Deepfake Evidence in Court: A New Database and the Proof Problem

The database and the pattern

Wisevoter's database of AI in court evidence now lists 26 cases from around the world ^[1]. They include forged contracts in property disputes, fabricated emails in divorce proceedings and AI-generated documents submitted to courts in criminal cases. In one, a lawyer used ChatGPT to draft a brief and submitted case citations the model had made up ^[1].

The pattern is the same. Someone submits a document, an email or a screenshot. The other side says it's fake. The court has to decide, and often the only test is whether the story around it holds up. Forensic analysis can sometimes show that an image has been edited, but AI-generated text typed into a real email client leaves no trail.

It's a bit like someone turning up with a printed bank statement. The paper looks real, the numbers are formatted correctly and the logo is in the right place. Proving it wasn't issued by the bank takes work—and if the person who printed it also controls the account, it takes even more.

What the AI Act says

Since 2 August 2026, Article 50 of the AI Act has required that people must as a rule be told when they are interacting with an AI system, and that AI-generated content must carry a machine-readable marking ^[2]. The idea is transparency: if you're talking to a bot, or looking at a generated image, you should know.

The rule has teeth. Breaches can lead to fines of up to 15 million euros or 3 per cent of global annual turnover, whichever is higher ^[2]. For small and medium-sized enterprises, the lower amount applies.

But the marking requirement doesn't solve the evidence problem. Marks can be stripped when a file is converted, compressed or shared. A machine-readable watermark in an image doesn't survive a screenshot. And Article 50 says nothing about how to prove that a particular document existed on a particular day, or that it hasn't been altered since.

The proof problem

In Swedish courts, the rule is free assessment of evidence under Chapter 35, Section 1 of the Code of Judicial Procedure (RB 35:1) ^[3]. Anything may be submitted, and the court decides how much weight to give it. A forged email can be admitted as evidence. Whether it's believed is another question.

The challenge is this: when a convincing fake can be generated in seconds, the burden shifts. It's no longer enough to produce a document. You have to show that it's genuine, and that means answering two questions. First, did this document exist at the time you say it did? Second, has it changed since?

That's harder than it sounds. If you keep the only copy, and you could edit it, why should anyone believe you didn't? A timestamp from your own server proves nothing. A saved file on your laptop has a modification date you control. Even a PDF signed with a certificate only shows who signed it, not when the content was created or whether it was altered before signing.

What this means for you

If your organisation handles contracts, decisions, correspondence or any record that might be disputed, here are three questions worth asking:

1. Could you show that a document existed on a given day, in a way the other side can check?
2. If someone challenges a record you keep—a log, an email, a decision—what evidence do you have that it hasn't been edited?
3. When you generate content with AI, do you keep a record of what was generated, by which system and when?

The database Wisevoter has compiled shows what's coming. Courts are already seeing fabricated evidence. The side with a checkable record starts ahead.

How a witness would work here

This is where a digital witness fits. The thing to seal is the document, email or decision record at the moment it's created or sent. Sealing takes a fingerprint of the content, not the content itself, so what you're handling stays with you. The fingerprint is posted to WitniumChain ^[4], which anyone can read but nobody—including us—can rewrite.

Months later, if the document is challenged, anyone you share it with can calculate the fingerprint again and check it against the sealed one. If it matches, the document hasn't changed. Change a single character and it won't match.

Who checks: the other party, their lawyer, an expert witness or the court.

What it shows: that the document is exactly what existed at the time it was sealed.

What it doesn't do: it doesn't show that the content is true, that a signature is valid or that the document was sent to the right person. It answers one question—has this been altered?—and leaves the rest to the usual rules of evidence.

Your system sends the fingerprint through the REST API ^[5], the SDK or the MCP server, and gets a witness ID back. The first 2,000 witnesses a month are free.

How it would work

1. A contract is finalised and sent. At that moment, your system takes a fingerprint and seals it.
2. Six months later, the other party claims a clause was different. You share the contract and the witness ID.
3. They calculate the fingerprint. It matches. The clause hasn't changed since the day it was sent.

What to watch

The AI Act's transparency requirements apply now, and courts around the world are starting to write the rules on AI-generated evidence as cases come in. Expect more guidance on what counts as adequate proof that a document is genuine, and on what duties organisations have to keep checkable records of what their systems generate.

In Sweden, a new Product Liability Act is proposed to enter into force on 9 December 2026 ^[6]. It will treat software, including AI, as a product. If software causes harm, the question of which version was running, and what it did, becomes a matter for courts and insurers. Sealed version records and decision logs move from nice-to-have to part of the defence.

Sources

1. Wisevoter, "Database Tracks AI Deepfakes in Court Evidence", 25 October 2026, https://wisevoter.com/world/2026/10/25/database-tracks-ai-deepfakes-court
2. European Commission AI Act Service Desk, "Article 50: Transparency obligations for certain AI systems", Regulation (EU) 2024/1689, https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50
3. Riksdagen, Rättegångsbalken (1942:740), 35 kap. 1 §, https://data.riksdagen.se/dokument/sfs-1942-740.html
4. WitniumChain documentation, https://witniumchain.com/docs
5. WitniumChain REST API, https://witniumchain.com/docs
6. Regeringen, "En ny produktansvarslag", lagrådsremiss, August 2026, https://www.regeringen.se/rattsliga-dokument/lagradsremiss/2026/08/en-ny-produktansvarslag/