The Witness Who Wasn't There
Judge Victoria Kolakowski was reviewing video testimony in a California case called Mendones v. Cushman & Wakefield when something felt off. The witness's face sat nearly motionless. There were strange cuts in the footage, a kind of looping repetition in how she moved her hands. Kolakowski spotted it — an AI-generated deepfake of a real person, submitted as authentic testimony by self-represented plaintiffs.
It was one of the first times a deepfake had been caught in court. But here's what keeps me up at night: the technology used was primitive by today's standards. We're talking about glitches any decent video editor would spot in seconds. And yet the judge only got lucky — experts noted she might have missed it if the artifacts hadn't been so obvious.
Fast forward to now. Anyone with an inexpensive monthly subscription to a consumer AI service can produce something far more convincing in minutes. The Mendones deepfake was the warning shot. We're past the point where a judge's trained eye is enough to catch these things.
This isn't hypothetical. In Florida, a woman spent two days in jail because her ex-boyfriend fabricated AI-generated text messages and used them to get her arrested for violating a protective order. No one verified the evidence. Prosecutors eventually dropped the charges — but only after eight months of legal proceedings that destroyed her life in the meantime.
Eight months. For a crime that didn't exist.
Why Video Evidence Has Always Been King
There's a reason we say "seeing is believing." Humans are visual creatures — we process visual information faster than text or abstract argument, and we give it more emotional weight. When you watch someone on a screen say something, you feel a powerful sense of presence. Researchers call this the "veridicality heuristic" — your brain's intuitive assumption that perception equals truth.
This is precisely why video evidence has always been so persuasive in courtrooms. Juries trust it. Judges rely on it. A recording, unlike a human witness, can't misremember. Can't be coached. Can't change its story.
For most of legal history, video was considered close to infallible. It could be a smoking gun — incontrovertible proof, dispositive. That assumption is now crumbling.
Think about what that means. We've built decades of legal safeguards around the fallibility of human memory. Witnesses lie. They misremember. They're influenced by leading questions and post-event information. The system has procedures for all of that — cross-examination, jury instructions about witness credibility, the whole machinery. But we never built safeguards for the fallibility of perception itself. Because perception was supposed to be reliable. Your eyes don't lie, right?
Wrong. Not anymore.
The Deepfake Defense — And Its Mirror
Defense attorneys have started invoking what's being called "the deepfake defense." Here's how it works: a real recording exists — a voicemail, a video, a text message chain. But the accused claims it's AI-generated. A fabrication.
The problem isn't that this defense always works. The problem is what it does to the system over time. When jurors hear "deepfake defense" repeatedly, they start getting skeptical of all digital evidence. Legitimate recordings lose their persuasive power. But here's the cruel twist — those same jurors remain susceptible to sophisticated fakes.
It creates what legal scholars are calling "epistemological paralysis" — a courtroom where no one quite knows what to trust. You end up with jurors who dismiss genuine evidence while remaining vulnerable to convincing fabrications. The psychological double bind is brutal.
The stakes aren't abstract. Judge Erica Yew of California's Santa Clara County Superior Court raised a scenario that should terrify anyone who uses the legal system: someone could generate a false vehicle title record using AI, present it to a county clerk who lacks the expertise to identify it as fraudulent, and have it entered into official records. Then obtain a certified copy and present it as authentic documentation.
"Now do I, as a judge, have to question a source of evidence that has traditionally been reliable?" Yew asked. "We're in a whole new frontier."
Chief Judge Anna Blackburne-Rigsby of the District of Columbia Court of Appeals put it even more directly: "If you're in a trial court presenting a case and you're afraid as a litigant or as a party, maybe the other side is using evidence that's been altered by artificial intelligence — does the judge know? Does the judge understand how this could impact my case?" The issue, she said, cuts to whether people believe the legal process is fair.
It does.
The Memory Contamination Problem
Decades of memory research offer a useful parallel. Psychologist John Wixted of UC San Diego — whose work has influenced how police conduct lineups across the country — has argued that eyewitness memory is no different than DNA or fingerprints in one crucial way: it can be contaminated.
Here's what he means. Forensic evidence gets contaminated when it travels too far from the crime scene without proper chain of custody. DNA samples degrade. Fingerprints get smudged. The police rush to collect uncontaminated evidence because they know it gets worse the longer you wait.
The same logic applies to digital evidence. A recording becomes harder to trust the further it travels from its original source without verified chain of custody. A text message gets forwarded, screenshotted, re-sent. Each hop introduces the possibility of alteration.
But here's where it gets worse than memory contamination. Wixted's research shows that eyewitnesses often don't know they have false memories. They remember clearly — except it's wrong. The memory feels absolutely true.
Generative AI works the same way. It uses two algorithms — one that generates content and one that evaluates its realism — in a constant feedback loop. The result is an arms race that human perception isn't equipped to win. Detection tools show high accuracy on clean datasets but collapse when faced with real-world fakes. They need recalibration every time a new AI tool drops.
We're asking human brains to do something they were never designed for: distinguish between authentic and AI-generated content in real time, under the pressure of a courtroom.
It's not going to work.
What Needs to Change
The proposed Federal Rule of Evidence 707 would extend the Rule 702 reliability standard to machine-generated evidence — treating AI output more like expert testimony, where its methodology can be challenged. The Advisory Committee on Evidence Rules published this in August 2025, and the comment window closed in February 2026. It's an important step.
But critics have a legitimate concern: Rule 707 only applies to evidence that the proponent acknowledges was created by AI. It does nothing for undisclosed deepfakes — the ones you don't know exist until someone raises the question. And honestly, that's probably going to be most of the problem cases.
California's Judicial Council AI Task Force is developing guidance for evaluating AI-generated evidence. The National Center for State Courts has published bench cards to help judges assess both acknowledged and unacknowledged AI materials. These are real, concrete steps in the right direction.
But I think we're focusing too much on the legal framework and not enough on the psychological education piece. Research has shown that jurors who lack accurate knowledge of how memory and perception work are far more likely to make flawed evidentiary judgments. The same will be true of AI literacy.
A juror who understands, conceptually, that a video can be fabricated may approach digital evidence with the same healthy skepticism they'd apply to any other testimony — weighing it, rather than simply believing it. That kind of education might be as urgent as any rule change.
There's also the self-represented litigant problem. Self-represented parties cite nonexistent cases, statutes, or quotations generated by AI tools far more frequently than lawyers do — over 350 documented cases in the US alone. Legal professionals aren't entirely exempt either, with more than 200 instances of false citations from AI. The system needs to account for the fact that most people using it don't have a lawyer who can catch these things.
We've spent decades building legal safeguards around the fallibility of human memory. We're now at the beginning of a longer, harder project: building safeguards around the fallibility of perception itself. Just as we must work to reclaim our focus in a world filled with digital distraction, we need to actively build AI literacy to navigate this new courtroom reality.
The Mendones deepfake was caught because it used obsolete technology. Future fakes won't be so obvious.