The current rules of evidence were not drafted for a world in which a party's opponent can generate, in under a minute, a photorealistic video of the opposing party confessing to a crime or a plausible-sounding audio recording of a key witness recanting testimony — and the federal judiciary has noticed.
I. The Authentication Framework Under Rule 901
Federal Rule of Evidence 901(a) states the foundational proposition: "To satisfy the requirement of authenticating or identifying an item of evidence, the proponent must produce evidence sufficient to support a finding that the item is what the proponent claims it is." The burden is deliberately low — the proponent need only produce evidence sufficient to support a finding of authenticity; the ultimate question of credibility remains with the trier of fact. Rule 901(b) provides a non-exhaustive list of examples satisfying the requirement, several of which bear directly on AI-generated content:
- Rule 901(b)(1) — Testimony of a witness with knowledge. For unaltered digital media, a custodian or creator can lay the foundation. For AI-generated content, the "creator" may be a synthetic system, complicating this path.
- Rule 901(b)(5) — Opinion about a voice. Traditionally satisfied by a witness familiar with the alleged speaker's voice. A realistic AI voice clone designed to fool even familiar listeners defeats this method at the trier-of-fact stage.
- Rule 901(b)(9) — Evidence about a process or system, showing that it produces an accurate result. This provision is the natural vehicle for AI-system authentication: the proponent would demonstrate that the AI tool generating the evidence has been validated for accuracy. It is equally available to challenge authentication of disputed AI-generated content.
Rule 901 asks only whether authentication is satisfied as a threshold matter — it does not purport to resolve whether the content is actually genuine. That gap is wide enough to admit fabricated evidence that passes a low threshold showing, subject to subsequent credibility attack but potentially after significant prejudice has already occurred.
II. The Advisory Committee's Proposed Rule 901(c)
The Federal Judiciary's Advisory Committee on Evidence Rules has been studying deepfake evidence since at least 2023. The December 2024 Advisory Committee Report outlined — but did not yet vote to adopt — proposed language for a new Rule 901(c) that would create a burden-shifting procedure specifically for AI-fabricated evidence challenges.
As reported by multiple legal publications covering the Committee's June 2025 agenda book, the proposed Rule 901(c) would provide:
(c) Potentially Fabricated Evidence Created by Artificial Intelligence. (1) A party challenging the authenticity of an item of evidence on the ground that it has been fabricated, in whole or in part, by generative artificial intelligence must present evidence sufficient to support a finding of such fabrication to warrant an inquiry by the court. (2) If the opponent meets the requirement of (1), the item of evidence will be admissible only if the proponent demonstrates to the court that it is more likely than not authentic. (3) This rule applies to items offered under either Rule 901 or 902.
This proposed text — which as of the date of this article remains a working draft, not a promulgated amendment — would effect a significant change in authentication mechanics. Under current Rule 901(a), the proponent bears the burden of production to authenticate. The proposed 901(c) would shift the burden to the proponent, at a preponderance standard, once a challenger makes a threshold showing that a reasonable jury could find the evidence was AI-fabricated. The court would then function as an evidentiary gatekeeper — analogous in form, though not in doctrinal basis, to its Daubert role under Rule 702 — screening out deepfakes before they reach the jury.
The Advisory Committee has also developed (and deferred) proposed language for Rule 901(b), which would replace the current requirement that evidence about a process or system be shown to produce an "accurate" result with language demanding a "valid and reliable result" — a standard arguably borrowed from Daubert-land and designed to raise the bar for AI-system authentication. The Committee ultimately adopted a "wait-and-see" posture on this amendment as well, preserving the flexibility of Rule 901's existing framework while acknowledging that the current text was drafted without AI fabrication in mind.
Practitioners litigating in federal court today should track the Advisory Committee's proceedings closely. Rules amendments proposed by the Committee proceed through Supreme Court approval and Congressional review before taking effect; the timeline from proposed text to effective amendment typically runs two to three years.
III. Deepfake Detection: The Scientific Landscape
The legal analysis cannot be divorced from the underlying technology. Modern deepfake detection occupies an adversarial scientific environment: detection methods improve, then generation methods improve to defeat detectors, then detection methods improve again. Several structural points warrant attention for litigators.
Perceptual versus algorithmic detection. Human perception-based detection of audio deepfakes is poor. Studies have documented error rates exceeding 70% for lay listeners evaluating sophisticated voice clones. Trained forensic experts perform better but still achieve error rates that could be exploited in cross-examination. Algorithmic detection tools (examining spectral patterns, generative artifacts, or compression signatures) currently outperform human perception, but their reliability is model-specific and degrades as synthesis technology advances. Expert testimony in deepfake challenges will need to address which detection tool was used, its error rate on the specific type of content at issue, and whether the synthesis method used to create the challenged content was within the detection tool's training distribution.
Provenance-based authentication. The most reliable authentication method is an unbroken chain of custody combined with content provenance metadata. The Coalition for Content Provenance and Authenticity (C2PA) standard provides a framework for embedding cryptographically signed provenance metadata at the point of content creation. For litigation, a recording or video bearing an unaltered C2PA manifest linking it to a specific camera, timestamp, and GPS coordinate is significantly more resistant to deepfake challenge than an unprovenanced file. Practitioners advising clients who routinely generate evidence (surveillance systems, corporate communications) should explore C2PA-compliant recording infrastructure now.
No presumption of authenticity for unprovenanced digital files. Under current Rule 901, the proponent of a document's admission bears the burden of authentication, but digital files lacking metadata or provenance still routinely clear this bar based on surrounding circumstances and witness testimony. The proposed Rule 901(c) framework would require the proponent to affirmatively establish authenticity at a preponderance standard once a challenger makes a facial showing of possible AI fabrication — a significantly harder bar.
IV. State Court Analogs
Most state evidence codes follow the Federal Rules closely, and analogous authentication frameworks apply in Florida and Alabama proceedings.
Florida. Fla. Stat. § 90.901 provides that authentication is "a condition precedent to admissibility" satisfied by "evidence sufficient to support a finding that the matter in question is what its proponent claims it to be." The language tracks FRE 901(a) closely. Florida has not yet enacted deepfake-specific authentication provisions for civil proceedings; practitioners in Florida state court must work within the general § 90.901 framework supplemented by expert testimony and chain-of-custody evidence. Florida has, however, enacted a criminal statute addressing deepfakes in elections, reflecting legislative awareness of the issue even where civil rules have not been updated.
Alabama. Ala. R. Evid. 901 mirrors the federal rule verbatim in its general provision and substantially tracks FRE 901(b)'s illustrative list. As with Florida, Alabama has no deepfake-specific civil authentication rule. The standard method for challenging AI-fabricated evidence in Alabama courts currently relies on Rule 901(b)(9) (process-or-system accuracy showing) combined with expert testimony attacking the reliability of the challenged evidence's source.
V. Practice Notes
Challenging AI-generated evidence. Under current law, challenge AI-generated evidence at three stages: (1) pre-trial, through a motion in limine supported by expert testimony identifying specific AI-generation artifacts in the challenged file; (2) at the time of offer, objecting to foundation under Rule 901 and demanding the proponent make a threshold showing of authenticity; and (3) with the jury, through cross-examination of the proponent's authenticating witnesses and affirmative expert testimony about detection methodology.
Preserving the record. If a court admits challenged AI-generated evidence over objection, preserve the record by specifying on the record: (a) the specific authentication deficiency, (b) the technical basis for the challenge (expert opinion, detection report), and (c) an offer of proof of the evidence the court excluded in ruling on your motion. The proposed Rule 901(c) framework, if adopted, would provide the appellate standard against which the ruling could be measured going forward.
Authentication of your own AI-generated evidence. Clients increasingly use AI tools to generate demonstrative exhibits, timelines, and even expert analyses. The proponent's burden under Rule 901(a) attaches to these as much as to any document. Best practice is to document the AI tool's specifications, version, training data if known, and prompt used to generate the output — records that can support a Rule 901(b)(9) foundation and, if challenged, demonstrate that the tool produces accurate results through validation testing.
Expert witness qualifications. AI forensics is an emerging field with no credentialing body. Daubert challenges to deepfake-detection experts should probe: (a) the error rate and peer-reviewed validation of the detection methodology, (b) whether the expert's method has been applied to the specific type of media at issue (audio vs. video vs. image), and (c) whether the detection tool's training data included examples of the synthesis method allegedly used in the challenged content.
VI. Open Questions
Judicial gatekeeping versus jury evaluation. The Advisory Committee's proposed Rule 901(c) contemplates court-level screening of deepfake challenges, removing the question of AI fabrication from the jury in cases where the proponent fails to establish authenticity at a preponderance standard. Critics argue this exceeds the appropriate judicial role in a system that generally leaves credibility to the trier of fact. Proponents argue that leaving plainly fake evidence before the jury — even with credibility attack — risks unacceptable prejudice. How courts resolve this tension will shape litigation strategy for years.
Self-authenticating AI provenance. Rule 902's self-authentication provisions (certified records, official publications, trade inscriptions) do not currently reach AI-generated content. If C2PA provenance metadata standards become widespread, an argument exists for treating C2PA-certified content as self-authenticating under Rule 902(13)'s provision for certified electronic records. The Advisory Committee has not yet addressed this extension.
Criminal proceedings. This article focuses on civil litigation, but the deepfake authentication problem is arguably more urgent in criminal proceedings where constitutional due process and confrontation clause concerns overlay the evidentiary framework. Brady obligations to disclose exculpatory AI-fabricated evidence, and the prosecution's burden to authenticate evidence it offers, add constitutional dimensions beyond the scope of evidentiary rule amendments.
VII. Closing
Rule 901's flexible framework has absorbed new media technologies across decades. The deepfake challenge is different in degree: the authentication defenses available to judges and juries are perceptually overwhelmed by well-constructed AI fabrications in ways that prior media manipulations could not achieve. The Advisory Committee's proposed Rule 901(c) is a measured, threshold-burden-shifting response that preserves judicial efficiency while providing a mechanism to keep fabricated evidence out of the record before the jury sees it. Practitioners should build deepfake-authentication protocols into their document preservation procedures now, before a high-stakes case depends on defending or challenging a file for which no provenance record exists.
Talk to Yates Anderson
If you are litigating a matter in this area — or weighing whether to — the working analysis above only goes so far. Request a case evaluation and a Yates Anderson attorney will respond within one business day.
Informational only. Not legal advice. No attorney-client relationship is created by reading this post. Consult a licensed attorney in your jurisdiction.