AI hallucination in hotels: why assistants invent facts and how to catch them

An AI hallucination is a confident answer with nothing behind it: the assistant produces a plausible sentence about your hotel that no document, rate sheet or record supports. It happens because these systems generate likely text rather than look up facts, and it is reduced, not eliminated, by answering only from written facts and showing the source.

A guest messages at 23:40 to ask whether the pool is still open. The draft reply is warm, specific and completely wrong: it says the pool closes at midnight. The sign by the door says 19:00, and the guest is already in a towel. After reading this article you will be able to explain why this happens, predict where it will bite your hotel, and run a short weekly routine that catches most invented answers before a guest does.

What is an AI hallucination, in plain words?#

A hallucination is a fluent, confident answer that has no source behind it. The assistant did not look anything up and did not find a fact; it produced the sentence that most plausibly continues the conversation. If you want the mechanics in more depth, read why fluent language models are not automatically right.

Why fluent is not the same as right#

These systems generate text word by word, picking what is likely to come next, not by reading your rate sheet and quoting it. Fluent and correct are two different properties, and one does not imply the other. An assistant can be right most of the day and wrong exactly where money or safety sits, because the wrong answer is often the one that sounds most specific.

There is no intent. A hallucination is not a lie, and treating it as dishonesty leads hotels to the wrong remedy: blame rather than source control. The word is borrowed from perception, where seeing something that is not there has a name. In software it simply means output with no grounding.

Hallucinations are not limited to numbers. They include invented amenities, invented staff names, invented policies and invented reasons for a policy. A made-up explanation of why checkout is at 12:00 is as much a hallucination as a made-up rate.

Three failures that look like one#

Before you fix anything, separate the failures that look alike:

• An invented fact: nothing behind it anywhere in your documents.
• A stale fact: a real document that is out of date, like last season's spa price list.
• A plumbing error: the right rate copied into the wrong field, or the wrong room type attached to the right price.

Each needs a different fix. Invention calls for writing the fact down; staleness calls for a dated review; a plumbing error calls for a systems check. If you treat all three as "the AI lied", you will fix none of them.

Where does hallucination bite a hotel? Four examples#

It bites wherever a guest will hold the hotel to the sentence: opening hours, cancellation terms, transfers and child policies. Those are the four categories where an invented detail turns into a complaint, a refund or a safety problem. In each case the guest is not reading a suggestion. They are reading what they believe is the hotel's own statement, in the hotel's own voice.

The four categories that cost money#

The table below shows illustrative examples; none of them happened, they show the pattern.

The cost is not only the refund. It is the argument at the desk, the review, and the staff time spent defending a sentence nobody on the team wrote. The most expensive single sentence an assistant can produce is calling a non-refundable rate free cancellation, because it converts a policy question into a money question with a witness.

Why does it happen more when facts are missing, old or contradictory?#

Because an assistant with no fact to read still has to produce a sentence, and the likeliest sentence is the one that sounds normal for a hotel. Gaps get filled, and old or conflicting documents give it something plausible to pick from. The cleanup work that prevents this is described in cleaning up hotel data before an AI assistant reads it.

Gaps, contradictions and stale documents#

A gap is an invitation. If the pool hours are not in any document the assistant can read, it will produce typical pool hours rather than nothing. Facts that live only in a person's head, the ones a receptionist knows by heart, are invisible to the assistant and get invented instead.

Old facts do not announce themselves. A pool closed for renovation last winter stays in the text long after it reopened, and the assistant has no sense of which document is current. Long documents bury the relevant line, too: if the child policy sits on page nine of a PDF, the answer is more likely to come from the pattern than from the page.

Language matters. If the Turkish fact sheet is thinner than the English one, the Turkish answer has more room to invent, and the guest reading Turkish gets the weaker reply. Check each language against the same source document, not against the English answer.

What actually reduces hallucinations in a hotel assistant?#

Three things, in this order: answer only from the hotel's written facts, show the source next to the answer, and make "I do not know, I will ask a colleague" an acceptable reply. None of them eliminates the problem; together they shrink it and make the rest visible.

Grounding, sources and the right to refuse#

Grounding means the assistant reads the hotel's own documents and answers from them, instead of answering from general knowledge about what hotels usually do. A visible source means staff can see which document produced the sentence. If there is no document, there should be no sentence.

A permitted refusal matters just as much. An assistant that must always answer will invent, so give it an explicit handover path and count a handover as a success, not a failure. Scope limits belong to the same family: visa rules, competitor prices and medical advice are outside the assistant's ground, and saying so is part of the design, not a weakness.

Two structural rules complete the picture. A person confirms anything sensitive: money, legal terms, safety and complaints go to a human before they are sent. And the hotel keeps one canonical fact sheet, dated and owned: every fact carries the date it was confirmed and the person who confirmed it, and expires if nobody renews it. Finally, test with the questions guests actually ask, including the awkward ones, rather than the questions the hotel wishes they asked. The testing routine itself is covered in how to test a hotel AI assistant before staff or guests use it.

How do you spot-check AI drafts and automated replies?#

Check by risk category, not at random. Read the replies that touch money, safety, legal duties and access first, because those are the ones where an invented detail becomes a refund, an incident or a complaint.

Sample the risky categories every week: rates and deposits, cancellation and refunds, pool and beach safety, check-in and identity rules, transfers and parking, pets and children. For each sampled reply, ask one question: where is the document this sentence came from? If nobody can point to it, treat the reply as invented even when it happens to be right.

Read the draft as the guest who will hold you to it, not as the colleague who knows what was meant. Check languages separately, because a fact can be correct in English and wrong in Turkish, Arabic, Russian or German, and the guest reads only one of them. Watch for the confident-but-vague pattern: "yes, we can arrange that" with no detail is often a gap being papered over.

A weekly fifteen-minute review#

Keep a correction log. Every fix should end in the fact sheet, not only in the conversation, otherwise the same wrong answer returns next month. If one category produces repeated errors, move that category behind human approval and say so in the tool's settings rather than in a meeting.

A wrong answer already reached a guest. What now?#

Fix the guest first, then the fact. The guest is holding a sentence in the hotel's voice, and arguing about which system wrote it makes the hotel look worse, not the software. The handover question, when a reply should reach a person instead of the assistant in the first place, is covered in when an AI reply should be handed over to hotel staff.

Reply quickly and plainly: what the correct position is, what the hotel will do about it, and by when. Do not explain the technology to a guest who is standing in the lobby. Decide the remedy within the authority you have, and escalate rather than improvise if the promise was large.

Write down exactly what was promised. A promise made in the hotel's name may have to be honored even when no one on the team made it. Then correct the fact sheet the same day, with a date, so the same answer cannot be produced again tomorrow.

Search the message log for the same wrong answer going to other guests. One hallucination is rarely a single event. Finally, tell the whole team, not only the manager, because the next guest will ask the receptionist, not the person who fixed the document.

What to do next#

This week, write down the ten facts guests ask about most, give each one a date and an owner, and make "I do not know, I will ask a colleague" an allowed answer in every channel. Hallucination is not solved by a better model; it is managed by controlling what the assistant is allowed to read, showing where each answer came from, and checking the risky categories before a guest does.