HOW DEEPFAKES ARE REWRITING TRUTH IN AMERICA’S COURTROOMS
Author: Jason Gerstorff, Senior Editor
Imagine yourself in a custody dispute with a former spouse, and they offer a video of you engaged in shocking behavior, sure to sway the court in their favor. You know the video is fabricated, and claim it is an AI deepfake. Or perhaps the video is real, and you falsely claim it is a deepfake, hoping to fool the court.[i] This dilemma has been coined “the liar’s dividend.”[ii] A version of this scenario happened in the 2021 homicide trial of Kyle Rittenhouse.[iii] During the trial, the defense counsel objected to the prosecution’s pinch-to-zoom feature, arguing that Apple uses artificial intelligence to “create what [the algorithms] believe is happening.”[iv] The trial judge ultimately sided with the defense, ruling that the prosecution had the burden of proving the footage was not altered.[v] The prosecution team could not do so, and after the footage was shown on a large television instead, Rittenhouse was acquitted of all charges.[vi] How are courts and parties to deal with fabricated evidence and “the liar’s dividend?”
Under the Rule 901 standards requiring only that a proponent produce evidence sufficient to support a finding that the video is what they say it is, courts and juries may believe deepfake videos are authentic just because they have met the fairly relaxed evidentiary standard, leading to inaccurate trial results.[vii] The authentication of audio evidence, for example, could be satisfied using an opinion identifying a person’s voice.[viii] Video may require a witness present when the video was made or verification that the video came from a surveillance camera.[ix] The rule is supplemented by Rule 902, which allows certain categories of digital evidence to be self-authenticating.[x] Forensic computer experts have ability to verify that electronic data has not been altered by someone, using advanced cryptographic algorithms.[xi] However, these advanced methods of authentication do not guarantee a video was not AI-generated from the start, only that it was not modified after its inception.
The first danger from the hypothetical custody dispute above is the admission of fake evidence into trial. In USA v. Khalilian the defense moved to exclude a voice recording of the defendant, arguing the recording could be a deepfake.[xii] The court concluded that verification from a witness familiar with the defendant’s voice was “probably enough to get it in.”[xiii] But is this level of scrutiny enough to adjudicate the hyper-realism of deep-faked audio? Professor Delfino argues no, stating that these deepfakes do not merely distort reality but mimic individuals with near-perfect accuracy, necessitating additional proof of reliability beyond the requirements of Rule 901.[xiv]
In the reverse scenario, real evidence gets challenged without basis. In Huang v. Tesla, discovery revealed a video of Elon Musk making statements about the safety of Tesla’s Autopilot.[xv] Tesla refused to admit the authenticity of the video, seemingly without any legitimate basis for doing so, by claiming that Musk could be targeted for deepfakes due to his high profile.[xvi] While the court reproached the refusal, this “deepfake defense” allows lawyers to raise questions, objections, and arguments to challenge even genuine evidence in bad faith.[xvii] The seeds of doubt planted in jurors’ minds tends to stain the authenticity of other genuine evidence.[xviii]
The Advisory Committee on Evidence Rules has been actively studying these issues. Judge Paul Grim and Dr. Maura Grossman’s proposal for Rule 901(c) would allow judges to revoke the jury’s duty of determining the authenticity of synthetic evidence if there were a real controversy between the parties.[xix] Professor Rebecca Delfino offered a stricter proposal requiring judges to engage in mandatory authentication evaluations outside of the presence of juries, eliminating the prejudicial aspect of the “Liar’s Dividend.”[xx] In a November 2024 meeting, the committee did consider a proposed rule 901(c) change governing “potentially fabricated or altered electronic evidence.”[xxi] The contemplated rule required a preliminary showing of evidence before claiming evidence is a deepfake and a showing of authenticity once that burden is met.[xxii] The committee ultimately declined to adopt the proposal, observing that 901 was not the appropriate vehicle for reliability standards and the new rule would be “mixing apples and oranges.”[xxiii]
Instead, it developed a new proposed rule, Rule 707 “Machine Generated Evidence,” designed to address concerns about the reliability of AI-generated content.[xxiv] Rule 707 would require AI and other machine-generated evidence be subjected to the same reliability standards as expert witnesses.[xxv] The new rule was available for public comment until February 16, 2026, however there is an important catch.[xxvi] The rule would only apply to evidence the proponent acknowledges was created by AI, doing little to validate evidence whose authenticity is in dispute.[xxvii]
The National Center for State Courts recently warned that AI deepfake evidence is a serious threat to public trust in the court system.[xxviii] One California judge luckily identified that self-represented plaintiffs in a case had submitted AI-generated evidence.[xxix] But what is to be done? It is no longer reasonable to expect a reasonable person to identify the authenticity of video evidence. Rule 707 could be helpful, but it does not solve the problem of deception, which is the prominent issue. One approach would leverage existing civil and evidential rules more proactively to get ahead of disputes in pretrial conferences.[xxx] Both sides should disclose audiovisual evidence they intent to use and flag authenticity challenges before trial, allowing time for proper discovery and expert retention.[xxxi] Courts could then hold a dedicated evidentiary hearing in advance where parties can bring corroborating evidence (such as geolocation phone data proving they were not at the location at the time suggested by the video), retain and depose deepfake detection experts, and file Daubert motions to challenge expert methodologies.[xxxii] Rule 403 would provide a backstop, given that people who see deepfakes could be significantly prejudiced even if the evidence otherwise appears admissible.[xxxiii] Leaving authenticity in the hands of judges is likely the best current option given that such forms of evidence are “technically complex and highly prejudicial to jury deliberations.”[xxxiv]
The Rittenhouse trial may have seemed like a bizarre sidebar in 2021, but it foreshadowed even more dramatic evidentiary scenarios in years since. Courts can no longer take a “wait-and-see” approach to the resolution of deepfake authenticity as synthetic media is now available to everyone using new video generation tools from OpenAI, Google, and other companies. Until the legislative landscape catches up to the technology, courts must take a proactive approach to managing and authenticating potentially synthetic audiovisual evidence.
[i] See State v. Rittenhouse, No. 2020CF983 (Wis. Cir. Ct. Kenosha Cnty. Nov. 19, 2021).
[ii] Shannon Bond, People Are Trying to Claim Real Videos Are Deepfakes. The Courts Are Not Amused, NPR (May 8, 2023, 5:01 AM ET), https://www.npr.org/2023/05/08/1174132413/people-are-trying-to-claim-real-videos-are-deepfakes-the-courts-are-not-amused.
[iii] Matthew Ferraro & Brent Gurney, The Other Side Says Your Evidence Is a Deepfake. Now What?, WilmerHale 2-3 (Dec. 21, 2022), https://www.wilmerhale.com/-/media/files/shared_content/editorial/publications/documents/2022-12-21-the-other-side-says-your-evidence-is-a-deepfake-now-what.pdf.
[iv] Id. at 2.
[v] Id. at 3.
[vi] Id. at 3.
[vii] Daniel J. Capra, Deepfakes Reach the Advisory Committee on Evidence Rules, 92 Fordham L. Rev. 2492 (2024).
[viii] Abhishek Dalal et al., Deepfakes in Court: How Judges Can Proactively Manage Alleged AI-Generated Material in National Security Cases, 2024 U. Chi. Legal F. 200.
[ix] Id.
[x] Ali Rind, Video Evidence Authentication: Standards Courts Expect in 2026, Vidizmo Dems (Feb. 25, 2026), https://digitalevidence.ai/blog/video-evidence-authentication-standards-courts.
[xi] Id.
[xii] Audrey Mitchell, Deepfaked Evidence: What Case Law Tells Us About How the Rules of Authenticity Needs to Change, Berkeley Tech. L.J. Blog (June 23, 2025), https://btlj.org/2025/06/deepfaked-evidence-what-case-law-tells-us-about-how-the-rules-of-authenticity-needs-to-change/.
[xiii] See Id.
[xiv] See Rebecca A. Delfino, Deepfakes on Trial 2.0: A Revised Proposal for a New Federal Rule of Evidence to Mitigate Deepfake Deceptions in Court 7 (Draft, 2025), https://www.uscourts.gov/sites/default/files/2025-04/25-ev-a_suggestion_from_prof._rebecca_delfino_-_rule_901.pdf.
[xv] Mitchell, supra note xii.
[xvi] Id.
[xvii] Rebecca A. Delfino, The Deepfake Defense—Exploring the Limits of the Law and Ethical Norms in Protecting Legal Proceedings from Lying Lawyers, 84 Ohio St. L.J. 1068, 1078 (2024).
[xviii] Id.
[xix] James Bickford, AI Is Coming, But the Rules Aren't Ready, Geo. L. Tech. Rev. Blog (Jan. 2025), https://georgetawnlawtechreview.org/ai-is-coming-but-the-rules-arent-ready/GLTR-01-2025/.
[xx] Id.
[xxi] Frank Young, A Deepfake Evidentiary Rule (Just in Case), UIC John Marshall Law Library (July 3, 2025), https://library.law.uic.edu/news-stories/a-deepfake-evidentiary-rule-just-in-case/.
[xxii]See Id.
[xxiii] Quinn Emanuel Urquhart & Sullivan, LLP, Adapting the Rules of Evidence for the Age of AI, Quinn Emanuel (Nov. 6, 2025), https://www.quinnemanuel.com/the-firm/publications/adapting-the-rules-of-evidence-for-the-age-of-ai/.
[xxiv] See Id.
[xxv] Id.
[xxvi] See Id.
[xxvii] Id.
[xxviii] See Connor Heaton, Shay Cleary & Michael Navin, AI-Generated Evidence Is a Threat to Public Trust in the Courts, Nat'l Ctr. for State Cts. (Feb. 24, 2026), https://www.ncsc.org/resources-courts/ai-generated-evidence-threat-public-trust-courts.
[xxix] Id.
[xxx] Dalal et al., supra note vii.
[xxxi] Id.
[xxxii] See Id. (explaining the Daubert factors could help identify error notoriously high rates with unreliable AI-detection technologies).
[xxxiii] See Id.
[xxxiv] Id.