Federal Rule of Evidence 901 and its state counterparts have long governed how litigants authenticate the photographs, recordings, and documents they seek to introduce at trial.[1] Those rules ask a straightforward question: has the proponent offered evidence “sufficient to support a finding that the item is what the proponent claims it is”?[2] That standard was written for a world of negatives, tape recorders, and paper trails—not for a world in which a few minutes of publicly available audio can be transformed into a wholly fabricated recording of a person’s voice, or in which a blurry surveillance video can be “enhanced” by an algorithm into images that never actually existed.
As deepfakes and other AI-generated media become easier and more convincing to produce, courts and attorneys alike are forced to rely on a decades-old authentication framework to determine whether evidence has been modified by artificial intelligence.
The term “deepfake” refers broadly to synthetic images, video, or audio that have been generated or manipulated using artificial intelligence to appear authentic. The danger deepfakes pose to the litigation process is not merely theoretical. Convincing fabricated evidence risks being admitted before a factfinder who has no reason to doubt it. Courts across the world have already begun to confront this issue, and these early decisions offer a useful roadmap.
In perhaps the most egregious case involving deepfake evidence, a California court in Mendones v. Cushman & Wakefield, Inc. reviewing the pro se plaintiffs’ motion for summary judgment grew suspicious of video “testimonials” submitted from certain witnesses.[3] At a hearing, the plaintiffs conceded that some of the individuals depicted in the videos were either deceased or could not be located.[4]
The court then ordered the plaintiffs to produce metadata for the videos, including file format, creation and modification dates, the device and lens used, and shutter speed, and ultimately held a show-cause hearing on whether terminating sanctions—or even a referral for criminal prosecution—were warranted.[5] The court imposed terminating sanctions, explaining that deepfake evidence compromises the court’s fact-finding function, damages public trust in the judicial process, and forces already under-resourced courts to serve as forensic gatekeepers.[6]
Mendones illustrates the practical signs that should put counsel on notice: deepfake videos sometimes display an unnatural lack of facial expression, a monotone delivery, missing pauses where a real speaker would naturally pause, and an overall robotic quality, while fabricated text evidence can often be identified through inconsistencies in font style, font size, spacing, or sharpness. But as technology advances, even these indicia are becoming harder to spot.
Deepfakes are not confined to video evidence. In a recent UK family court proceeding, a mother sought to limit her husband’s access to their children by offering an audio recording in which he purportedly threatened her during a phone call.[7] The father denied ever making the statements. It later emerged that the mother had used widely available editing software and online tutorials to splice and alter an authentic recording of the father’s voice, inserting threatening language he had never actually spoken.[8]
The court excluded the recording after the parties recovered the original audio file and its metadata confirmed the fabrication.[9] The case offers a useful checklist for any litigator confronting suspicious audio or video: begin from a position of skepticism rather than assuming authenticity simply because a recording “sounds real.” Litigators should therefore independently verify the surrounding facts, including timestamps and phone records, retain a digital forensic expert, and consider whether the party offering the evidence had unusual access to source recordings of the person supposedly depicted.
Not every authentication problem stems from outright fabrication, however. Sometimes the underlying recording is genuine, and it is a party’s attempt to enhance that recording with artificial intelligence that raises the concern.
In Washington v. Puloka, a bystander’s cell phone captured ten seconds of a shooting, but the resulting video was low-resolution and blurry.[10] The defense sought to introduce an AI-enhanced version of that same video, created by an expert who used AI video-editing software to sharpen the footage, smooth its edges, and add definition to the objects depicted.[11] The prosecution’s rebuttal expert testified that the AI process had not merely clarified the video—it had invented image detail that was never captured by the camera, altering shapes and stripping away the very artifacts that forensic analysts rely on to authenticate a recording.[12] The trial court excluded the enhanced video under three separate evidentiary rules.
Under Rule 702, the court found that AI-based video enhancement is not generally accepted within the forensic video community and lacks support in peer-reviewed literature.[13] Under Rule 401, the court reasoned that the enhanced video did not reliably depict what actually happened, but rather what the AI model predicted should be shown.[14] And under Rule 403, the court concluded that admitting the video risked confusing the jury, inviting a “trial within a trial” over the reliability of the AI tool, and unfairly overshadowing the eyewitness testimony given the enhanced video’s visual clarity.[15]
Puloka serves as a warning that AI-enhanced evidence—even when it originates from an authentic source—can distort the very evidence it purports to clarify, and that litigators confronted with such evidence should immediately consult a forensic expert to assess its reliability.
The issues posed by fabricated audio also extend beyond the courtroom. In 2019, cybercriminals used a cloned voice recording of a German parent company’s chief executive—complete with his distinctive accent and speech patterns—to convince the chief executive of its UK-based subsidiary to wire $243,000 to a fraudulent supplier over the course of three separate phone calls.[16]
The scheme succeeded in part because the executive overlooked a telling detail: the calls originated from an Austrian phone number, even though his boss was based in Germany.[17] The same technique could just as easily be used against a law firm—for example, a fraudster spoofing an opposing counsel’s voice to leave a voicemail requesting a wire transfer of settlement funds or a client’s personally identifiable information.
As voice-cloning and video-generation tools continue to improve, attorneys should treat any unexpected request to wire funds or disclose confidential information with heightened scrutiny and should independently call back a known number to confirm the request before acting on it.
The Federal Rules of Evidence and their state counterparts will need to be revisited to address a category of evidence their drafters could not have anticipated. In the interim, courts and litigators must remain vigilant and utilize a consistent set of practices: treat unusually clean, convenient, or favorable evidence with skepticism, retain a qualified forensic expert as soon as authenticity is in doubt, press both your own experts and opposing experts on any use of artificial intelligence in forming their opinions, and preserve a clear record of the inputs behind any AI-assisted work product.
As deepfake technology becomes more sophisticated and more accessible, the burden of separating real evidence from synthetic evidence will increasingly fall on trial counsel long before it ever reaches a judge.
[1] Fed. R. Evid. 901; see also R.I. R. Evid. 901; Mass. G. Evid. § 901.
[2] Fed. R. Evid. 901(a).
[3] Mendones v. Cushman & Wakefield, Inc., No. 23CV028772, 2025 WL 2613764 (Cal. Super. Ct. Sept. 9, 2025).
[4] Id. at *1.
[5] Id. at *2.
[6] Id. at *5-7.
[7] Deepfake Audio Evidence Used in UK Court to Discredit Father, Cyfor Forensics Digital Evidence (last visited Sept. 16, 2026), https://cyfor.co.uk/deepfake-audio-evidence-used-in-uk-court-to-discredit-father/.
[8] Id.
[9] Id.
[10] Findings of Fact and Conclusions of Law RE: Frye Hearing on Admissibility of Videos Enhanced by Artificial Intelligence at 2, Washington v. Puloka, No. 21-1-045851-2KNT (Wash. Super. Ct. Mar. 29, 2024).
[11] Id. at 3.
[12] Id. at 3-4.
[13] Id. at 5-6.
[14] Id. at 6.
[15] Id.
[16] Jesse Damiani, A Voice Deepfake Was Used to Scam a CEO Out of $243,000, Forbes (Sept. 3, 2019), https://www.forbes.com/sites/jessedamiani/2019/09/03/a-voice-deepfake-was-used-to-scam-a-ceo-out-of-243000/.
[17] Id.
About The Author
Lucas G. Spremulli
Lucas is an associate with Adler Pollock & Sheehan and a member of the firm’s litigation group. A trial lawyer admitted to…
