How to Tell if a PDF Has Been Edited: Six Things to Inspect
How to tell if a PDF has been edited using document properties, dates, producer fields, saves, signatures, and fonts, and why each can mislead.
TL;DR: A fake receipt is proof of purchase that was edited after a real transaction or generated for a purchase that never happened. AppZen platform data reported by Accounting Today shows AI-generated images rising from 0% to 70.8% of the fake receipts it flagged between March 2025 and mid-May 2026, which describes one vendor's flagged files and not overall prevalence. Visual review cannot settle a receipt, so the deciding test is a match against card data or the merchant's own records.
A fake receipt is a record of a purchase that did not happen as shown. Reviewers meet it in expense reports, insurance claims for lost or damaged goods, retail returns, benefit reimbursements, and loan files that take proof of purchase. The document fails in the same two ways in each queue.
Edited-after-creation starts from a genuine receipt and changes the amount, the date, the merchant, or the items. Generated-from-scratch describes a transaction with nothing behind it. Emburse's 2026 survey, fielded by Atomik Research, split its U.S. respondents along similar lines: 19% had used AI to fabricate a purchase, 15% to inflate the value of something they bought, and 6% to replace a lost receipt for a genuine purchase. That last group matters for policy. A recreated receipt is a false document even when the expense was real, and a written rule should say how the team treats it.
A useful baseline for what a receipt must show comes from tax substantiation. IRS Publication 463 says documentary evidence is ordinarily adequate if it shows the amount, date, place, and essential character of the expense. The same publication generally does not require documentary evidence for non-lodging expenses under $75, and an employer's policy can be stricter. Those four elements work as a presence check in any queue. A receipt that carries all four can still be fake.
Accounting Today reported that AppZen's 70.8% figure rests on 1,471 AI-generated fake receipts from 745 employees at 174 companies, claiming $148,143 in total. Those receipts averaged about $100, with a median near $32. Fakes built from older templates averaged $182. The article is careful on one point: the figure shows the share of flagged fakes that were AI-generated, and it does not measure how much fraud grew.
The same report covers Emburse's poll of 2,000 U.S. and U.K. workers, where 34% said they had used AI to generate a fake receipt for a business expense, including 40% of U.S. respondents. Both datasets come from companies that sell expense software, and the survey is self-reported, so they show direction better than prevalence.
Independent casework explains why small claims deserve attention. The ACFE's Occupational Fraud 2026 report found expense reimbursement schemes in 13% of the 2,402 cases studied, with a $36,000 median loss and a median of 18 months before detection, against 12 months across all cases. A review program that only examines claims above a dollar line will rarely see a $32 receipt. Random sampling below the line closes that gap.
A digital receipt and a photographed receipt are different evidence. An e-receipt or order confirmation exported as a PDF carries a production history: the software that wrote it, a text layer, and timestamps. A reviewer can ask whether that history fits a merchant's system. A photo or screenshot is only pixels. Camera metadata may be absent after messaging or upload, so the file often says little about its origin.
Paper receipts reach the queue as photos, and that is ordinary. Missing metadata on a phone image is not a finding. The harder consequence is that a generated receipt and a genuine photo arrive in the same container, because an image model and a camera both output pixels. What flattening discards applies here in full: an image-only file supports fewer checks, so the conclusion should carry less weight.
What survives on an image is still useful: the arithmetic, the type across lines, compression around an altered region, and whether the image appears on another claim. Where a merchant issues digital copies, a reasonable policy asks for one above a materiality line.
Three tells keep working because they test the transaction instead of the page.
Merchant identity asks whether this business exists, operates where the receipt says, and sells what the receipt lists. A business registry or the merchant's own site, found independently, can answer that. Trade names, parent companies, and payment processors make names differ legitimately, so a mismatch calls for a question and does not end the review.
Tax arithmetic asks whether the subtotal, tax, and total agree, and whether the tax is plausible for the place of sale. Exempt items, mixed-rate baskets, and rounding all produce honest variation. A total that cannot be rebuilt is a real finding. A total that rebuilds proves little, since a generated receipt can be composed so that its lines agree.
Timing against card data asks whether the receipt's date, amount, and merchant line up with a card authorization. Posting dates lag, and tips can change a restaurant total after the receipt prints, so small differences have innocent explanations. A receipt with no matching transaction at all is the strongest of the three signals.
The receipt's story is that this merchant recorded this sale, for this amount, at this time. Six classes of signal test whether the file supports it.
Production method asks how the file was made: a merchant export, a later edit of that export, or an image with no document underneath. Timestamp trail asks whether creation and modification fit the claimed purchase date, and a file created days after the claimed sale needs a reason. Issuer matching asks whether the file resembles what this merchant's systems would produce.
Consistency tests the numbers and the type against each other. Model artifacts are the rendering traces that generation leaves in an image. Recycled patterns look across files, because one reviewer holding one receipt will never see its twin in another team's queue. Serial reuse detection needs that wider view. On a photo, the first two signals may have little to read.
File forensics can show that a document was altered. No inspection of a file can prove a purchase took place. That proof comes from a record the submitter did not produce.
A card feed match is the most direct source. A corporate card program gives the reviewer the transaction without asking the employee for anything. On a personal card, a statement line helps, though the statement is itself a document under test. Cash purchases have no feed, so they lean on merchant records and deserve closer sampling.
Merchant records answer the question directly. A returns desk can look up its own point-of-sale system. An insurer can contact the merchant with the transaction number, using contact details the insurer found independently and not those printed on the receipt. A lender financing equipment can confirm the sale with the vendor, the same discipline described in how to spot a fake invoice. Small merchants may hold no searchable record, and a failed lookup alone proves nothing.
Each confirmation has limits. A card match shows payment. It cannot show a business purpose, and it does not reveal a later refund. In an insurance claim, a confirmed purchase establishes that the item was bought and leaves the loss itself to be investigated.
Hold the claim before paying it. Redo the arithmetic, then check the merchant and the card record. If the claim rests on an image, ask for the digital original or a statement line. Treat a request for more evidence as routine, because honest claimants lose paper receipts.
Record which class the concern falls into, edited or generated, and what evidence supports it. When the numbers agree and the file still looks wrong, check the file for a scored read of its production history, then return to the card feed or the merchant for the outside answer. Route confirmed cases to HR, the SIU, or the fraud team under the existing policy, and apply that policy the same way to every submitter.
Sphinx examines the file and shows a reviewer why it scored the way it did, across the six checks above. Sphinx's published figures are a 94.3% correct verdict and under 28 seconds to clear clean files, at $0.45 per document with no seats or platform fee. Sphinx does not connect to bank accounts, so the transaction match stays with the team's expense or claims system, and the team keeps final approval. The free playground covers 50 documents with no card.
Rebuild the subtotal, tax, and total, confirm the merchant exists and sells the item, and match the date and amount to a card transaction or the merchant's own record. Print quality and fonts are weak signals, since genuine receipts fade and competent fakes look clean.
Often, though certainty from the image alone is rare. Model artifacts and reuse across claims can flag a generated receipt, while an image-only file offers less to examine than a merchant's original PDF. A receipt with no matching card transaction or merchant record is stronger evidence than any visual cue.
Many expense policies accept photos, and paper receipts usually arrive that way. A photo supports fewer forensic checks than a digital original, so higher-value claims benefit from a card match or a digital copy from the merchant. IRS Publication 463 treats documentary evidence as ordinarily adequate when it shows amount, date, place, and essential character.
A card match shows that a payment of that amount reached that merchant. It does not show what was bought, whether the purchase had a business purpose, or whether it was later refunded. The receipt can still be edited on line items even when the total matches the card.