In January 2026, a federal court in the Western District of Texas logged the most consequential deepfake to reach an American courtroom so far, and almost nobody in the room argued about it. The file was offered as audiovisual evidence. It helped support an indictment. No one challenged it as synthetic. The fabrication surfaced only at the sentencing hearing of the confidential informant who produced it, after he pleaded guilty to obstruction of justice. The person the fake helped indict had already had the case dismissed, without explanation.
The case is United States v. Hernandez, No. 3:25-cr-2384, and it turns up in the Federal Judicial Center’s 2026 survey of federal judges as the survey’s most striking entry. That is the detail worth sitting with. Every dramatic deepfake story in the press involves someone spotting the tell: the frozen face, the looping footage, the mouth that does not match the words. The Hernandez case had no spotter. It had a plea deal, roughly two years later.
The deepfake that worked
Compare that with the case everyone cites. In Mendones v. Cushman & Wakefield, a housing dispute in Alameda County, California, Judge Victoria Kolakowski noticed something off about two video exhibits purporting to be testimony from a witness named Geri Haas. The face barely moved. The cadence was monotone. The footage looped. She compared the exhibits against a third video of the same witness already in the record and concluded the “person” in the first two was not the same human being.
On September 9, 2025, she struck the second amended complaint and dismissed the case with prejudice, writing that the sanction would send a message of “zero tolerance with attempting to pass deepfakes as evidence.” It appears to be one of the first instances in which a court affirmatively identified AI-generated video submitted as genuine testimony, and it has become the example everyone reaches for.
Both stories are true. They point in opposite directions, and the difference between them is not the quality of the fake. It is who was looking, and whether anyone had a reason to look.
What authentication actually asks (and why the answer is “probably”)
Federal Rule of Evidence 901(a) is the door every exhibit walks through: the proponent must produce evidence “sufficient to support a finding that the item is what the proponent claims it is.” That is a low bar on purpose. Authentication is a screening function, not a truth-finding one, and for most of its history the screening worked because faking a document left fingerprints that faking a video no longer does.
The doctrinal wrinkle that matters is who decides. Under the traditional reading, authenticity is a question of conditional relevance for the jury under Rule 104(b). The judge asks only whether a reasonable juror could find the item authentic; the fight over whether it actually is belongs to the jury. The draft federal rule on deepfakes would move that fight to the judge and raise the standard to a preponderance. That shift, not the word “deepfake,” is the real innovation.
Two other rules do more work in practice than most people realize, and neither solves the problem. Rule 902(13) and 902(14) let a proponent certify the integrity of a file as captured through a hash, a process, and a custodian. Those certifications say something about whether the file in court is the file that was seized. They say nothing about whether the events in that file ever happened. A perfectly preserved deepfake certifies beautifully.
Two studies, one uncomfortable conclusion
If authentication is a human judgment, the obvious question is how good the humans are. Two lines of research from the past two years answer it, and neither answer is flattering.
The first is courtroom-specific. A study submitted to the Advisory Committee on Evidence Rules by Justin Sevier, Henry Zhuhao Wang, and Gino Cerrone tested 1,783 people, including judges, lawyers, law students, and laypeople, on 28 authentic and AI-generated images, videos, audio clips, and text passages. Overall accuracy was 47.4%. That is worse than a coin flip. Legal training conferred no advantage. Participants were at chance on images and audio and below chance on video and text, and their confidence was badly calibrated: the modality they judged worst was the one they felt best about.
The second is technical, and it is worse. A 2026 benchmark called DF26, built by S. Barrington, M. Bohacek, and Hany Farid, assembled 271 real videos and 2,420 synthetic ones from seven modern text-to-video and image-to-video models. Human participants identified the fakes with 52.6% accuracy on the DF26 benchmark, barely above chance, while performing fine on the real videos at 76%. State-of-the-art automated detectors dropped to near chance on the same set. On older benchmarks, humans and machines score in the 70s and 80s. On current generators, the tells that made those benchmarks easy no longer exist.
So the Mendones outcome was not proof that judges can spot deepfakes. It was proof that a judge can spot bad deepfakes, in a case where the fabricator left the face frozen and the audio flat. The gap between a bad deepfake and a good one is now a subscription tier.
Washington has a draft rule. It has had one for three years.
The Advisory Committee on Evidence Rules has been working on this since 2023, across at least seven meetings. Its working draft, now styled Rule 901(c), does two things in sequence. First, the party challenging the exhibit must put forward evidence sufficient to support a finding that it was fabricated by generative AI. A bare “that’s a deepfake” does not earn an inquiry. Second, if the challenger clears that threshold, the proponent must prove authenticity by a preponderance of the evidence, which means the judge, not the jury, decides whether the exhibit survives.
As of the Committee’s May 2026 report to the Standing Committee, the rule is not going anywhere. The Committee declined to recommend an amendment to Rule 901, reasoning that “few examples exist of courts having to address the possibility of deepfakes,” and kept the draft in abeyance so it can move quickly if that changes. A companion proposal, Rule 707, which would treat machine-generated evidence more like expert testimony, was also held back for further study.
That decision rests on a survey number that has been quoted ever since. The Federal Judicial Center polled every federal district, magistrate, and bankruptcy judge. Of 931 who responded, a 45% response rate, fifteen had encountered a deepfake challenge at all, roughly 2%. Most of those fifteen had seen exactly one, in 2024 or 2025, usually in a civil case.
I think that figure is being read backwards. “Only fifteen judges” is treated as evidence the problem is small. It is equally consistent with a problem that resolves itself before an opinion is ever written: a plea, a withdrawal, a settlement, an exhibit quietly dropped. The Hernandez prosecution produced no reported authentication ruling. It produced a dismissed indictment.
What judges say they would actually demand
The same Federal Judicial Center survey asked the 914 judges who had never seen a deepfake challenge what they would require before entertaining one. The answers are more useful than the headline count, because they describe a threshold courts are already improvising.
| Forum | What triggers a closer look | Standard once triggered | Status |
|---|---|---|---|
| Federal courts, proposed Rule 901(c) | Challenger presents evidence sufficient to support a finding of AI fabrication | Proponent must show the item is more likely than not authentic; judge decides | Drafted and held in abeyance; not recommended for amendment as of May 2026 |
| Montana, proposed Rule 901 amendment | Challenger shows a reasonable factfinder could find fabrication | Proponent must show it is more likely than not authentic | Revised after public comment on 5 October 2026; pending |
| New York, Matter of M.S. (M.H.) | No special trigger; deepfake risk folded into the ordinary foundation analysis | Proponent must lay a sufficient foundation; ambiguity is no longer automatically a weight question | Court of Appeals, 17 February 2026, 4–3 |
| Iowa, Amyda | Speculation about AI is not enough on its own | Circumstantial evidence can authenticate; unsupported deepfake claims go to weight | Court of Appeals decision |
Sources: Advisory Committee on Evidence Rules reports (May 2026 and October 2026); the Montana proposal as amended; and the Iowa State Bar’s practitioner analysis. Statuses current as of October 2026.
Within that survey, 679 judges wrote out what the threshold should look like. About 42% would accept a minimal showing, such as an affidavit or technical evidence pointing toward fabrication. Another 35% wanted a good-faith, articulable reason beyond speculation. Only 3% would require qualified expert testimony before the inquiry even began, and another 3% wanted sworn statements.
A written affidavit, a specific visual anomaly, an alibi, or a technical inconsistency clears the bar in most federal courtrooms. A retained expert does not. That is a lower trigger than some defense lawyers assume and a higher one than some plaintiffs fear.
New York did not wait for a rule
On 17 February 2026, the New York Court of Appeals reversed a child-protection finding in a 4–3 decision that had nothing to do with AI on its face and everything to do with it underneath. In Matter of M.S. (M.H.), 2026 NY Slip Op 00825, the only evidence of abuse was three video clips that had surfaced on the computer of a Syracuse man under investigation for trading child sexual abuse material, who said he had hacked the family’s security cameras. The clips were excerpts, not a continuous feed. Nobody explained what had been selected or what had been deleted. The alleged creator of the recordings never testified, and neither did the hacker; his account came in as hearsay through an FBI agent. Roughly two and a half years passed between the clips being taken and their recovery, a gap the majority called 65 times as long as the one that sank the video in People v. Patterson.
The county’s case rested on two things: an investigator matching the living room furniture to the room in the clips, and an agent saying he saw no signs of tampering. The agent had never been qualified as an expert in video authentication and was asked one leading question on the subject. The court held the foundation legally insufficient and dismissed the petitions.
The deepfake language, though, is what will be quoted for a decade. Citing the risk that fabricated videos now “leverage” real details of real places, the majority wrote that matching circumstantial details to personal observation has become “a more suspect form of authentication.” That is a genuinely new proposition in a state that had treated video authentication as a low, forgiving threshold.
My read is that the majority got the outcome right and the reasoning partly wrong. These videos failed for unglamorous reasons: hearsay, a broken chain of custody, a witness nobody bothered to qualify. You do not need generative AI to exclude them. The deepfake discussion is, strictly speaking, unnecessary to the holding, and it now functions as a thumb on the scale in every New York foundation fight.
The dissents make the strongest version of the counterargument, and they deserve to be taken seriously. Judge Troutman, joined by Judges Garcia and Singas, pointed out that nobody said the word “deepfake” in Family Court and that the theory appeared for the first time on appeal. Her objection is structural: a party can now posit a fake with no evidence at all and force the proponent to disprove it, and she asked where a county child protective agency is supposed to find affordable detection experts. Judge Singas argued separately that the fabrication the majority feared was not technically feasible in 2019 or 2022.
The quieter threat: the deepfake defense
Here is where I part company with most of the commentary. The deepfake that fools a court is terrifying, but it is still rare, and it is still a crime to offer one. The deepfake defense is neither rare nor self-punishing. It costs a sentence to assert and it attaches to real evidence, which is exactly the evidence most likely to convict.
Researchers call this the liar’s dividend: the same technology that makes fakes cheap makes it cheap to call true things fake. In a 2023 Tesla crash lawsuit, defense counsel tried, unsuccessfully, to knock out a video by arguing it could be a deepfake. That is the template. Defense attorneys have already begun building strategy around it.
The human cost is not hypothetical. In Florida, a woman spent two days in jail after her ex-boyfriend allegedly fabricated AI-generated text messages that led to her arrest for violating a protective order. Prosecutors dropped the charges eight months later. As the National Center for State Courts put it, no one verified the evidence before it was acted on.
The proposed burden-shift is, whatever else it does, a partial answer to exactly that problem. It forces the challenger to say something specific before the court spends a week on it. That is why I think the structure is right even though the rule is stalled.
Provenance is a paper trail, not a polygraph
The structural fix that gets the most attention is provenance. The C2PA content credentials standard attaches cryptographically signed metadata at the point of capture, recording the device, the edit history, and any AI involvement. Some forensic hardware now signs at capture, so a fingerprint image or crime-scene photo carries a verifiable origin. If the chain is intact, a court can trace an exhibit to its source instead of relying on a sponsoring witness’s memory.
Do not treat it as a cure. An independent security analysis published in April 2026 by Enis Golaszewski and coauthors found that current C2PA specifications fail to achieve their claimed security goals, including weak file integrity and inconsistent validation. The authors’ conclusion is blunt: C2PA is promising, but should not yet be relied on for high-stakes uses such as legal evidence. Absence of credentials also does not make an exhibit inadmissible, and their presence is not conclusive. It is one input.
The practical consequence is boring and important. Ask for the original file, not the copy someone re-exported through a messaging app. Provenance metadata is the first thing stripped by a screenshot, a re-upload, or a group chat.
What I would tell a lawyer walking in
If you are offering the exhibit, the foundation that survives 2026 looks different from the one that survived 2016. Produce the complete recording, not a clip. Document the chain of custody by name and date. Put on a witness who can speak to the capture device itself, not just to what the video depicts. Ask that witness about tampering directly and specifically, because the New York case turned on an agent who was never asked. And if content credentials exist, preserve them in the native file.
If you are challenging the exhibit, bring an anomaly. A timestamp that jumps. A shadow that runs the wrong way. A reflection that should not be there. A location the depicted person provably was not. The judges in the survey were clear that a bare assertion earns nothing.
If you need a first-pass read before deciding whether to spend money, a free AI detector will give you a signal in seconds, and that is all it is: a lead. Given that published detectors fall to near chance on current generation models, a score is a reason to look closer, not a finding you can hand a court.
What would change my mind
Two things. If reported opinions start showing deepfake challenges as routine rather than exceptional, the case for amending Rule 901 becomes much stronger than fifteen judges suggest. And if the Federal Judicial Center re-runs its survey in two years and the count is still in the teens, then the Committee was right to wait, and the burden-shift I favor is a solution to a problem that has not arrived.
What would not change my mind is another demonstration video. The technology’s capability has never been the bottleneck. The bottleneck is a doctrine built for a world where seeing was believing.
FAQ: what litigants actually ask
Can a judge exclude a deepfake from evidence?
Yes. A fabricated exhibit is not what its proponent claims it is, so it fails authentication under Rule 901 regardless of whether a special deepfake rule exists. Courts have excluded video on ordinary foundation grounds for decades. The disputed question is not whether they can exclude, but how much evidence a challenger must produce before the court has to engage.
Do you need an expert witness to challenge a video as AI-generated?
In most federal courtrooms, no. Only about 3% of judges surveyed by the Federal Judicial Center said they would require qualified expert testimony at the threshold. Most would accept an affidavit, technical evidence pointing to fabrication, or a specific factual reason to doubt the exhibit. An expert becomes far more important once the challenge is actually litigated.
Does missing metadata mean a video is fake?
No. Provenance credentials are stripped routinely by messaging apps, screenshots, and re-exports, and older or cheaper devices never generated them in the first place. Missing credentials shift weight toward forensic examination and corroboration, but they do not establish fabrication and do not make an exhibit inadmissible.
What is the “deepfake defense”?
It is the practice of attacking authentic evidence by suggesting it might be synthetic. Because asserting it is cheap and disproving it is expensive, it functions as a doubt-generating tactic even when it fails. The proposed Rule 901(c) threshold, and the notice provisions discussed alongside it, are aimed partly at this problem.
Can a lawyer be sanctioned for submitting AI-generated evidence?
Yes, and state courts have started doing it. In the Alameda County case, the court struck the complaint and dismissed with prejudice under California’s sanction statute, having found that two exhibits were products of generative AI. On the criminal side, the Texas informant who supplied fabricated audiovisual material pleaded guilty to obstruction of justice.
Do the federal evidence rules govern state court?
No. Each state has its own evidence code, and the federal draft rule binds no state court. Montana is currently considering a similar amendment to its own Rule 901, and New York has effectively raised its authentication bar through case law instead of rulemaking. Whether a deepfake challenge clears the threshold can therefore turn entirely on which courthouse you are standing in.
How this article was put together
I worked from primary documents rather than summaries: the Advisory Committee on Evidence Rules reports from May and October 2026, the committee’s own draft rule language, the Federal Judicial Center survey results as reported by both the committee and practitioner analyses, the Alameda County sanctions order, and the New York Court of Appeals opinion in Matter of M.S. (M.H.). The two accuracy figures come from the underlying studies, including the DF26 benchmark paper, not from press coverage of them. The C2PA security assessment is an independent academic analysis, not an industry release. Everything here was checked in October 2026. Rule statuses and the Montana proposal will need rechecking before the next rules cycle.