How to detect AI-generated documents

What generative AI changes for document fraud

In a November 2024 alert, FinCEN, the US financial crimes authority, says generative AI tools have greatly reduced the resources needed to produce realistic content. Since 2023 it has seen more reports from financial institutions describing fake identity documents created or altered this way. That finding concerns the United States and identity documents, but the mechanism applies to any supporting document. In practice, three things change:

• Less skill is needed. Producing an image of a document no longer takes mastery of an image editor.
• Less repetition. A fake no longer has to be a copy of one template, so comparing it with a known example is not enough.
• Well-written text. A flawless letter or certificate is no longer reassuring: spelling has stopped being a clue.

It is still worth keeping a sense of proportion. A late 2025 preprint concludes that current models imitate the surface look of identity documents without reproducing their structural and forensic authenticity, and that the risk of expert-grade forgeries may be overestimated. In other words, AI fools a quick glance more easily than a methodical check. Nor has it replaced ordinary edits to a genuine document, which you still need to look for.

Visible signs, and their limits

Some clues are worth a look. None of them is proof, and their absence guarantees nothing:

• Text: unreadable small print, legal notices that mean nothing, warped words in a secondary area.
• Identifiers: a company number, IBAN or reference whose format or length is not what you would expect.
• Logos, stamps and signatures: approximate shapes, or a stamp without a single printing flaw.
• Layout: columns that do not follow a grid, table rows that drift.
• Photos: FinCEN cites a photo that is internally inconsistent or does not fit other information, such as a date of birth. The report from NIST, the US standards institute, mentions asymmetric details, such as earrings or reflections in the eyes. On an insurance claim photo, also look at shadows, reflections and lettering.

These signs have three limits. First, they depend on today’s models, which keep improving. Second, the human eye is unreliable: in a study of generated images, video and audio, participants did little better than chance, and the authors consider human perception no longer viable as a defense. Third, a poor scan or heavy compression produces similar flaws, so a genuine document can look suspicious.

Provenance marks: C2PA and watermarks

Rather than guessing from the image, you can read what the file declares. The C2PA (Coalition for Content Provenance and Authenticity) is a group of companies and organizations behind an open technical standard. Its Content Credentials are signed data attached to a file: how it was made, with what tools, when, and how it changed. Its specification covers images and PDF among others, and provides a way to state that content was produced by an AI model.

This data does not say the content is true. The C2PA’s principles rule out any judgment on whether the information is good or bad: validation only establishes that it is tied to the file, correctly formed and free from tampering. And the C2PA itself acknowledges that a manifest, the block holding this signed data, can be separated from the file.

Invisible watermarks

Some tools also embed an invisible signal in the content. NIST points out that detecting one always carries some probability of error, that metadata may be inaccurate, and that no technique solves the problem alone.

What EU law changes

Article 50 of the EU AI Act has required providers of generative systems to mark their outputs in a machine-readable format since 2 August 2026. Systems already on the market before that date have until 2 December 2026, and there are exceptions, for instance for simple assistance with standard editing. These marks should therefore spread, without covering every case: FinCEN notes that some tools are open source and that their safeguards can be circumvented.

In practice, read these marks as follows:

• A mark indicates AI generation: ask about it, without assuming fraud. An illustration made with AI does not turn a document into a fake.
• A valid mark describes a consistent origin: useful information, to be cross-checked like any other.
• No mark at all: conclude nothing. NIST notes that this absence could become a signal if such data becomes widespread, and then only in some contexts.

Checking a suspicious document: the method

1. Get the original file

   A screenshot or a document passed along by a messaging app may have lost its information. Ask for the file as it was issued.

2. Read what the file declares

   Properties, software, dates, any provenance mark. Our free tool reads the properties of a PDF or an image in your browser, and our guide explains how to tell if a PDF was edited.

3. Trace it back to the issuer

   AI can imitate a page, but it cannot create a record at the issuer. Ask the issuer through its official contact details, or use a checking service where one exists, such as a national business register or the French Kbis check and tax notice service (both in French).

4. Read any signed codes

   A generator can draw a code, but it cannot produce a valid signature. On French documents, a 2D-Doc can be checked with a recognized reader, and its data compared with what is printed.

5. Look for other copies of a photo

   FinCEN notes that a reverse image search may reveal that an identity photo comes from a gallery of generated faces. The same habit is useful for a photo of damage.

6. Review the whole case file again

   FinCEN notes that institutions often detect this content by re-reviewing account opening documents. Compare the documents with one another: that is where inconsistencies appear.

Why cross-checking remains the decisive check

When appearance costs nothing to produce, its value as evidence falls. What counts is whatever a forger does not control: registers, issuers, money flows, other documents. A few examples:

• A payslip: does the net pay appear on the bank statement, and does the employer exist?
• A proof of address: does the address match the other documents, and is there a signed code?
• An invoice: does it match an order, and does the VAT number check out?
• An insurance claim photo: does it agree with the quote, the expert’s report and the stated circumstances?

FinCEN puts it clearly: no single indicator is enough, and the surrounding facts and circumstances must be weighed. Its indicators include inconsistencies between several documents supplied by the same person: it is the case file, more than the single document, that gives a fake away. Our main guide to document authenticity sets out a four-level method.

Mistakes to avoid

• Accusing on the word of a detector. FinCEN lists such results among indicators, none of which is enough alone, and NIST stresses that wrongly labeling human content as AI-generated can do serious harm.
• Treating text written with AI as a forgery. Drafting a letter with an assistant is not fraud. What matters is that the facts and the issuer are real.
• Taking an unmarked document as genuine. A mark can be lost, and not all tools apply one.
• Only looking for AI. Many checks hold however the fake was made, and an ordinary edit shows none of the signs of a generated image.

Legally, the tool makes no difference. In France, for example, article 441-1 of the Criminal Code (in French) covers any fraudulent alteration of the truth liable to cause harm, “by whatever means”. Forgery and the use of forged documents are punished by three years in prison and a €45,000 fine, and our guide on document fraud says more.

How Docusan helps

Among its analyses, Docusan, a French document fraud detection service, looks at the pixel level for AI-generated images and deepfakes, as well as content inserted into an image. It also examines the file’s metadata and structure, whether PDF, PNG, JPEG or TIFF, then gives a verdict: No detection, Warning or Alert. The anomalies are located on the document, and the decision always stays with your teams.

Docusan does not contact the issuer or government databases: it complements the cross-checks above and does not replace them. You can use it from the platform or integrate it into your tools with its API, and your documents are never used to train our models (security and privacy).