When *Free Evaluation of Evidence* Meets AI: What Courts Actually Check

The case: AI transcripts nobody could vouch for

In a property dispute heard by the Civil Resolution Tribunal of British Columbia in 2024, one party tried to submit transcripts and images they said had been generated by artificial intelligence.^[1] The tribunal chair refused to admit them. The ruling was short: the party had not shown that the transcripts or images "authentically represent the speech or image" of the people they purported to depict. Without that foundation, the exhibits had no weight.

Think of it like a photocopy of a contract. If you turn up with a photocopy and the other side says "that's not what I signed", you need to show it matches the original. If you cannot, the photocopy is just paper. AI-generated evidence has the same problem, only sharper: there is often no original to compare it with, and the person offering the evidence may not know how the AI arrived at what it produced.

The tribunal did not say AI evidence is inadmissible. It said this AI evidence, offered in this way, was not reliable enough to be given weight. That is free evaluation of evidence in action.

What free evaluation of evidence means

Free evaluation of evidence is the rule in most civil law systems, including Sweden's, where it is set out in RB 35:1.^[2] Courts are not bound by formal rules about what counts as evidence or how much weight to give it. Anything can be offered, and the court decides how convincing it is.

That freedom cuts both ways. It means you can bring a video, a log file, a screenshot or an AI-generated transcript, and the court will hear you out. It also means the court will ask: where did this come from, who made it, when, and can I trust it? If the answers are vague, the evidence loses value.

AI complicates the picture because it is often hard to say exactly what happened inside the system. A transcript generated by a speech model might be mostly accurate, but if you cannot show which recording it came from, which version of the model processed it, or whether anyone edited the output, a court has reason to doubt it. The British Columbia tribunal's refusal was not a blanket rejection of AI. It was a recognition that reliability requires a foundation, and in that case the foundation was missing.

Why courts care about authenticity

Authenticity is not the same as truth. A document can be authentic—meaning it is what it purports to be—without being accurate. A forged letter is inauthentic even if everything it says is true. Courts want to know they are looking at the thing the party says they are looking at, before they decide what it means.

For AI-generated material, authenticity is a two-part question. First, did the AI system in question actually produce this output? Second, has it changed since then? Those are the questions a party must answer before the court even gets to whether the content is reliable.

In the British Columbia case, the party could not answer the first question. There was no record linking the transcripts and images to a particular AI system, input or moment in time. Without that link, the tribunal had no way to assess whether the exhibits were what the party said they were, so it gave them no weight.

Under Swedish law, the reasoning would be similar. RB 35:1 allows a party to submit whatever they like, but the court is free to disregard it if it is not convincing. An AI transcript with no provenance is not convincing.

The evidence question: what did the AI produce, and when?

When someone offers AI-generated evidence, the other side will ask the obvious questions. Which system made this? What was the input? When was it generated? Has anyone edited it? If those questions do not have clear answers, the evidence is easy to challenge.

Say a company uses a speech recognition model to transcribe customer calls, and a call recording later becomes disputed. The customer says the transcript is wrong. The company says it is accurate. Who has the better case? It depends on what the company can show. If it has a sealed record of the audio file, the model version, the time of transcription and the output as it was generated, the transcript is harder to dismiss. If it just has a text file someone exported from a system, with no record of where it came from, the transcript is easier to attack.

The point is not that every AI output needs forensic-grade documentation. The point is that courts apply free evaluation, and under free evaluation, evidence without a clear foundation gets less weight. That is true whether the evidence is a photograph, a contract or an AI-generated transcript.

How a digital witness fits this problem

This is a case where a digital witness fits well, because the core problem is showing what an AI system produced and that it has not been altered.

Here, the thing to seal is a record of each generation event: the input reference, the model version, the timestamp and a fingerprint of the output. Sealing takes a fingerprint of that record, not the content itself, so sensitive data never leaves your systems. The record is sealed the moment the AI produces the output. Later, if the output is questioned, anyone you share the record with can check that the fingerprint still matches. If it does, the record has not changed since that moment.

What the witness shows is that the record linking an output to a system and a time existed when you say it did, and has not been edited. What it does not show is that the AI's output was accurate, that the input was authentic, or that the model worked as intended. Those are separate questions. A sealed record helps you answer the foundation question—did this system produce this output at this time—so the court can move on to whether the output is reliable.

You can start with WitniumChain, which offers a REST API, an SDK and an MCP server. The content you want to seal stays with you. A fingerprint is sent to the chain, and you get a witness ID back. To check, anyone calculates the fingerprint again and compares it with the sealed one. The first 2,000 witnesses a month are free, and the documentation is public.

How it would work

1. The AI generates a transcript. Your system records the audio fingerprint, the model version, the timestamp and the output.
2. That record is sealed immediately. You keep the record; the chain keeps the fingerprint.
3. Months later, if the transcript is disputed, you share the sealed record. Anyone can check that it matches the fingerprint from the day it was generated.

What this means for you

If you use AI to generate material that might later be questioned—transcripts, summaries, images, decisions—ask yourself these questions now:

• Can you show which system produced a given output, and when?
• Do you have a record of the model version and input that was used?
• If someone challenges an AI-generated document or transcript six months from now, could you demonstrate it has not been altered?
• Who in your organisation would be responsible for answering those questions in a dispute?

The British Columbia ruling is a reminder that courts do not accept AI evidence just because it exists. They ask where it came from, and if you cannot answer that clearly, the evidence loses value. Free evaluation of evidence means you can submit anything, but it also means the court is free to give it no weight.

What to watch

Article 50 of the AI Act requires AI-generated content to carry a machine-readable marking, and deployers must disclose deepfakes and certain AI-generated text published to inform the public.^[3] Those duties apply from 2 August 2026. The marking is about transparency, not authentication, but it creates an expectation: if content is marked as AI-generated, people will want to know it is the marking that was applied at generation, not one added or removed later. That is another place where a sealed record of what was generated, and when, can matter.

Courts in Europe and beyond are still working out how to assess AI evidence. The principles—authenticity, reliability, free evaluation—are not new, but applying them to generative AI is. Expect more rulings like this one, where the material is rejected not because it is AI-generated, but because the party offering it could not show where it came from.