When your chatbot promises the wrong thing, who pays?

The case

In November 2022, Jake Moffatt's grandmother died. He needed to fly to her funeral and asked Air Canada's website about bereavement fares. The airline's chatbot told him to buy a full-price ticket and apply for a bereavement discount afterwards. He did. Air Canada later refused the discount, saying its policy required him to apply before travel. Moffatt took the airline to British Columbia's Civil Resolution Tribunal, with a screenshot of what the chatbot had told him ^[1].

Think of it like this: if the person at the ticket desk gives you the wrong instructions, the airline doesn't get to say the desk clerk was freelancing. The chatbot worked the same desk.

The airline argued it couldn't be held responsible for information provided by "agents" such as chatbots, suggesting the bot was in effect a separate legal entity responsible for its own actions. The tribunal called that remarkable and didn't accept it. Everything on Air Canada's website is the airline's responsibility, whether it comes from a static page or a bot ^[1]. An airline representative had already admitted the chatbot's wording was misleading, and Air Canada offered no evidence to contradict the screenshot. The tribunal took that to mean such evidence wouldn't have helped the airline's case ^[1]. Air Canada had to pay.

What the ruling says

The tribunal's decision, published in February 2024, is short and clear. The airline is responsible for all the information on its website. It doesn't matter whether a page is written by hand or generated by an AI. The customer relied on what the chatbot said, the chatbot's answer was wrong, and Air Canada had already agreed the wording was misleading ^[1].

The interesting part is the evidence. Moffatt had a screenshot. Air Canada had nothing that contradicted it, so the tribunal assumed the screenshot was accurate. That assumption decided the case.

Article 50 and the duty to inform

Article 50 of the AI Act sets transparency rules for AI that interacts with people or generates content. Since 2 August 2026, providers must ensure that people are informed when they are interacting with an AI system, unless it is obvious from the context ^[2]. The rule is meant to let people know they are not talking to a human, so they can decide how much weight to give the answers.

The article also requires AI-generated content to be marked in a machine-readable way, and deployers must disclose when they publish deepfakes or AI-generated text meant to inform the public, unless a human has reviewed it and takes responsibility for it ^[2]. The penalty for failing to inform or mark properly can be up to 15 million euro or 3 per cent of worldwide annual turnover, whichever is higher; for small and medium enterprises the lower amount applies ^[2].

Article 50 doesn't say how to keep a record of what the AI said. It says people have to be told, not that you have to prove it afterwards. The proof question is a different matter.

The evidence gap

In the Air Canada case, the customer kept a screenshot and the airline had no record of its own. That is the pattern in most chatbot disputes: the customer has something, and the company does not. A screenshot can be edited, but if nobody challenges it and the other side has nothing to show, the screenshot wins.

The question for anyone running a chatbot is straightforward: if a customer turns up six months later with a screenshot of something your bot supposedly promised, what have you got? If the answer is nothing, you are asking the customer, a regulator or a court to take your word that the screenshot is wrong. That is not a strong position.

Logs of what the chatbot actually said are the starting point, but logs kept by the organisation that runs the bot are easier to challenge than an independent record. If you can edit the log, someone else can argue you did.

How a digital witness would work here

In a case like this, the thing to seal is each answer the chatbot gives, at the moment it is sent. The content stays on your servers; what gets sealed is a fingerprint. If a customer later disputes what the bot said, you share the answer and the witness ID. Anyone can check that the fingerprint of the answer you are showing matches the one sealed on that date. Change a single character and it won't match.

Who can check: customer service, the customer, an ombudsman or a court. What it shows: that the record of what the chatbot said, and when, hasn't changed since it was sealed. What it doesn't do: it doesn't show that the answer was correct, or settle whether the company is legally bound by what the bot promised. It just gives both sides the same factual starting point.

Your system sends the fingerprint, not the answer itself, to WitniumChain through the REST API, the SDK or the MCP server, and gets a witness ID back. To check later, you or anyone you share the record with calculates the fingerprint again and compares it. The first 2,000 witnesses a month are free. Read the docs.

How it would work

1. Customer asks the chatbot a question.
2. Chatbot generates an answer and your system seals a fingerprint of it, with the timestamp, before sending the answer.
3. Six months later the customer disputes what the bot said. You share the sealed answer and the witness ID. They check the fingerprint matches.

What this means for you

If you run a customer-facing chatbot, here are three questions worth asking this week:

• Could we show what our chatbot said on a specific day, to a specific customer, if they turned up with a screenshot?
• Who keeps that record, and could that person edit it?
• If Article 50 already requires us to tell people they are talking to an AI, have we actually done that, and can we show we did?

The Air Canada case is a useful reminder that the bot speaks for the company. Its promises are your promises, and in a dispute the side without records starts on the back foot. Sealing answers as they happen is one way to keep your own evidence, independent of anyone's word.

What to watch

Article 50's transparency requirements have applied since 2 August 2026. Penalties for failing to inform people they are talking to an AI, or for not marking generated content properly, can reach 15 million euro or 3 per cent of global turnover ^[2]. National supervisory authorities are beginning enforcement.

For high-risk AI systems under Annex III of the AI Act, logging requirements in Article 12 apply from 2 December 2027. Those logs must record events over the system's lifetime, though the article does not specify how to protect them ^[3]. If your chatbot makes decisions that could be questioned later, it is worth deciding now how you will keep those logs trustworthy.