Your chatbot promised a discount: is the hotel bound by what AI tells a guest?

Chatbot liability means the business, not the software, answers for what its assistant tells a guest. In Moffatt v. Air Canada (2024 BCCRT 149, 14 February 2024), a tribunal rejected the idea that a chatbot is a separate entity. No Turkish decision was found; counsel decides, starting from the Code of Obligations and consumer law.

A guest asks your website assistant whether late checkout is free. It answers yes. She books a 4 p.m. flight, arrives, and reception asks for a fee. Nobody at the desk wrote that sentence, and the guest does not care who did. This article walks through the one well-documented case on chatbot liability, the Turkish texts a lawyer would open, and the operational choices that shrink the question before it becomes a dispute.

What does a wrong promise from a hotel chatbot look like?#

It is almost always one of four things: a price, a policy, an inclusion, or a timing. The guest treats the answer as the hotel's word, because that is what it is.

A price promise is a discount, a returning-guest rate, a child-free stay, or a rate that no longer exists in the system. A policy promise covers cancellation windows, refund rules, pet rules, minimum age at check-in and deposit amounts. An inclusion promise states that breakfast, airport transfer, parking, spa access or beach service is included when it is an extra. A timing promise covers early check-in, late checkout, luggage storage or a room type change on arrival.

Before any legal question, establish three facts: the exact words, the date and time, and whether the guest saw the answer before booking or only before arriving. The first operational step is always the same: pull the transcript and the version of the knowledge source the assistant was answering from that day. If you are still deciding what the assistant may discuss at all, start with scope, knowledge and handover for a hotel website chatbot.

What happened in Moffatt v. Air Canada?#

A Canadian tribunal held the airline responsible for a wrong answer its own website chatbot gave about a bereavement fare, and rejected the argument that a chatbot is something the airline cannot answer for (British Columbia Civil Resolution Tribunal, Moffatt v. Air Canada, 2024 BCCRT 149, 14 February 2024 (opens in a new tab)).

In November 2022, after a grandparent's death, Jake Moffatt used a chatbot on Air Canada's website. It said a traveler who had already flown could apply for a reduced bereavement rate within 90 days of the ticket's issue date. The airline's linked "Bereavement travel" page said the opposite: the policy did not apply after travel was completed. Moffatt bought two one-way tickets at the regular fare, then applied, and the airline refused, although an employee admitted in the exchange that the chatbot had given "misleading words".

The tribunal treated the claim as negligent misrepresentation, found a duty of care, and held that the airline had not taken reasonable care to keep the chatbot accurate. It called the argument that the airline could not be liable for a chatbot's information a "remarkable submission", and said it makes no difference whether information comes from a static page or a chatbot: both are part of the airline's website. It found Moffatt's reliance reasonable and said customers should not have to check one part of a website against another. The airline's tariff defense failed because it did not produce the tariff text.

The order was CAD 812.02 within 14 days: CAD 650.88 in damages for the fare difference, CAD 36.14 in pre-judgment interest and CAD 125 in tribunal fees.

Does a Canadian tribunal decision bind a hotel in Türkiye?#

No. It is a small-claims decision from British Columbia, decided on 14 February 2024 by Tribunal Member Christopher C. Rivers, and it binds nobody outside that dispute. No Turkish court decision on this question was found.

The British Columbia Civil Resolution Tribunal is a small-claims body, not a court whose reasoning travels to Turkish disputes. What does travel is the shape of the reasoning: a business answers for the words on its own website, including the words a machine generated there. What does not travel is the remedy, the procedure, the standard of care and the amounts, all of which belong to that jurisdiction and that dispute.

The absence of a Turkish decision is not the absence of a rule. It means the answer comes from general law applied to your facts, which is why the next section matters.

Read the case for one practical lesson: the defense that failed was not about technology, it was about a business declining to own its own website.

Which Turkish texts would a lawyer start from?#

Two families of rules: the Code of Obligations, which asks whether the hotel made a binding declaration and whether it is at fault, and consumer law, which asks whether the exchange was a misleading commercial practice.

The Code of Obligations#

The Turkish Code of Obligations (Law No. 6098) starts with how a contract forms (consolidated text, mevzuat.gov.tr, 7 November 2026 (opens in a new tab)). Article 1 says a contract forms when the parties declare their wills to each other, mutually and consistently, and a declaration may be express or implied. A chat answer is a candidate declaration, not automatically a contract term.

Article 4(2) says an offer made by telephone, computer or similar means counts as made between persons who are present, which changes the timing rules a lawyer applies. Article 8 says an offer does not bind if the offeror expressly reserved that right, or if the nature of the matter shows no intention to be bound; displayed prices and price lists count as offers unless the contrary is clearly and easily understood.

On fault, Article 112 says a debtor who does not perform, or performs badly, must compensate the loss unless it proves it bears no fault. Article 115 makes an advance agreement excluding liability for gross fault absolutely void. Article 116 makes a debtor answer for auxiliaries such as employees carrying out the work; that liability can be excluded by prior agreement, but such an agreement is absolutely void where the service requires expertise and may be carried out only with a permit from the law or the authorities. Whether a software assistant is an auxiliary, and whether that exception applies to your operation, is a question for counsel.

Consumer protection and advertising rules#

Consumer law pulls in a second direction (Law No. 6502 and the advertising regulation, consolidated texts, mevzuat.gov.tr, 7 November 2026 (opens in a new tab)). Under Law No. 6502, advertisements must be truthful and honest (Article 61(2)); deception and exploitation of a consumer's lack of knowledge are banned (Article 61(3)); the advertiser must prove the truth of its claims (Article 61(6)); and unfair commercial practices are prohibited, with the trader bearing the burden of proving a challenged practice is not unfair (Article 62).

The Commercial Advertising and Unfair Commercial Practices Regulation adds two specific concepts. Article 29 covers misleading action, including the scope of the trader's commitments, the price and how it is calculated, even where the information is factually correct. Article 30 covers misleading omission, including hiding material information or giving it at an unsuitable time, with the limits of the medium taken into account.

Two open questions a lawyer answers, not this article: whether a chat exchange counts as an advertisement or a commercial practice at all, and whether the guest in front of you is a consumer in the sense the law uses.

Why is "the AI made a mistake" weak in front of a guest?#

Because the guest contracted with the hotel, not with a model, a vendor or an integrator, and the tribunal in the Canadian case treated exactly that separation as no defense at all.

The guest's counterparty is the business whose name is on the website and the confirmation email. Everything behind that name is the hotel's own arrangement. The tribunal's phrase, "remarkable submission", is the clearest public signal of how a decision-maker reacts when a business disowns its own website.

Staff saying "the system made a mistake" at the desk sounds like an admission and gives the guest nothing. The Air Canada employee's admission that the chatbot used "misleading words" did not help the airline. The reputational cost lands on the hotel's name; the vendor's name never appears in the guest's review. The argument also spends time: the disputed amount in the Canadian case was CAD 650.88 in damages, and the argument around it lasted months.

A better sentence at the desk names the decision, not the technology: what the hotel will do now, and by when. For the routine behind that sentence, see an incident routine for when the AI gets it wrong with a guest.

Do disclaimers protect the hotel?#

Partly, and only under conditions. A disclaimer can set the scope of what the assistant is doing; it cannot erase a promise the guest reasonably relied on, and it cannot exclude liability for gross fault.

What a disclaimer can do: say the assistant gives general information, that rates and policies are confirmed in the booking confirmation, and which page or document is authoritative. Where it stops: it must be seen before the guest relies, it cannot override mandatory consumer protection, and Article 115 of the Code of Obligations makes an advance exclusion of gross fault absolutely void.

A generic line in a footer is close to useless. The same line shown in the first message and repeated next to any price or policy answer is a different thing. Specific beats generic: naming the authoritative source ("cancellation terms are those in your confirmation") does more work than "information may be inaccurate". A disclaimer contradicted by the assistant's own confident answer is weak evidence, because the guest read the answer, not the caveat. Keep the disclaimer version with the transcript, so that a later dispute shows what the guest actually saw on that date.

For what the assistant should not be allowed to say in the first place, see guardrails for a guest-facing AI assistant.

Who pays in the end, and when should the hotel honor the promise?#

The guest's claim runs against the hotel; the hotel's recovery runs against its vendor only as far as the contract allows. So the cheaper question is decided before the dispute: which promises the hotel honors on the spot.

The vendor contract#

The vendor contract decides recovery, not liability to the guest. Look for accuracy commitments, notice of model or knowledge-base changes, indemnity, access to logs, and a duty to cooperate in a dispute. If the contract is silent on wrong answers, the hotel carries the guest-facing cost alone, which is exactly the gap the contract should close. For the clauses to ask for, see AI clauses in hotel software contracts.

A standing rule for the front desk#

A standing rule works better than case-by-case argument. Set three tiers in advance: honor immediately, escalate to a manager within a stated time, or decline.

Honor immediately where the promise is small, specific and inside what the assistant was authorized to discuss: a breakfast, a late checkout, a modest rate adjustment. Escalate where the amount is material or the promise touches a rate structure, a group block or a refund rule. Decline only where the promise was clearly outside the assistant's scope and the guest could see that, and say so in writing with the transcript attached.

Whatever the outcome, fix the source the same day so the same answer cannot be given twice, and record the decision so the next shift does not reopen it.

What to do next#

Before the next guest message arrives, do one thing: open the transcript tool and the vendor contract side by side, and write a one-page rule that says which promises the desk honors on the spot, which go to a manager, and who fixes the source afterwards. The legal question stays open until a Turkish decision or your adviser closes it; the operational question does not have to wait, and answering it well is what keeps a small wrong sentence from becoming a dispute.