AI-generated documents: what to look for in 2026
Image models such as GPT-image and Nano Banana now produce invoices, receipts and certificates as images that look photographed. Here is what characterises them, and how they are caught.
Updated 2026-08-31 · Veriflied
What has changed
Until recently AI image models could not write legible text. Now they can. A prompt like 'photo of a receipt from an electronics store for a laptop, dated 12 March' yields an image with the right logo, legible lines and credible shadows. Documents have become the fastest-growing area of AI fraud, because they trigger money directly.
The older method still exists too: asking a text model to produce a PDF invoice. That yields a 'real' PDF file, but made with a generic library. The two methods leave completely different traces.
Signs of AI images of documents
Text that is almost right
Read everything. AI models misspell small print, mix characters in long numbers and repeat patterns. Look especially at footers, tax lines and addresses.
Numbers that do not add up
The model draws numbers, it does not calculate. Line amounts, tax and total rarely agree on an AI invoice.
Perfect yet wrong layout
Logo and structure resemble the real retailer, but details deviate: font, field order, a missing barcode, a QR code with no content.
Paper and light
Too smooth a paper surface, shadows that do not match the light source, receipt edges that bend wrongly.
No camera chain
A real phone photo carries make, model, capture time and often GPS in the file. An AI image typically has nothing, or has metadata that does not match the image's technical properties.
Signs of AI-made PDFs
PDFs from AI tools are made with a small handful of libraries: ReportLab, fpdf, WeasyPrint, matplotlib or a headless browser. None are used by billing systems. The files are made in a single pass, created and modified in the same second, without an XMP packet and often without a document id. The text is real text (it can be selected), but the fonts are standard fonts, not the retailer's.
How they are caught
AI images of documents are caught by an AI detector trained specifically on document images. It assesses whether the pixels were generated, regardless of whether the text is legible, and a heatmap shows where. AI PDFs are caught by document forensics: producer chain, timestamps and missing issuer identity. With both in the same analysis, both methods are covered, and the result can be cross-checked against the case dates: a document created after the claim was filed, or years after the claimed purchase, is a finding however it was made.
Frequently asked questions
Can an AI detector tell a real scan from an AI image?
Yes. A scan carries the scanner's noise pattern and compression, an AI image carries the generator's. The detector is trained on both and on the intermediate forms that arise when images are shared through apps.
What about screenshots of a real invoice?
A screenshot is not AI, but it has no camera chain or PDF traces either. The analysis marks it as a 'digital rendering' and relies on content and case dates. Always ask for the original file.
Do you catch the newest models too?
Detection is retrained continuously, and new generators are added to the training material shortly after release. It is a race, which is why the analysis combines several independent layers instead of relying on one.
Can I test with my own examples?
Yes. Create an account and analyse 10 documents for free.
Try it on your own document
Create an account and analyse your first 10 images or documents for free. You get a report with the concrete findings, not just a score. All processing in the EU.