Courts grapple with AI-altered evidence as *manipulation beats detection*

The problem lands in court

A lawyer presents a photo in court. It shows damage to a car, taken at the scene. The other side questions whether the photo is real. The lawyer insists it is. The judge asks: how do you know?

That question, once rare, is becoming routine. AI tools can now alter photos, fake voices and generate video that looks real enough to fool most people. The New Jersey State Bar Association is running a session this week titled Truth or Tech: Navigating AI Generated Evidence in the Courtroom ^[1], and the title gives away the anxiety. When a piece of evidence might be AI-generated or AI-altered, how should a court decide whether to trust it?

Detection doesn't keep up

The first instinct is to run the evidence through a detector. Several tools claim to spot AI-generated images, and some work well on obvious fakes. The trouble is that manipulation is often more subtle than generation, and a skilled edit to a real photo can leave no trace a detector will catch.

Think of it like a fake banknote. If someone prints a whole note from scratch, the paper and the ink might give it away. But if they take a real note and carefully change one number, you need to compare it with the original to see the difference. Without the original, you're guessing.

Detection tools look for patterns in how an image was made. AI-generated images often have statistical quirks a detector can spot. But an AI-altered image starts with a real photo, and the alteration can be so precise that the quirks blend into the noise. Research published in 2024 found that commercially available detection tools missed a significant proportion of sophisticated alterations ^[2]. In short, if someone wants to fool a detector and knows how, they probably can.

Free evaluation, harder questions

In jurisdictions that follow free evaluation of evidence, like Sweden and many other European countries, a court isn't bound by fixed rules about what counts as proof ^[3]. Anything can be presented, and the judge weighs it based on how reliable it seems. That flexibility is useful, but it also means the judge has to form a view on whether a photo or a recording has been tampered with.

The traditional test is circumstantial. Does the evidence fit with the rest of the story? Is there a reason to doubt it? If the answer to both is yes, the court might ask for more. The problem with AI-altered material is that it can fit the story perfectly, because that's what it was made to do.

Say a claimant presents a photo of a damaged container, allegedly taken on the day of the incident. The photo shows the damage in detail, the lighting looks right for the time of day, and there's nothing obviously wrong with it. The defendant says the damage wasn't that bad and asks how the claimant can prove the photo hasn't been edited. The claimant points to the metadata, which shows the date and the camera model. Metadata can be rewritten. The defendant presses: is there anything that shows this exact image existed on that day, unchanged?

Often, there isn't.

The burden creeps backwards

Under normal rules, the party making a claim has to prove it. But when the other side questions the authenticity of a piece of evidence, the question shifts. It's no longer just whether the photo shows damage. It's whether the photo itself is what it claims to be.

In the British Columbia tribunal case involving Air Canada's chatbot, covered in an earlier piece in this series, the airline offered no evidence to contradict the customer's screenshot, so the tribunal assumed such evidence wouldn't have helped the airline ^[4]. The customer had a record, the airline didn't, and that decided it. The airline could have argued the screenshot was fake, but without its own log of what the chatbot actually said, the argument would have been weak.

The pattern is becoming common. The side with a contemporaneous, unalterable record has the stronger position. The side relying on something that could have been edited, even if it wasn't, starts behind.

What this means for you

If you present evidence in court, or advise someone who does, here are three questions to ask early:

• Do you have a record made at the time? A photo taken later, or a document prepared after the event, is easier to challenge than one captured as it happened.
• Can you show it hasn't changed? Metadata isn't enough. Ask whether there's a way to demonstrate that the file you're presenting is exactly what existed on the day in question.
• What would you say if the other side questions it? Practise the answer before the hearing, because the question is coming.

For organisations that generate or handle evidence routinely—insurers, surveyors, regulated firms—the question is operational. When a photo is taken, a decision is logged or a document is approved, what creates a record that will hold up months or years later, if it's disputed?

Sealing what you capture

A digital witness can help here. When a photo, a document or a log entry is captured, the system takes a fingerprint of it and seals the fingerprint on a record where any later change would show. The content stays with you. The fingerprint goes onto a chain where even the operator can't quietly rewrite it.

Say a claims adjuster photographs a damaged container on the quay. At the moment the photo is taken, a fingerprint is sealed with the time. A year later, when the claim is disputed, anyone the adjuster shares the photo with can calculate the fingerprint again and check it against the sealed one. If they match, the photo hasn't changed since that moment. If they don't, something has.

What it shows is narrow but useful: that the file you're presenting is the one that existed when it was sealed. It doesn't show that the photo depicts what you say it depicts, or that it was taken where you say it was. But it does mean the other side can't credibly argue you've edited it since.

The thing to seal is each piece of evidence at capture: photos, video, documents and, if you're logging decisions, each log entry as it's written. Sealing takes a fingerprint, not the data, so the evidence itself never leaves your systems. To check, the person you share it with calculates the fingerprint themselves and compares it. No trust required.

That's done with WitniumChain, through a REST API, an SDK or an MCP server. The content stays with you. The first 2,000 seals a month are free. Read the documentation to see how it fits your workflow.

How it would work

1. At capture: the app or system takes a fingerprint of the photo, document or log entry and sends it to WitniumChain. A witness ID comes back.
2. Later: when the evidence is questioned, you share the file and the witness ID. The recipient calculates the fingerprint and checks it.
3. The check: if the fingerprints match, the file is exactly what existed when it was sealed. If they don't, it isn't.

What to watch

Courts haven't settled on a standard for authenticating digital evidence when AI manipulation is possible. Until they do, the practical test will be whether you can show a piece of evidence existed at a particular time and hasn't changed. The party that can do that, simply and independently, will have an easier hearing.

Meanwhile, AI tools are getting better at altering content in ways that leave no visible trace. Detection will improve, but so will manipulation. The gap between the two isn't closing.

For now, the lesson is straightforward: if evidence matters to your case or your business, decide how you'll prove it's unchanged before someone asks you to.