How to spot fake receipts made with AI

What changed: receipts on demand

A fake receipt used to take some effort: a template website, a photo editor, a doctored PDF. Now a chatbot will produce a photo of a crumpled receipt from a named restaurant, with a fold, a coffee ring and a believable total, in seconds and for nothing.

Expense and audit teams report that AI-made fakes overtook template fakes in expense systems during 2025 and 2026. The amounts are often small, sized to slip under auto-approval limits, which makes each one easy to miss and the total expensive.

It helps to separate the three kinds of fake receipt, because each needs a different check:

• Generated from scratch: a chatbot or image model makes the whole photo, paper, print, table and all.
• Edited: a real receipt with the amount, the date, the merchant or a line item changed, by hand or with an AI edit tool.
• Reused: a real receipt claimed twice, by two people, or for a trip it doesn't belong to.

Only the first is a question about pixels. The other two are caught by records, and the records catch most generated receipts as well.

Why looking closely no longer works

Approvers were trained to spot fakes by eye: a wrong font, a blurry logo, numbers that didn't line up. Generated receipts have moved past most of that. Here is what still repays attention, and what no longer does:

The point is not that fakes are flawless. It is that their flaws are no longer in the parts people look at. A generated receipt is made to look right; it is rarely made to agree with a bank feed, a booking and the merchant's own records.

Step 1: match the receipt to the money

The strongest check sits outside the receipt. A receipt is a claim that a payment happened, and the payment leaves a record of its own.

1. 1

   Match it to the card or bank record

   For company cards, the card feed shows the merchant, the date and the amount. A receipt with no matching transaction, or with an amount that differs from it, needs an explanation before it is paid.

2. 2

   Treat personal-card and cash claims differently

   When someone paid with their own card or in cash, there is no feed to match. Ask for the statement line for larger amounts, and keep an eye on how often cash claims appear for each person.

3. 3

   Compare with the rest of the trip

   Flights, hotel bookings, mileage and calendar entries put a person in a place at a time. A dinner receipt from a city they had already left, or a taxi on a day they were on leave, is a question to ask.

4. 4

   Look for duplicates and near-duplicates

   The same total, merchant and date from two people, or one receipt submitted again with a new date. Expense tools flag exact duplicates; near-duplicates need a rule or a human eye.

Step 2: check the merchant and the maths

A generated receipt has to invent details that real businesses publish. Each one can be checked in a minute:

• The merchant exists: search the name and the address, and check the opening hours for the date and time printed on the receipt.
• The tax number: VAT, GST and other tax registration numbers follow a set format in each country, and many countries let you look one up. A number in the wrong format, or one that belongs to another business, is a strong sign.
• The phone number and address: they should match the merchant's own website or listing, not just look plausible.
• The maths: add up the line items, apply the tax rate that was in force on that date, then check the service charge, the tip and the total.
• The items: does the restaurant serve that dish, does the hotel charge for that extra, is the price in the right range for the place?
• The small print: receipt numbers that run in sequence across visits weeks apart, a till or table number that never changes, or card digits that don't match the employee's card.

Generators are good at making a receipt look right and poor at making it agree with the rest of the world. Every detail that can be checked against a public record or another document is worth more than a minute spent zooming in on the print.

Step 3: ask for the original

When something doesn't fit, ask. Most honest people can answer quickly, and most fakes don't survive the question.

• Ask for the paper receipt, or a new photo of it taken now from another angle.
• Ask for the emailed receipt or the booking confirmation the merchant sent.
• For card payments, ask for the matching line on the card or bank statement.
• For larger amounts, contact the merchant and ask them to confirm the transaction, the date and the total.

Where Perspicax fits

Perspicax combines proven computer vision techniques with TrueEye 2.3, our state-of-the-art detector. It examines every region of an image, down to the pixels, and answers in one of three bands:

What it can tell you depends on what the receipt image is:

• A photo of a receipt, on a table or held in a hand, is a photo-like image. If a chatbot generated the whole picture, that is exactly what Perspicax is built for.
• A flat scan or a PDF is a document. Perspicax still answers, but it marks the answer as a weaker signal, because it is tuned for photos.
• A real receipt with an edited amount is an edit. Most of its pixels came from a camera, so it may come back likely real or inconclusive.

Used that way, a detector is good for triage. It can raise a generated receipt photo before anyone spends time on the maths, and it costs an approver nothing when the answer is likely real. Expense and audit tools can call it on every upload through the Perspicax API and store the answer next to the claim.

Policies that stop fakes before they start

The cheapest fake to catch is the one nobody submits. A few rules make generated receipts pointless:

• Use company cards with a transaction feed wherever you can, so most claims match automatically.
• Sample small claims at random, including those under the auto-approval limit, and tell people that you do.
• Ask for the original or the merchant's emailed receipt above a set amount.
• Keep duplicate and near-duplicate checks running across people and across months.
• Have a missing-receipt route, such as a signed declaration with the statement line up to a limit, so honest people never feel pushed to invent one.
• Make it plain in your expense policy that a fabricated or altered receipt is misconduct, whatever tool made it.

Small businesses without an expense system can do the same with a spreadsheet: once a month, match every receipt to the bank statement and ask about anything that doesn't match.

Refund and insurance teams face the same problem with photos of damage. See AI photos in refund and insurance claims, and for receipts that were real but changed, how to tell if a photo was edited with AI.