Canada case is that control being tested. It broke.
It broke for a structural reason. A link only helps if the customer clicks it, and the entire point of a conversational interface is to spare the customer from reading the page. That's the product. That's why it was built, why it was funded, and why it sits on the homepage instead of a search box. So the safeguard only works in the moment the interface fails at its job. When the interface works as designed, the safeguard does nothing.
A control reduces the chance of harm. This reduces the chance of blame. It sits there, unclicked, until something goes wrong, and then the organisation points at it and says: it was right there.
That's a liability transfer, and the customer is the one carrying it.
He wasn't trusting the bot
It's tempting to read this as user error. He had the link. He didn't click.
But he didn't put his faith in the bot. He put it in the Air Canada brand. The bot was just the surface the airline happened to be wearing that day, in the airline's colours, on the airline's domain, answering in the airline's name.
Trust is what customers give you. Trustworthiness is what you are. The philosopher Onora O'Neill has spent decades on this point: we can't make people trust us, we can only give them the grounds to. And customers give their trust to organisations, not to components. They don't hold separate accounts for your chatbot, your help centre and your call centre, each marked to a different standard. They hold one account, in your name, and every surface draws on it.
That's why "he should have clicked" has no answer to Rivers' question. The moment you ask a customer to verify one part of your organisation against another, you've told them something expensive: that keeping track of which bits of you are reliable is now their job.
Friction, removed at exactly the wrong moment
There's a second reading, and it runs against a decade of experience design.
Clicking that link is friction. An extra step, a page load, a paragraph of policy prose between a person and a booking. Every instinct in modern design says remove it, and Air Canada did.
But friction is a signal to be calibrated, not a quantity to be minimised. There are moments where slowing someone down is the most caring thing a system can do: when a decision is expensive, when it can't be undone, when the person making it is in no state to be careful. Moffatt's booking was all of those. He was arranging his grandmother's funeral.
Air Canada removed the friction at exactly the moment friction protects someone, then blamed him for taking the smooth path it had built for him. Strip out every bit of 'friction' and you don't get a faster journey. You get a journey with nothing in it to catch a mistake.
The error was the chatbot's. The choice wasn't.
The most useful thing about this case is that the hallucination is the least interesting part.
Air Canada had at least three chances to stop. When its representative admitted the words were misleading. When Moffatt filed his claim. When somebody sat down to draft the defence. Each time it chose the policy over the person, and the third time it chose to argue in public that its own website wasn't its responsibility.
The AI made an error. The organisation made a decision. Only one of those is a trustworthiness problem.
In the trust model we use at Sideways, Integrity acts as a gate, not a score. An organisation can be reliable, competent and warm, and none of it survives a moment where it declines to own what it did. Air Canada's chatbot cost it $650.88. Its answer to the chatbot is why you're reading about it more than two years later.
Australia's AI Ethics Principles list contestability and accountability among their eight. This case failed both, and not in the model layer. Moffatt had to go to a tribunal to contest an $800 error, and the airline's formal position was that accountability belonged to someone else.
The work sits upstream
If a link isn't a control, what is?
The honest answer is that the work happens before the customer arrives, not at the point of failure. A safeguard that activates at the moment of harm is too late by definition. By the time Moffatt read that answer, the harm had landed. He believed a wrong thing about his grandmother's funeral, and everything after that was cleanup.
Belinda Donald at Portable asks the question I'd bring to any AI deployment: who is the manager of the AI? If a person gave this answer all day, in your name, they'd have a manager. Someone would own their accuracy, review their work, and answer for it when they got it wrong. Name that person for your AI and most of the rest follows: what gets logged, what gets sampled, what gets escalated, who picks up the phone.
Air Canada's answer, submitted in writing, was that nobody manages it, because it isn't ours.
You don't have to solve this at the model layer. Start small. Find the three answers your AI gives most often where being wrong would cost someone something, and check them against the page you'd point at in your defence. If they disagree, you've found your work. It won't take long, and it's considerably cheaper than a tribunal.
Sideways is a behavioural design practice for regulated businesses. We make trustworthiness measurable and designable, so the behaviours that hold customer relationships together can be built into systems rather than left to good intentions.