Deepfake evidence in court: how judges actually decide what is real

Have deepfakes actually appeared in real courtrooms yet?

Yes, though documented incidents are still scarce. The most cited example is Mendones v. Cushman & Wakefield in Alameda County Superior Court, where a judge detected a deepfake video exhibit that had been offered as evidence, making it among the first documented courtroom deepfake incidents ^[2]. The case matters less for its particular facts than for what it demonstrated: fabricated audiovisual exhibits are no longer a law review hypothetical. A synthetic video reached a real judge in a real civil matter, and it was a human on the bench, not an automated scanner, who caught it.

That detail deserves emphasis, because it cuts against the intuition many people bring to the subject. The instinct is to assume that fake video will be defeated by better fake-video detectors, an arms race of algorithms. The first documented incident points elsewhere. A judge exercising ordinary judicial skepticism, aided by the specifics of the case, was the safeguard that worked. That is consistent with how courts have long policed authenticity, and it is a useful anchor for everything that follows: the courtroom response to synthetic media is being built on top of existing evidentiary habits, not invented from scratch.

The rarity is confirmed by the Federal Judicial Center's own numbers, with only around 2% of surveyed federal judges reporting they had ever seen a deepfake challenge ^[1]. But two pressures are converging. Generation tools keep getting cheaper and more convincing, which raises the odds of fabricated exhibits. And the same technology invites what commentators call the opposite abuse: attacking genuine recordings as fake in the hope of manufacturing doubt. Courts appear to be preparing for both failure modes at once, which is precisely why the showing they demand from challengers has become the central question.

Scarcity today is not a forecast for tomorrow. The value of understanding the courtroom rules now is that a firm can build habits before the volume of disputes rises, rather than after.

What the Federal Judicial Center's 2026 survey of 931 judges found

The Federal Judicial Center survey, released March 25, 2026, drew responses from 931 federal judges, roughly a 45% response rate, according to reporting by Esquire Deposition Solutions ^[1]. Two findings stand out. First, actual encounters remain rare: only about 2% of responding judges had ever faced a deepfake evidence challenge. Second, and more consequential for practitioners, roughly four in five judges said they would require a specific initial showing before entertaining such a challenge at all. In other words, the federal bench has largely settled on a gatekeeping posture before the wave arrives.

A response rate near 45% is worth pausing on. It is high enough to take the results seriously as a signal of judicial attitude, while not so complete that the figures should be read as a precise census of every federal judge's view. The honest reading is directional: a large, deliberately surveyed group of federal judges converged on a gatekeeping instinct, and that convergence is the news, more than any single decimal point. Practitioners who cite the survey should cite it for its direction and its documented figures, not for a precision it does not claim.

That posture is worth reading carefully. Judges are not saying deepfake concerns are illegitimate; they are saying the price of admission is specificity. The examples judges pointed to, such as a demonstrable file anomaly or an external fact that contradicts the recording, are concrete and checkable ^[1]. This mirrors how courts have historically handled other authenticity disputes: generalized suspicion does not reopen the evidence, particulars do. For law firms, the practical translation is that a deepfake objection is an evidentiary project requiring preparation, not a rhetorical move to deploy at the podium. The survey findings are best traced to the FJC report itself, but the reported figures are consistent on these points.

The two failure modes courts are guarding against at once

Fabricated exhibits and opportunistic denial of genuine ones are opposite abuses of the same technology. A specific-showing requirement is the single filter that constrains both.

It helps to name the two problems separately, because they call for different responses even though a single record can address both. The first failure mode is the obvious one: a fabricated exhibit offered as if it were real, the synthetic video or audio that never depicted a true event. The Mendones incident is the documented example of this mode reaching a courtroom ^[2]. The second failure mode is quieter and, in the long run, may be more corrosive. It is the tactic of attacking a genuine recording as though it were fake, exploiting general awareness that fabrication is now possible in order to manufacture reasonable doubt about authentic evidence ^[1].

The reason judges have converged on a specific-showing requirement is that the requirement disciplines both modes at the same time. If a party must point to a file anomaly or a contradicting external fact before a challenge is entertained, then a genuine recording cannot be casually smeared as fake, and a fabricated exhibit still faces scrutiny once a challenger produces the particulars. The gatekeeping posture is not hostility to authenticity concerns; it is a filter that lets substantiated concerns through and holds unsubstantiated ones back. Understanding that dual purpose explains why the survey result is stable rather than accidental: a specificity requirement is the natural equilibrium for a court that must protect against forgery and against opportunistic denial in the same breath.

The showing judges require before taking a deepfake challenge seriously

What does a "specific initial showing" look like in practice? Based on the examples surfaced in the FJC survey reporting, judges want something they can examine rather than an atmosphere of doubt ^[1]. A bare "that could be a deepfake" objection, without more, is the kind of challenge four in five judges indicated they would not entertain. The specifics that appear to move the needle are the ones that connect to verifiable facts about the file or the world it claims to depict. Counsel preparing a challenge would be wise to assemble items such as:

Notice what this list rewards: documentation that exists independently of the disputed exhibit. Every item is easier to establish when reliable records were being kept before the dispute began. That is why the deepfake problem, examined closely, keeps collapsing into a chain-of-custody and provenance problem. The party who can show where a file came from, when it existed, and whether it changed holds the specifics judges are asking for. The party relying on memory and assertion does not. None of this is legal advice, and the right showing varies by court and posture, but the direction of judicial expectations is clear.

It also helps to see how these items reinforce one another. A file anomaly on its own can be explained away as an artifact of ordinary re-encoding or transfer. An external contradiction on its own can be attributed to mistaken recollection. But an anomaly that lines up with an external contradiction, and with a gap in the chain of custody, begins to look like a pattern rather than a coincidence, and a pattern is what persuades a court to take a closer look. Expert analysis then does its proper work, which is to explain and connect those particulars for the finder of fact, not to substitute a confident conclusion for the particulars themselves.

• A demonstrable file anomaly, for example metadata, encoding, or edit artifacts inconsistent with the claimed origin ^[1]
• An external fact that contradicts the recording, such as reliable evidence the depicted person was elsewhere ^[1]
• Chain-of-custody gaps between the claimed capture and the exhibit produced in discovery
• Qualified expert analysis, offered as support for the specifics rather than a substitute for them

How to prepare a deepfake challenge that meets the bar

The survey describes what judges want; it does not walk counsel through assembling it. The following sequence is a way to translate the gatekeeping standard into a workflow, drawn from the same specifics the FJC reporting surfaced ^[1]. It is general information rather than legal advice, and the order and emphasis will shift with the jurisdiction and the posture of the case.

The through line across every step is that the challenge should rest on things a court can inspect. When the record is built this way, the objection stops being a rhetorical gamble and becomes an evidentiary submission with a spine. That is the difference between a challenge four in five judges would wave off and one they would agree to examine.

1. Fix the claimed originPin down exactly what the proponent says the exhibit is, who captured it, when, and on what device, so any anomaly can be measured against a stated baseline rather than a vague impression.
2. Examine the file's own recordLook for metadata, encoding, or edit artifacts that are inconsistent with the claimed origin. A concrete, describable anomaly is the kind of particular judges said they would examine ^[1].
3. Find the external contradictionIdentify reliable, independent facts at odds with what the exhibit depicts, such as evidence the person shown was elsewhere. External contradictions are among the specifics the survey highlighted ^[1].
4. Map the chain of custodyTrace the exhibit from claimed capture to production in discovery and document any gaps. Missing links convert a hunch into an articulable weakness a court can weigh.
5. Frame expert analysis as supportRetain qualified analysis to explain and connect the particulars, not to replace them. Offered as reinforcement for concrete specifics, expert work carries more weight than a standalone opinion.

This sequence is a preparation aid, not a rule of procedure. Consult a licensed attorney about how to raise an authenticity challenge in a specific court.

How Rule 901 and authentication law are evolving

Federal Rule of Evidence 901 sets a modest threshold: the proponent must produce evidence sufficient to support a finding that the item is what the proponent claims it is. That standard was built for a world where forging a convincing video was expensive and rare. Whether it still fits is now under active discussion. Professor Rebecca Delfino has proposed amending Rule 901 to address deepfake authentication, in a suggestion submitted to the Advisory Committee on Evidence Rules ^[3]. Proposals like this contemplate recalibrating who must show what, and when, once a genuine question of synthetic fabrication is raised.

It is worth being precise about what a proposal is and is not. A suggestion to an advisory committee is the opening of a deliberative process, not a change in the law. The existing 901 threshold governs today, and it is deliberately low, designed to send most authenticity questions to the finder of fact rather than resolve them at the threshold. The interesting question the Delfino suggestion raises is whether synthetic media strains that design, because a low bar to admission assumes that fabrication is hard and therefore uncommon ^[3]. As that assumption weakens, the pressure to adjust the burden architecture grows, which is why the proposal exists at all.

As of this writing, no deepfake-specific amendment has been adopted, and courts continue to work within the existing framework, so practitioners should verify the current state of the rule before relying on any proposed change. But the trajectory matters for planning. If the burden architecture shifts, parties who can affirmatively document the origin and integrity of their exhibits will be positioned for whatever showing the amended rule requires, while parties who cannot will be negotiating from weakness. The FJC survey suggests judges are already applying that intuition informally ^[1]: they want specifics from challengers, and they will predictably want credible provenance from proponents in return.

Why the deepfake problem is really a chain-of-custody problem

Authentication has always turned on an item's history. Synthetic media raises the stakes of that inquiry without changing its shape, which is why provenance and custody remain the center of gravity.

Strip away the novelty of synthetic media and a familiar structure appears underneath. Courts have always had to decide whether a document, a photograph, or a recording is what its proponent says it is, and they have always resolved those questions by looking at where the item came from and how it was handled. Deepfakes raise the stakes of that inquiry, because a convincing fake is easier to make than it used to be, but they do not change its shape. The question a court asks about a suspicious video is the same question it asks about a contract with a disputed signature: can you show me its history?

That is why the specifics judges reward map so cleanly onto custody and provenance ^[1]. A file anomaly is a statement about how the file was produced or altered. An external contradiction is a statement about whether the depicted event could have happened as shown. A chain-of-custody gap is a statement about who touched the file and when. Each of these is a question about the record surrounding the exhibit, not merely the pixels inside it. Counsel who treat authentication as a records discipline, established well before any dispute, are working with the grain of how courts actually decide. Counsel who treat it as a last-minute forensic scramble are working against it.

Why provenance records beat after-the-fact deepfake detection

Detection tools analyze a file after a dispute exists and render a probabilistic opinion: this recording appears synthetic, or appears genuine, to some confidence level. That has real value, and nothing here suggests detection is useless. But it arrives late, it can be contested by an opposing expert running a different model, and it examines only the artifact, not its history. Provenance records work the other way around. A record created and sealed at capture or creation time, anchored to a tamper-evident, post-quantum-signed transparency log, documents what existed and when, before anyone had a motive to fabricate or to cry fake ^[1].

The contrast is easiest to see side by side. Detection is retrospective, probabilistic, and contestable, and it speaks only to the file in front of it. Provenance is contemporaneous, documentary, and independently checkable, and it speaks to the file's origin and integrity over time. The two are not mutually exclusive, and a well-prepared party may use both. But they occupy different positions in a dispute, and the survey's emphasis on specific, checkable facts favors the kind of evidence that was fixed before the fight began ^[1].

That is exactly the shape of evidence the FJC survey says judges reward: specific, checkable facts rather than dueling impressions ^[1]. A sealed provenance trail gives counsel a contradicting external fact when an exhibit is forged, and an affirmative foundation when a genuine exhibit is attacked. RankShield Legal provides that provenance layer for law firms, sealing records to an independently verifiable log at the moment they are created. To be direct about the limits: provenance records strengthen an authentication showing, but they do not guarantee admissibility, and no vendor can. Courts decide authenticity. The goal is to walk in with proof of what existed when, instead of arguing about it afterward.

The limits: what provenance can and cannot do

It would be easy, and wrong, to present provenance records as a solution to the deepfake problem. They are not. A provenance trail improves the quality of an authentication showing; it does not remove the court from the decision. Under Rule 901 and the practice built around it, authenticity is a question for the judge and, ultimately, the finder of fact, and no record, no seal, and no vendor can pre-decide it ^[3]. A firm that oversells provenance to a client sets up a disappointment that undermines the very credibility the records were meant to build.

The honest description is narrower and more durable. Provenance gives counsel the specific, checkable facts that the FJC survey says judges look for, on both sides of the two failure modes: a contradicting external fact when a fabricated exhibit is offered, and an affirmative foundation when a genuine one is attacked ^[1]. It shifts the argument from unaided memory and assertion toward a documented record that existed before anyone had a reason to shape it. That is a meaningful advantage, and it is worth stating plainly precisely because it is bounded. The value is in walking into a dispute already able to show what existed and when, not in a promise that the court's judgment can be bypassed.

None of the foregoing is legal advice. It is a description of how the available reporting says judges are approaching synthetic media, and how records created at the moment of capture fit the specifics those judges reward. The right showing for any particular exhibit depends on the jurisdiction, the posture, and facts a licensed attorney is positioned to assess.

Provenance records strengthen an authentication showing. They do not guarantee admissibility, and courts, not vendors, decide authenticity.