In November 2022, Jake Moffatt’s grandmother died, and he went to Air Canada’s website to book a flight to the funeral. He asked the airline’s chatbot about bereavement fares. It told him he could buy his ticket now and apply for the reduced fare afterwards, within 90 days of the purchase. The words “bereavement fares” in that answer were a link to Air Canada’s own Bereavement travel page, which said the policy did not apply to requests made after travel was complete. Moffatt took the chatbot at its word, paid about C$1,630 for his flights, and applied for the discount later. Air Canada refused.
The case went to British Columbia’s Civil Resolution Tribunal. Among its defenses, Air Canada argued that it could not be held liable for information provided by its agents, servants or representatives, its chatbot included. Tribunal member Christopher Rivers read that as a claim that the chatbot was, in effect, a separate legal entity responsible for its own actions, and called it “a remarkable submission.” His answer was short: “It should be obvious to Air Canada that it is responsible for all the information on its website.” In February 2024 he ordered the airline to pay Moffatt C$650.88, the difference between what he paid and the bereavement fare, plus interest and fees.
The legal question was easy. The behavioral question underneath is a bit more complex. Air Canada’s position required the customer to know which part of its website to believe. The chatbot and the policy page were both the airline speaking, on the same site, about the same fare. Rivers made the point himself: the airline did not explain why customers should have to double-check information found in one part of its website against another part. To protect himself, Moffatt would have needed to trust the policy page and distrust the chatbot, on a question where nothing on his screen told him the two differed in reliability.
That is a harder request than it sounds, because of how trust works. In 1995, Roger Mayer, James Davis and David Schoorman proposed that trusting someone rests on three separate judgments. Ability is whether they can do the thing. Benevolence is whether they want good things for you. Integrity is whether they hold to principles you accept. Researchers who study automation adapted the model for machines. John Lee and Neville Moray recast the three judgments as performance, what a system does; process, how it does it; and purpose, why it was built. In a 2004 review, Lee and Katrina See set the goal as appropriate trust and broke it into parts. Trust should be calibrated, meaning it matches what the system can actually do. It should also be specific, separating the functions a system performs well from the ones it does not.
Specific trust is how people already treat each other. You can rely on a colleague’s numbers and not on their sense of a deadline. You can accept a doctor’s diagnosis and still get a second opinion on the surgery. Specificity is also what lets trust survive a mistake. When the colleague misses a deadline, you discount their deadlines and keep using their numbers.
A chatbot makes that hard to do. Flight status, baggage rules, refund eligibility, and a customer’s legal entitlements all arrive in the same window. A human agent gives off cues about which answers to lean on. They read one answer off a screen, hesitate on another, and say they will check with a supervisor on a third. My hypothesis is that a language model, built to produce fluent answers, gives off almost none of those cues, so the customer has nothing to calibrate against. The interface asks for one level of trust across every question, which is the general trust Lee and See warn against.
Research on how people treat algorithms shows what follows. In 2015, Berkeley Dietvorst, Joseph Simmons and Cade Massey ran five studies in which participants forecast real outcomes, such as how MBA students would perform, and chose whether to tie their bonus to a statistical model or to a human forecaster. The model was more accurate in every study. In the MBA task, human forecasts carried 15 to 29% more errors. Yet participants who had watched the model make forecasts, and therefore make mistakes, became less confident in it and less likely to choose it. Most of the participants who watched both forecasters saw the model win, and they were still among the least likely to bet on it. The authors traced this to an asymmetry: after seeing an algorithm and a person make the same mistake, people lose confidence in the algorithm faster. They called it algorithm aversion.
A second finding sits beside it. In 2019, Jennifer Logg, Julia Minson and Don Moore found across six experiments that lay people followed advice more closely when they were told it came from an algorithm than when they were told it came from a person. Read together, the two results describe trust given in general. People start out generous with a system they have not yet seen fail, and they pull their trust back from the whole system once it does. Neither pattern is calibrated, because both treat the algorithm as one thing to be trusted or not.
For a company running a chatbot, that is the cost of asking for trust in general. A customer with no way to tell the reliable answers from the unreliable ones will extend the same trust to all of them. When one goes wrong, as one did for Moffatt, the damage will not stay with that answer. My inference, which neither study tested on customer service, is that the customer stops trusting the chat window for everything, including the questions it handles well, and ultimately leaves.
The fix is to make the trust specific, and the Air Canada case reflects that. The chatbot already linked to the policy page. A response built for specific trust would have quoted that page instead of paraphrasing against it, and said in plain words what kind of answer it was giving: this is the policy, and here is where it is written; this is general guidance, and a person has to confirm it before you spend money. When a question carries money or a deadline, the answer should say so and offer a person. And the company has to stand behind what its chatbot says. That is the integrity judgment in Mayer’s model, and a customer cannot make it in a company’s favor while the company reserves the right to disown its own website.
How an AI treats a person includes whether it tells them how far to trust it. Most deployments are built to sound equally sure about everything, because even confidence reads as competence in a demo. Moffatt had nothing on his screen to tell him which answer to stake C$1,600 on, and the tribunal decided he should not have needed anything. Trust is earned in specific ways. A system that asks for it in general will lose it the same way.