/>i
A victim reports receiving a series of threatening voicemail messages left by a former partner. The victim sends the recordings to police by email. The call logs from the victim’s phone show the threatening calls originated from a blocked number. The victim's original phone is no longer available. The defendant denies making the calls and claims the recordings are deepfake imitations of his voice.
In the criminal trial for threats, the prosecutor seeks to admit the recordings as evidence. The defendant objects, challenging their authenticity. Before the jury hears the audio evidence, the judge must decide if it is admissible. Generative AI has significantly increased the practical difficulty of carrying out the judge's role as evidentiary gatekeeper.
As the ability to create convincing synthetic audio, video, and images becomes widely accessible, the risk of fabricated evidence distorting the administration of justice grows more pressing by the day. Yet, the rapid evolution and growing accessibility of generative AI do not diminish the continued applicability of existing admissibility doctrines. As a result, the most effective response today is to ensure judges are equipped, through education and institutional support, to apply existing principles to evidence as generative AI becomes commonplace.
The rules of evidence provide a framework for deciding admissibility, but they alone do not equip judges to assess the capabilities and limitations of AI to produce and detect synthetic evidence. With as little as one minute of clear audio as a sample, commercially available software can clone a person's voice and generate audio saying anything the creator types into a prompt.1 Never has the fabrication of evidence been so easy and the outcome so convincing.
Our system of justice depends on judges and juries applying procedures designed to maximize the likelihood of finding the truth. Public confidence in verdicts rests on the historic reliability of those procedures. The existing rules and doctrines of evidence have proven reliable over generations at separating truth from fabrication, and the public trusts outcomes because the process that produces them has earned that trust.
Synthetic evidence threatens that foundation in two distinct ways. First, fabricated evidence may be credited as genuine. Second, genuine evidence may be challenged as fabricated without any basis for the claim. Each failure carries the potential consequence of a verdict resting on something other than the truth. The cumulative risk is that repeated failures of either kind would erode the public's confidence that courts can distinguish real evidence from manufactured evidence, and ultimately their ability to find the truth.
The advent of AI has made the core function of a judge, to apply law to facts, more difficult as the authenticity of those facts is now harder to verify. Questions of evidence admissibility and authenticity are not new; judges make them every day. But the tools available to help them are increasingly insufficient, and given the trajectory of technological advances since deepfakes emerged as a recognized phenomenon in 2017,2 it is unlikely that deepfake identification tools will outperform deepfake creation ones anytime soon.3 That makes it critical that judges be well-versed in current AI capabilities, so they can reliably make evidentiary rulings under the existing frameworks.
In 1988, before DNA evidence was introduced by expert testimony for the first time in a U.S. criminal trial, a Florida court in Andrews v. State held an evidentiary hearing and considered application of a standard set in 1923. There, the court pondered the applicability of the Frye4 standard to determine the admissibility of DNA evidence as a case of first impression. In doing so, Andrews built upon Frye and affirmed "…relevancy as the linchpin of admissibility…" when considering novel evidence.5 The Frye standard was developed sixty years earlier to determine the relevancy of the systolic blood pressure deception test, a precursor to the modern-day polygraph. The introduction of DNA evidence did not require a bespoke rule upon which to determine admissibility. Instead, it relied on the practical evolution of then-existing principles.
The Frye standard eventually gave way fully to Rule 702 of the Federal Rules of Evidence. In Daubert v. Merrell Dow Pharmaceuticals, Inc.,6 the Supreme Court further elevated the "gatekeeper" function of the judge in its interpretation of Rule 702. While those standards apply to scientific evidence specifically, Rule 901 applies to authentication generally. It is under Rule 901 that the judge's gatekeeper function is tested by synthetic evidence. It is through the application of law to facts under Rule 901 that a judge with accumulated experience and understanding of the technology can separate the irrelevant from the relevant evidence before it makes it to a jury.
The law has historically applied, and later adapted, existing standards to new evidence types through the imposition of high bars to admissibility by requiring an increased demonstration of the reliability and authenticity of evidence, and assigning the preliminary fact determination to the judge. This is the historical roadmap that will guide the judicial system as it navigates evidence admissibility in the AI era.
In our hypothetical trial above, the defense challenge to the audio recording offered by the prosecutor must be decided by the judge as gatekeeper of evidence before it can reach the jury. Rule 901(b) provides the non-exhaustive and intentionally broad methods in which evidence may be authenticated. Most relevant to our example here is Rule 901(b)(5), which requires “[a]n opinion identifying a person’s voice – whether heard firsthand or through mechanical or electronic transmission or recording – based on hearing the voice at any time under circumstances that connect it with the alleged speaker.”7 Evidence is admissible if it is relevant, meaning that it tends to prove or disprove a material fact at issue. However, subsumed in the analysis of relevancy is authentication. Evidence cannot have a tendency to make the existence of a disputed fact more or less likely if the evidence is not what its proponent claims it to be. Application of Rule 901(b)(3), permitting comparison by an expert witness, and Rule 901(b)(9), permitting authentication through evidence of a reliable process or system, can further help establish the recording's true origin. Authentication therefore serves as the first, and currently best, safeguard against false evidence reaching the jury.
However, the application of the rules is only as good as the practitioner applying them. If the judge is the gatekeeper, then the judge's understanding of the technology is the gate itself. A gate with a broken hinge does not stop functioning; it simply stops functioning reliably, swinging open when it should hold and holding when it should open. A judge who does not understand the current capability of generative AI to create believable deepfakes cannot fulfill their obligations under the existing rules to determine the authenticity of the evidence. An uninformed judge may apply the correct rules, but in a way that gets the ruling wrong.
The resulting error swings in both directions. On one end, a judge may allow highly persuasive but insufficiently authenticated synthetic evidence to reach the jury before fulfilling the court's gatekeeping obligation under existing authentication rules, thereby making jurors the first meaningful reviewers of authenticity. Once a fabricated recording reaches the jury box, the damage may already be done. Conversely, a judge may improperly exclude authentic evidence based on unsupported allegations that it is AI-generated, giving effect to what has become known as the "liar's dividend," in which the existence of deepfakes provides dishonest actors with plausible deniability for genuine evidence.8 The cost of either failure is substantial.
In order to meaningfully apply existing rules and case law in an age of deepfakes, judges must understand the technology. Judges need not become experts in AI, but they must possess sufficient technological literacy to fulfill their gatekeeping responsibilities. They should understand that generative AI can now produce highly realistic text, images, audio, video, and documents with minimal cost or technical expertise. They should also recognize that human observation alone is no longer a reliable means of distinguishing authentic from synthetic evidence.9 Necessary resources should be deployed to assist jurists in achieving this standard of understanding the current state of technology. Judicial education can take many forms, including continuing judicial education programs and bench books summarizing current AI capabilities and evidentiary implications.10 Courts should also consider building rosters of vetted technical advisors who can assist judges in evaluating novel authentication challenges. This is a mechanism particularly needed for disputed authenticity cases that fall outside the proposed Federal Rule of Evidence 707, which currently applies only where the proponent acknowledges the evidence was AI-generated.11
This foundational knowledge enables judges to assess whether the proponent has presented sufficient evidence to support a finding that an item is what it purports to be and to determine when traditional methods of authentication should be supplemented by evidence of provenance, metadata, chain of custody, corroborating evidence, or expert testimony. Without a basic understanding of AI's capabilities and limitations, courts risk applying established authentication principles based on outdated assumptions about the difficulty of fabricating convincing evidence.
Return to the voicemail recordings that opened this piece. The judge presiding over that case does not need a new rule of evidence to decide whether they are authentic. The judge needs to understand what a smartphone and a minute of recorded speech can now produce. A judge equipped with that understanding can apply Rule 901 as it stands and reach the right result. A judge without it will have more difficulty doing so. Improving outcomes in the courtroom comes not from creating new evidentiary standards, but from improving judicial education and resources to ensure that judges can faithfully apply existing authentication rules to evidence generated in an era of increasingly sophisticated artificial intelligence.
Endnotes
1. ElevenLabs, Instant Voice Cloning Documentation, https://elevenlabs.io/docs/eleven-creative/voices/voice-cloning/instant-voice-cloning (last visited July 8, 2026) (stating that users should "[r]ecord at least 1 minute of audio" for Instant Voice Cloning).
2. Meredith Somers, Deepfakes, Explained, MIT SLOAN SCH. OF MGMT.: IDEAS MADE TO MATTER (July 21, 2020), https://mitsloan.mit.edu/ideas-made-to-matter/deepfakes-explained.
3. Deepfake Detection – Methods, Tools and Research, Alliance for Forensic Integrity and Provenance, https://afip.org/research/deepfake-detection/ (last visited July 21, 2026); Hannah Lee et al., The Tug-of-War Between Deepfake Generation and Detection, ArXiv (2024), https://arxiv.org/abs/2407.06174.
4. Frye v. United States, 293 F. 1013 (D.C. Cir. 1923).
5. Andrews v. State, 533 So. 2d 841, 846 (Fla. Dist. Ct. App. 1988).
6. Daubert v. Merrell Dow Pharmaceuticals, 509 U.S. 579 (1993).
7. FED. R. EVID. 901(b)(5).
8. Robert Chesney & Danielle Keats Citron, Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security, 107 Calif. L. Rev. 1753 (2019). Also, Brennan Center for Justice, "Deepfakes, Elections, and Shrinking the Liar's Dividend" (Josh Goldstein & Andrew Lohn, Jan. 2024): https://www.brennancenter.org/our-work/research-reports/deepfakes-elections-and-shrinking-liars-dividend (last visited July 23, 2026).
9. Nils C. Köbis, Barbora Doležalová & Ivan Soraperra, Fooled Twice: People Cannot Detect Deepfakes But Think They Can, 24 iScience 103364 (2021).
10. National Center for State Courts, AI-Generated Evidence: Evaluating Unacknowledged AI-Generated Evidence, https://www.ncsc.org/resources-courts/evaluating-unacknowledged-ai-generated-evidence (last visited July 23, 2026).
11. Comm. on Rules of Practice & Procedure, Jud. Conf. of the U.S., Preliminary Draft of Proposed Amendments to the Federal Rules of Appellate, Bankruptcy, Civil, and Criminal Procedure, and the Federal Rules of Evidence 108–15 (Aug. 2025), https://www.uscourts.gov/sites/default/files/document/preliminary-draft-of-proposed-amendments-to-federal-rules_august2025.pdf.
