Courts are tightening authentication rules for AI evidence and deepfakes

When someone presents a video in court, the other side no longer just argues about what it shows. They argue about whether it is real.

Deepfakes have moved from a novelty to a courtroom problem. A judge in a US fraud case recently excluded a video because the party presenting it could not show it hadn't been manipulated. In another, an expert spent half a day explaining how to spot synthetic edits. Courts that once took digital evidence at face value are now asking: how do you know?

It is a bit like handing over a photocopy of a contract with the signature cut out. The contract might be genuine, but if you cannot show the original or explain where it came from, the other side will ask why you are hiding it.

Swedish courts follow the principle of free evaluation of evidence, set out in Chapter 35, Section 1 of the Code of Judicial Procedure ^[1]. Anything can be presented. The question is what weight the court gives it, and that depends on how reliable it appears. A grainy screenshot of a chatbot conversation is evidence. So is a sealed log with a verification link. The court decides which one to trust.

The authentication gap

Authentication means showing that a piece of evidence is what you say it is. For a paper document, that might be a witness who saw it signed. For digital evidence, it is harder.

A video file has metadata: the time it was created, the device, the software version. But metadata can be rewritten. A timestamp can be altered. An image can be run through editing software that leaves no obvious trace. Courts know this, and they are responding by raising the bar.

In one recent case, a party submitted dashcam footage to support a claim about a traffic incident. The other side hired a forensic analyst who found that the file's creation date had been changed twice. The footage itself might have been genuine, but the court gave it almost no weight, because the party presenting it could not explain the discrepancies ^[2].

Expert testimony is becoming the norm. The expert examines the file, looks for signs of editing, checks whether the metadata is consistent with the claimed timeline, and presents a report. That costs money and takes time, but it is often the only way to get digital evidence admitted when the other side objects.

What the AI Act says about transparency

Article 50 of the AI Act requires that people be told when they are interacting with an AI system, and that AI-generated content be marked in a machine-readable way ^[3]. The marking is less a warning label and more a barcode: machines read it, people barely notice it.

The article applies from 2 August 2026. It does not say how to prove the marking was there when the content was created, or that the content itself hasn't been altered. That is a separate question, and it matters.

Say a customer disputes what a chatbot told them about a refund policy. The company has a log showing the conversation. The customer has a screenshot showing something different. Under Article 50, the chatbot should have told the customer they were talking to an AI. But if the company cannot show that its log is the one written on the day, the screenshot may be more convincing.

The evidence stack

Authentication is the first layer. Once the court accepts that the evidence is what you say it is, the next question is what it proves.

A sealed photograph of a damaged container shows that the photograph existed at a certain time and hasn't been altered. It does not show that the damage happened when and where the photographer says it did, or that it was caused in the way claimed. Those are separate arguments, built on other evidence: witness statements, cargo manifests, weather reports.

But if the photograph itself is challenged, the argument never gets to those other layers. The party presenting it has to stop and prove the photograph is genuine. That is where a digital witness comes in.

How a witness fits here

When a court asks how you know a file hasn't changed, the answer needs to be more than "we kept it safe". Courts are sceptical of assurances, particularly when the party giving them has an interest in the outcome.

A digital witness seals a fingerprint of the file at the moment it is created or received. The fingerprint goes onto a chain where any later change would show, even one made by the organisation that sealed it. To check, anyone calculates the fingerprint again and compares it with the sealed one. If they match, the file is unchanged. Alter a single pixel and it will not match.

In the container case, the photograph would be sealed the moment it is taken, along with the time and, if wanted, the location. When the claim is disputed a year later, the surveyor can show the court the sealed record and the verification link. The other side can check it themselves, without having to trust the surveyor's word.

What it does not do is prove that the photograph shows what the surveyor says it shows, or that the damage was caused in the way claimed. It just removes one line of attack: the claim that the photograph has been doctored.

You seal the file through WitniumChain, using the REST API, the SDK or the MCP server. The content stays with you; only the fingerprint is sealed. Anyone you share the record with can check it through the verification link. The first 2,000 witnesses a month are free. Read the docs to see how it works in practice.

How it would work

1. Capture the evidence (a photo, a video, a chatbot log) and seal it immediately.
2. Store the file and the witness ID.
3. When it is questioned, share the file and the verification link. The other side checks that the fingerprint matches.

What this means for you

If your organisation creates or relies on digital evidence, ask these questions now:

• Could you show a court that a file, log or screenshot is exactly what was created on a given day?
• Who keeps the original, and how would you prove it hasn't been edited?
• If the other side presents a deepfake or an altered document, could you show your version is the genuine one?
• When your AI generates content or interacts with customers, do you keep a sealed record of what it said?

Authentication rules are tightening because courts have seen too many doctored files and too many metadata discrepancies. The trend is not reversing. It is spreading.

What to watch

Article 50 of the AI Act, on transparency and marking of AI-generated content, applies from 2 August 2026 ^[3]. Expect courts to start asking how parties can show that the marking was there, and that the content is the version created.

Jurisdictions that follow free evaluation of evidence, such as Sweden, give judges discretion on what weight to assign. That makes authentication even more important: the party with the weaker provenance loses ground before the substantive argument begins.

Evidence that can be checked beats evidence that has to be trusted. That is not new. What is new is that courts now expect you to bring the proof.