Was this made by AI? What detection tools can and cannot prove for a hotel

An AI detector is software that estimates whether a text or an image was machine-made from its style, and it returns a probability, not proof. For a hotel it is a prompt for a second look. What settles a review, a CV or a complaint photo is the record behind it: the reservation, the original file, the room, the date.

A guest sends a photo of a stained curtain with a refund request. An applicant sends a cover letter that reads like a template. A review appears that describes a room layout you do not have. In each case someone opens a detector, pastes the text or the image, and gets a number. This article explains what that number actually measures, why it fails most often on exactly the people hotels deal with, and what to check instead, so the next suspicious file leads to a decision you can defend.

Where does a hotel actually meet an AI detector?#

In four places: a review you suspect was generated, a CV or cover letter, a photo sent with a complaint or refund claim, and paperwork from a supplier or agency. In all four the detector is being asked a question it cannot answer.

On rating sites, a run of similar reviews, or a review with no stay details, makes someone want a verdict fast. In hiring, CVs and cover letters, especially for roles where writing is not the job, arrive polished and templated. In complaints, a damaged room, a stained towel or a broken fixture arrives attached to a request for money back. And in procurement, references, contracts and invoices sometimes look too clean.

The common shape is this: someone hands you a file and asks "is this real?", and the honest answer begins with "real in what sense?". Whether a file was machine-made is a different question from whether the claim in it is true. A guest can use an assistant to draft an honest complaint, and a fraudster can write a false one by hand.

The decision rule: name the decision you are about to make, refund, reject, hire, escalate, before you open any tool. The tool cannot answer that decision.

How does an AI detector work, and why is the score a guess?#

A text detector compares writing against patterns it learned from labelled human and machine text and returns a probability. It does not find a mark that says "machine".

In plain words, it measures how predictable the word choices are, and how evenly that predictability runs through the text. Human writing tends to vary more; machine writing tends to stay level. The output is a number dressed as a percentage. It is the model's estimate of style, not a measurement of origin, and the same text can score differently on two tools or after one edit.

There is no hidden signature unless a tool deliberately embedded an invisible signal when the text was generated and you can read that signal. In the ordinary case, nothing in the file records who wrote it.

Image detectors are a different family. They look for generator fingerprints, upsampling artifacts and impossible physics. A screenshot, a crop, a re-save or a messaging app's compression destroys most of that evidence before the file reaches you. By the time a photo has passed through a chat app, the detector is reading damage, not origin.

What that means at a front desk or in an HR office: a tool that misses most machine text and still accuses honest writers is a hint generator, not a judge. The one published test cited here came from the tool makers themselves, and the tool is no longer available.

Why do honest guests and honest applicants get flagged?#

Because detectors measure style, and style is shaped by language, templates, translation and length, not by who wrote the words.

Second-language writing is the clearest case. A guest writing in English uses a narrower, more predictable vocabulary and fewer idioms. To a detector that reads as machine-like. Formal and templated writing behaves the same way: a CV built from an online template, a complaint written in the register of a consumer-rights forum, a letter an accountant drafted.

Translation matters too. Text that passed through a translation tool carries machine-shaped sentences even when a person wrote the original, which is common in international guest correspondence. Short text is the fourth trap: a two-line review gives the detector almost nothing to measure, and it still prints a confident percentage. And editing blurs everything: a human-edited machine draft and a machine-edited human draft both land in the middle of the scale, exactly where the tool is least useful.

The asymmetry is what matters. A false positive costs a guest an accusation or an applicant a job. A false negative costs you nothing you can see. The tool's errors do not fall evenly on the two sides.

If the question is a photo, what can you check instead?#

Check the record, the room and the date, in that order. A photo on its own proves almost nothing; the reservation record behind it proves a great deal.

Start with the record: was there a stay, in that name, on those dates, in that room type? A photo attached to a claim with no matching reservation is a different conversation from one that matches. Then ask for the original file, as it left the camera or phone, not a screenshot or a forwarded copy. The file's embedded capture data (EXIF: date, device, sometimes location) can be edited, so it supports a story rather than proving one.

Then check the room itself. Compare the photo against your own pictures of that room number, the furniture, the fixtures, the view. Ask which room and which date; a mismatch is a fact you can state plainly. Ask for a second angle, or a picture with something in frame that places it in time, such as the key card or that day's newspaper. A genuine complainant usually has more than one picture.

Date logic settles more than software. A photo taken at 09:00 on departure day describes a room the guest slept in; a photo taken after check-out describes a room housekeeping has already entered. And watch the pattern: several claims from the same source, the same framing, the same wording, arriving close together. For money requests that arrive with documents attached, the wider routine is in our guide to verifying payment and bank-detail requests.

If the question is a review or a CV, what can you check instead?#

For a review the question is whether there was a stay; for a CV the question is whether the claims can be checked. A detector answers neither.

For a review, match the name and date against the reservation record, and use the platform's own booking check where it offers one. Look for details only a guest would know: the room number, the layout of the breakfast room, the name of a staff member. Look just as carefully for their absence in a review that otherwise describes everything. A review with no stay behind it is a platform matter, handled under the platform's rules and evidence requirements, not under a detector's percentage. And never answer publicly with "this was written by AI". Answer the content, or do not answer.

For a CV, verify the claim, not the prose. Ask for the certificate, call the reference, and ask a question that requires the experience, such as walking you through a night audit they closed. A short written task done live, or a structured interview, tells you more about capability than any score, and it is defensible if the candidate asks why they were rejected.

For the platform side, see what each platform removes and how to report a review. The one legitimate use of a detector is as a prompt for a second look, never as the stated reason for a decision.

What do you write to the guest, and what do you never write?#

Write what you can verify and what you need next. Never write that a detector says the guest is lying.

A reply that holds up has four parts. Acknowledge the claim. State the record: "we have your reservation for 12 to 15 August, room 412". Ask for the one specific thing you need, the original photo file or a second picture of the same corner. Give a date by which you will answer.

Never write a percentage. "Our AI detector shows your photo is 98% AI-generated" cannot be proved in the conversation, escalates it, and hands the guest a story about being accused by software. Never write an accusation of fraud either. Describe the discrepancy, not the motive: what the record shows, what the photo shows, what does not match.

Hold a refund decision to the same evidence standard as a card dispute: the reservation record, the registration, the photos, the correspondence. A score is not part of that file. If you draft the reply with an AI assistant, proofread it before it goes out; a templated reply reads as indifference to a guest who is already angry. Follow a proofreading checklist for hotel messages drafted with AI, and record the decision and the evidence internally so a second person can review it and the same guest's next claim meets the same standard.

What to do before the next suspicious file arrives#

Write down three checks and one sentence. The checks: the record behind the claim, the original file rather than a copy, and a second source that places the photo, the review or the CV claim in the real world. The sentence is the one you will not write, that a tool says the guest is lying. A detector can tell you where to look; only the record can tell you what happened.