Image models became good at rendering text in 2025, and expense fraud followed. A convincing restaurant receipt, crumpled and photographed on a table, now takes one prompt and a few seconds. Expense platforms have reported sharp increases in AI-generated receipts, and finance teams that relied on "does it look real?" have lost their main defence.
The good news: fake receipts still fail checks that do not depend on how the image looks.
Why looking harder no longer works
Older fakes had obvious tells: mismatched fonts, misaligned columns, numbers that did not add up. Modern generators produce believable wrinkles, thermal-paper fading, shadows and realistic logos. You might catch a bad one, but training staff to spot visual artefacts is a losing race.
Checks that still work
1. Do the maths. Generators are improving, but arithmetic errors still slip through: line items that do not sum to the subtotal, tax at a rate that does not exist in that city, or a tip calculated on the wrong base.
2. Check the merchant. Does the business exist at that address? Does its real receipt look like this? A two-minute search catches invented restaurants and wrong addresses.
3. Match against the card. The strongest control is to reconcile receipts against corporate card or bank transactions. A fake receipt for a real charge is still a problem, but a receipt with no matching transaction is a red flag.
4. Look at the file, not just the picture. Metadata is not proof, but it is a clue. A "photo" with no camera information, created by an image tool, or a PDF exported minutes before submission deserves a second look. See hidden data in PDFs explained.
5. Watch for patterns. Round amounts just under approval limits, many receipts from cash-only venues, and identical layouts across "different" merchants are behaviour signals no generator hides.
6. Look for content credentials. Some generators and cameras now attach C2PA content credentials that state how an image was made. Absence proves nothing, but presence can settle a question quickly. See C2PA content credentials for documents.
Controls that reduce the incentive
- Prefer card-linked expenses where the receipt is matched automatically.
- Ask for itemised receipts rather than card slips.
- Spot-check randomly, and say that you do. Visible audits deter more than hidden ones.
- Keep submissions in one archive so duplicates and repeated layouts are easy to find. See organizing expense receipts as PDFs.
What about AI detectors?
Detection tools exist and some expense platforms build them in. They help as one signal among several, but none is reliable enough to accuse an employee on its own. False positives on genuine, low-quality phone photos are common. Treat a detector flag as a reason to ask for the original card statement, not as a verdict.
Takeaway
AI made fake receipts look perfect, so stop judging by looks. Reconcile with transactions, check the merchant and the maths, and keep an archive that makes patterns visible. For invoice-side fraud aimed at accounts payable, see invoice fraud and business email compromise.
