Solutions/Fake Receipt Detector

Fake Receipt Fraud Detection

A receipt can be rewritten in seconds and still look perfect. Sightova checks receipts for AI watermarks, content credentials and pixel-level edits, then shows exactly which region of the receipt looks edited.

FORGERY HEATMAP
SYNTHID CHECK
C2PA + EXIF
sightova.com/dashboard/document-tampering
Photo of a restaurant receipt whose CHANGE amount was edited from 13,700 to 16,700
  1. SynthID watermark

    Google Gemini · OpenAI GPT-Image

    Checking
  2. Content credentials

    C2PA manifest · EXIF metadata

    Pending
  3. Document forensics model

    Pixel-level tampering analysis

    Pending
Analyzing receipt…
// REAL EXAMPLE

One number changed. Nothing else did.

This is the receipt from the scan above, a real restaurant receipt photographed on a phone. We asked an open-source AI image editor to change a single amount. Switch between the versions and try to spot it before you look at the heatmap.

Close-up of the original receipt: change 13,700Close-up of the edited receipt: change 16,700
edited copy

One amount rewritten. Can you see where?

Same photo, same paper grain, same blur and the same printer font. An AI image editor rewrote a single number in about ten seconds.

Check the maths

TOTAL
36,300
CASH
50,000
CHANGE · printed
16,700
CHANGE · should be
13,700

This forger forgot the arithmetic. A careful one won't, and then the numbers all add up. The pixels still give the edit away, which is why Sightova looks at the pixels.

Model output · same photo, twice

Original photo3%

Authentic

Edited copy98%

Tampering detected · 0.9% of the image flagged

Forgery probability from Sightova's document forensics model on these exact files.

Receipt photo from the CORD v2 dataset (NAVER Clova AI), licensed CC BY 4.0. Merchant details are blurred in the original dataset. We cropped the photo and edited the change amount with an open-source AI image editor for this demonstration.

// HOW RECEIPTS GET FAKED

Three kinds of fake, and what gives each one away

Each receipt goes through the same three checks: an invisible-watermark scan, a read of its content credentials and metadata, and a pixel-level forensics model. Each kind of fake trips a different one.

The lower half of the edited receipt with the forgery heatmap over the CHANGE line

01 · Edited original

A real receipt with one number changed

The most common fake: a genuine photo or scan with a total, date or amount rewritten in an editor or an AI inpainting tool. Everything else on the receipt is authentic.

SynthID watermarkClear
Document forensics98%
Tampering Detected

The real example from above: the heatmap marks the rewritten line.

BLUE HARBOR BISTRO

14 Quay Street · Table 6

Seafood linguine24.50
Grilled sea bass29.00
Sparkling water4.50
SUBTOTAL58.00
VAT 20%11.60
TOTAL69.60

Thank you for dining with us

SynthID · Gemini

02 · Generated from a prompt

A receipt that never existed

Image models now render convincing receipts from one sentence, down to the thermal print and the crumpled paper. Google's and OpenAI's models leave an invisible watermark in the pixels, and we check for both.

SynthID watermarkFound
Content credentialsAI declared
AI-Generated Document

Illustration. Merchant and amounts are fictional.

NORTHGATE HOTEL

Folio 20418 · 2 nights

Room 2 x 145.00290.00
Breakfast36.00
Minibar184.00
TOTAL DUE510.00

Card ending 0000 · approved

C2PA: edited in Photoshop

03 · Re-touched in an editor

An honest receipt, a dishonest total

Someone opens a real receipt in Photoshop, bumps the hotel bill and saves it again. Editors that write Content Credentials record that edit, and the pixel model shows where it happened.

Content credentialsEdited
Document forensicsRegion flagged
Tampering Detected

Illustration. Merchant and amounts are fictional.

// DOCUMENT-FORENSICS API

A receipt's pixels tell the story

The same scan is one API call. Send the receipt image and get back a verdict your expense, accounts payable or claims workflow can act on, with the evidence to show a reviewer why.

  • Verdict, forgery probability and the share of the document flagged
  • Optional 256 × 256 heatmap to draw the edited region in your own review screen
  • Content Credentials (C2PA) checked in the same request
  • One call per receipt, before the reimbursement or payout is approved
Read the API docs
POST /api/v1/detect/document-tampering
{
  "tampering": {
    "is_tampered": true,
    "verdict": "tampered",
    "forged_probability": 0.9761,
    "max_confidence": 0.9996,
    "tampered_percent": 0.9,
    "image_size": [864, 1296],
    "heatmap": [[0, 0, 0, …], …]  // 256 × 256
  }
}

The response for the edited receipt above, trimmed to the tampering block.

Receipt Fraud in the AI Era: Why Traditional Expense Auditing Is Failing

Expense fraud has always been a persistent drain on corporate finances, but the tools available to bad actors have undergone a radical transformation. Where fabricated receipts once required clumsy Photoshop edits that a trained auditor could spot at a glance, generative AI now produces documents with authentic thermal paper textures, accurate merchant formatting, and mathematically consistent line items. For organizations processing thousands of reimbursement claims per month, an ai image detector that screens receipt images for edits adds a strong extra check.

From Photoshop to Prompt: A New Class of Fake

Five years ago, a fabricated receipt meant someone with intermediate Photoshop skills spent twenty minutes cloning a template and adjusting the totals. Today, a single text prompt can generate a pixel perfect restaurant receipt, complete with appropriate tax calculations, a valid merchant ID format, and realistic thermal printer artifacts. The quality gap between authentic and fabricated documents has collapsed almost entirely.

This shift means that manual spot checks, the backbone of traditional expense auditing, are effectively obsolete for catching sophisticated fakes. The human eye cannot reliably distinguish a generative model's output from a genuine receipt when the model has been trained on millions of real examples. Organizations need forensic tools that operate at the pixel level, not the glance level.

The Real Cost of Looking the Other Way

The Association of Certified Fraud Examiners estimates that organizations lose 5% of annual revenue to fraud, with expense reimbursement schemes ranking among the most common. For a company with $200 million in revenue, that translates to $10 million in annual leakage. Most of it goes undetected because the per claim amounts are small enough to avoid manual review thresholds.

The ripple effects extend beyond direct financial loss. Lenders evaluating small business applicants increasingly rely on submitted financial documents to verify revenue, a process that AI generated receipts and invoices can systematically exploit. Alternative lending fraud detection is one area where this risk shows up. Meanwhile, accounts payable departments face growing pressure to show that fraud controls are in place.

When a Text Prompt Produces a Perfect Invoice

Fraud researchers expect AI generated financial documents to account for a growing share of expense fraud attempts, putting pressure on the manual processes most organizations still depend on. The same convergence of large language models and image generators will fuel fabricated claims against insurance fraud detection systems, where fabricated repair invoices and medical bills carry even higher stakes.

Pixel Level Forensics at the Moment of Submission

Sightova treats each submitted receipt as a forensic evidence sample. Each receipt goes through three checks: a scan for invisible watermarks (Google SynthID, OpenAI), a read of C2PA content credentials and EXIF metadata, and a pixel-level forensics model that flags regions that look edited. Metadata can name the software that last saved the file, so a reviewer can see whether a receipt came straight from a camera.

The pixel-level model looks for traces left where a receipt was edited or generated, such as a changed amount. Because it scores the image itself, it can flag an edit even when the numbers still add up. You set the threshold and decide what happens to each flagged receipt.

Scanning Before Reimbursement

Sightova is a REST API with an API key. Your developers can call it from your expense, accounts payable, or ERP workflow when a receipt is submitted, and use the returned verdict and score to decide whether the receipt goes to a reviewer before reimbursement is approved. Results come back in seconds, images are hosted in Germany and deleted after analysis, and they are not used for training. Sightova is tested on 250k+ benchmark images and 50+ generative models, with results published in our testing methodology. For organizations already using AI driven controls in their banking fraud detection stack, adding receipt screening is one more signal against fabricated expense claims, inflated invoices, and synthetic purchase proofs.

Flag Suspect Receipts

Scan receipts before a dollar leaves the company, and give reviewers a heatmap instead of a hunch.