Deepfake Nearly Indicted an Innocent Person. Courts Have Zero Protocols to Stop the Next One.

A judge in California caught a deepfake video that a plaintiff submitted as witness testimony in court. The judge spotted it — not because of any screening protocol, not because of any detection software, but because the person in the video blinked oddly and repeated the same facial expression over and over. That was it. Luck. Not procedure. Luck.

That case — Mendones v

That case — Mendones v. Cushman and Wakefield — is believed to be the first time a court identified A.I.-generated evidence after it was already submitted. Experts in digital forensics don't think it was the first time a deepfake made it into a courtroom. They think it was just the first time anyone noticed. And that distinction should sit with you for a second, whether you're a prosecutor building a case or a person who's ever been recorded on a doorbell camera. Because right now, no court in the country has a formal procedure — none — for testing whether video or audio evidence is real before it's used against someone. The legal standards governing what counts as admissible evidence were written before this technology existed. So the question running through this entire story is simple. If the system can't tell real from fake, what happens to the people inside that system?

Start with what the judge actually saw. In the Mendones case, the video was submitted as testimony from a witness. It looked like a real person speaking on camera. But the facial movements didn't track naturally. Expressions looped. The kind of thing you might miss if you watched it once on a laptop in a busy courtroom. The judge happened to look closely enough to see it. No algorithm flagged it. No authentication software caught it first. A human being got lucky.

Now widen the lens. Current generative A.I. tools can produce video that's nearly indistinguishable from footage of a real person. That means manual detection — just watching a video and deciding whether it looks right — is no longer a reliable strategy. And the tools built to detect A.I.-generated content have their own problems. According to researchers in digital forensics, those detection systems have proven unreliable and biased. They miss things. They flag things that are real. And humans aren't much better. Studies show people are poor judges of whether digital content is authentic or synthetic. So neither the machines nor the people watching the machines can consistently tell the difference. For anyone who's ever had a photo or video used in a legal matter — a car accident, a custody dispute, a workplace complaint — that gap is personal.

There's a concept legal scholars use to describe

There's a concept legal scholars use to describe what happens next. They call it the Liar's Dividend. It works like this. Once deepfakes become common enough, every piece of digital evidence becomes suspect — even the real stuff. A defense attorney can point to a legitimate video and say, "How do you know that's not A.I.?" And right now, there's no standard answer. That's a cascading credibility crisis. It doesn't just help people who fake evidence. It helps people who want to discredit evidence that's completely genuine. For investigators, that means a solid case can unravel over a question no one had to ask five years ago. For the rest of us, it means the security camera footage from your front porch could be challenged as fabricated in court, and the system has no default way to prove otherwise.

Some judges are starting to push back. According to reporting from W.J.L.A., multiple judges have now publicly acknowledged the risk that generative A.I. poses to evidence integrity. They're calling for changes to how courts evaluate digital submissions. Experts recommend a set of countermeasures — certified video companies that verify footage at the source, strict chain-of-custody rules for digital files, source stamping that tracks where a file originated, and in some cases, more reliance on in-person testimony. But none of these are standard yet. None of them are required. They're suggestions floating above a system that still runs on rules written before anyone could generate a fake human face in thirty seconds.

Critics of mandatory authentication say it introduces its own risks. Verification tools can produce errors. They can slow down proceedings. That's a fair concern. But the cost of one wrongful indictment — one innocent person dragged through prosecution because no one checked whether the evidence was real — outweighs the friction of building authentication into the process. And the approach doesn't have to rely on a single detection tool. Provenance infrastructure — tracking where a file came from, how it was stored, whether it was altered — is broader than any one piece of software. It's about treating digital evidence the way we already treat physical evidence. You wouldn't admit a blood sample without chain-of-custody documentation. Why would you admit a video without it?