When forged documents cost billions, can AI catch AI?

The scale of the problem

Document fraud costs European insurers an estimated €13 billion a year ^[1]. Banks lose billions more to forged invoices, manipulated statements and fake identity documents ^[2]. The industry has turned to AI to catch what human reviewers miss: altered numbers on a payslip, cloned signatures on a contract, synthetic faces on passports.

Companies like RIG provide detection tools to insurers, banks and authorities ^[3]. The technology is improving fast. The tools can spot a manipulated pixel or an inconsistent shadow, and they flag thousands of suspicious documents every month.

But here's the thing: catching a forgery is one job, and proving you caught it is another. When a claim is disputed or a decision is challenged, the question isn't just whether the document was forged. It's whether you can show what your system found, and when.

Why the record matters

Say a bank's AI flags an invoice as manipulated and the loan application is declined. Six months later, the applicant disputes the decision. The bank needs to show what the detection system saw: which version of the invoice was submitted, what the model flagged and which human reviewer approved the decision.

If the bank can't produce that record, or if the record itself can be questioned, the case gets harder. A log the bank's own IT team can edit is easier to challenge than one nobody can quietly rewrite.

It's like the difference between a receipt you wrote yourself and a till receipt with a timestamp. Both say you paid, but one is easier to trust.

What the AI Act asks for

From 2 December 2027, high-risk AI systems under Annex III must technically allow for automatic logging of events over their lifetime ^[4]. That includes systems used to assess creditworthiness of natural persons and to determine pricing in life and health insurance.

Article 12 says what to log. It doesn't say how to protect the logs, and it doesn't require them to be tamper-proof ^[5]. But a log you can edit is still a log someone will question.

The same goes for detection AI. If it's used to make decisions about people—declining a loan, flagging a suspicious claim—it may fall under the high-risk rules, and those rules ask for records of what the system did.

The gap between detection and proof

Detection AI answers one question: is this document real? But when the decision is challenged, three more questions follow:

1. What exactly did the system flag?
2. When did it flag it?
3. Has the record of what it found changed since then?

The original document may have been updated, deleted or replaced. The person disputing the decision may produce a different version and ask why the bank or insurer saw something else. At that point, the organisation needs its own unchanged record of what its system analysed and what it concluded.

That's the evidence question, and it comes up in audits, disputes and regulatory reviews. It isn't about the AI's accuracy. It's about whether the record of what it did can be trusted.

What this means for you

If you provide or use AI to detect forged or manipulated documents, here are three questions worth asking now:

• When a flagged document is questioned months later, can you show what your system saw at the time, unchanged?
• If the person affected produces a different version of the document, do you have an independent record of the one you analysed?
• When the AI Act's logging requirements apply, will your logs hold up if someone with the right access edits them?

Protecting logs isn't just about compliance. It's about being able to answer those questions without asking anyone to take your word for it.

How a digital witness fits here

This is a strong fit for a digital witness, because the problem is showing that a record of what the detection AI found hasn't changed.

Here, the thing to seal is each detection event: the document fingerprint, the model version, the flags raised and the reviewer's decision. Sealing takes a fingerprint of that record, not the document itself or its content, so sensitive data never leaves your systems.

When a decision is questioned, anyone you share the record with—an auditor, a court, the person affected—can check that it matches the sealed fingerprint. Change a single character in the record and it won't match.

The check is independent. It doesn't rely on trusting you or your logs. That's the point.

What it doesn't do: it doesn't show that the AI's conclusion was correct, that the document was in fact forged, or that the decision was fair. It shows that the record is the one written at the time.

You can seal records through WitniumChain with the REST API, the SDK or the MCP server. The content stays with you, and the first 2,000 witnesses a month are free. Read the docs to see how it works.

How it would work

1. Your detection system analyses a document and writes a decision record.
2. The record is sealed: a fingerprint is sent to WitniumChain and a witness ID comes back.
3. Months later, if the decision is questioned, you share the record and the witness ID. The other side calculates the fingerprint and checks that it matches.

What to watch

• 2 December 2027: high-risk requirements under Annex III of the AI Act apply, including automatic logging.
• January 2028: the European Commission is expected to publish guidelines on the practical application of Article 12.

Document fraud isn't going away, and neither are the questions that follow a disputed decision. The sharper the detection AI gets, the more it matters that you can show what it found.