Was this receipt made with ChatGPT or another AI?

Proof you can re-check, and opinion

Every signal in a result says what kind of evidence it is. A verifiable signal is proof you can check again without trusting us. An opinion is a judgement.

Proof

• Content Credentials (C2PA) in the file that say an AI model made the image, or that an AI tool signed it.
• Generator metadata that names an AI tool, such as ChatGPT, Gemini, Midjourney or Stable Diffusion, or that declares the image AI-made.
• Totals that don't add up: line items that don't sum to the subtotal, or a subtotal and tax that miss the total.
• A PDF edited after it was written: changed later than it was created, or saved again on top of itself.

To re-check, open the file's metadata in any viewer, or redo the sum. Receipt Audit reads what a C2PA manifest says; it doesn't validate the manifest's signature.

Opinion

• The vision model's judgement: the image looks generated, a region looks pasted or retouched, the fonts don't match.
• Facts that prove nothing on their own: an image editor named in the file, no camera data, an unusual tax rate, a placeholder phone number.

The risk score, from 0 to 1, combines both. Use it to decide what a person should look at, not to reject a receipt on its own.

What isn't proof

• A screenshot or a re-save strips Content Credentials. OpenAI says images made with ChatGPT and its API carry them, and that they "can be lost through everyday actions, such as taking a screenshot or converting a file" (OpenAI). Once they are gone, the file can't say how it was made.
• A clean result doesn't mean a receipt is genuine. Without metadata, and with amounts that add up, only the model's opinion is left, and it misses fakes. On RA-Bench's held-out set, 85.0% of fakes with their metadata intact scored 0.5 or more (n=20), against 31.7% with it stripped (n=20). See the benchmark.
• Missing camera data isn't proof either. Screenshots, scanners and messaging apps drop it too.
• Proof can still be wrong when the model misreads a printed amount, so a person should look before acting on it. On the same set, 2.5% of genuine receipts carried a finding labelled proof (n=81).

How to detect AI-generated receipts?

Start with what can be proven. Look in the file's metadata for Content Credentials (C2PA) or the name of an AI tool. Redo the sums: line items to subtotal, subtotal plus tax to total. Check that the tax rate exists in that country and that the phone number and tax ID aren't placeholders. Only then judge the image: fonts, alignment, a photo that looks too perfect. Receipt Audit does all of this in one check and labels each finding as proof or opinion.

Can AI detect fake receipts?

A vision model can give an opinion, and it can be wrong in both directions. On RA-Bench's held-out set, Receipt Audit gave a risk of 0.5 or more to 71.7% (n=20) of fully AI-generated fakes and to 2.1% (n=81) of genuine receipts. It found proof on 61.7% (n=20) of those fakes. A model's judgement is never proof. Proof comes from the file and the numbers: metadata that names the generator, or totals that don't add up.

Can AI-generated papers be detected?

Receipt Audit checks receipts and invoices, not essays or other papers. For any document image the same rule holds: metadata that names an AI tool is proof while it is still in the file, and a judgement of how the image looks is only an opinion.

Can ChatGPT generate a receipt?

Yes. In March 2025 TechCrunch showed ChatGPT's new image generator making convincing fake receipts. OpenAI says images made with ChatGPT or its API carry C2PA Content Credentials, so an untouched file can be traced. A screenshot or a re-save removes them. The receipts on this page were made with OpenAI's and Google's image models.

Sources: TechCrunch, 2025-03-31; OpenAI, provenance signals.

Check receipts from your app or agent

The same check over the API and MCP. Agents get a key on their own; every response labels each signal verifiable or opinion.

# Once: a key with 50 free checks, no sign-up
curl -X POST https://receiptaudit.com/api/v1/keys
# → { "api_key": "raud_…", "credits": 50, "claim_url": "…", … }

# Check a receipt by URL
curl -X POST https://receiptaudit.com/api/v1/receipts/check \
  -H "Authorization: Bearer raud_…" \
  -H "Content-Type: application/json" \
  -d '{"image_url": "https://example.com/receipt.jpg"}'

# or send the file itself
-d '{"image_base64": "'"$(base64 < receipt.jpg)"'"}'

claude mcp add --transport http receipt-audit https://receiptaudit.com/mcp

API documentation · MCP server setup