Authenticating AI-Generated Evidence Under Rule 901 in the Deepfake Era

When the U.S. Judicial Conference's Advisory Committee on Evidence Rules met on May 2, 2025 to weigh whether the Federal Rules of Evidence need new tools for artificial-intelligence fakery, it reached a notably cautious conclusion: a rule amendment was not necessary at that time, given the courts' existing methods for evaluating authenticity and the limited instances of deepfakes seen in courtrooms to date. Yet the same Committee did not simply walk away — it drafted a standby amendment, Rule 901(c), and put it on the shelf for future consideration should circumstances change, while a companion proposal on machine-generated evidence, Rule 707, advanced to public comment on August 16, 2025.

That split decision captures the current posture of the bench and the bar. The existing framework is holding, but the drafters are visibly preparing for the day it does not. This article offers practical guidance on how courts are applying the authentication rules to synthetic audio, video, and documents, in three parts: (1) what Rule 901 actually requires and why deepfakes strain it; (2) what the proposed Rule 901(c) and Rule 707 amendments would change; and (3) a working checklist for authenticating — or challenging — AI-era evidence.

What does Rule 901 actually require?

Authentication is a threshold, not a verdict. Rule 901(a) sets the bar low by design, asking only for enough proof to let a reasonable juror find the item genuine.

Historically, that prima facie showing could be made through the testimony of a witness with knowledge, comparison with an authenticated specimen by an expert or the trier of fact, or the item's own distinctive characteristics. The judge does not decide that the evidence is real; the judge decides only that a reasonable person could so find, and the genuine weight of the item is left to the jury. Rule 902 supplies a parallel track for items that are self-authenticating and require no extrinsic evidence at all. This is a deliberately permissive gate, and for decades it has done its work without much controversy.

Why do deepfakes strain the traditional framework?

Generative AI attacks the framework from two directions at once. The first is the obvious one: fabricated audio, video, or documents that are difficult to discern from reality can clear a low prima facie bar if no one is equipped to challenge them. The second is subtler and, in the long run, more corrosive. Because jurors now know that convincing fakes exist, a litigant can cast doubt on authentic evidence simply by crying deepfake — a phenomenon law professors Bobby Chesney and Danielle Citron labeled the "liar's dividend." The concern is real: the same technology that lets a wrongdoer manufacture a false recording lets a wrongdoer dismiss a true one.

The Advisory Committee was alert to both problems, but it drew a line between them. Its draft authentication amendment targets the fabrication problem directly. For the liar's-dividend problem — the risk that juries will distrust genuine evidence — the Committee pointed not to a new rule but to existing tools: Rule 403's power to exclude evidence or argument whose prejudice substantially outweighs its probative value, and the trial judge's traditional gatekeeping role in policing unsupported attorney assertions that nothing on the screen can be believed. In other words, the framework already contains an answer; the question is whether courts will use it consistently.

What would proposed Rule 901(c) change?

Proposed Rule 901(c), titled "Potentially Fabricated Evidence Created by Artificial Intelligence," would install a two-step, burden-shifting test for deepfake challenges, and it would apply to items offered under either Rule 901 or Rule 902. First, the party attacking the evidence must clear a gate of its own: it must present evidence sufficient to support a finding that the item was fabricated, in whole or in part, by generative AI. A bare assertion that something is a deepfake would not be enough to trigger judicial inquiry. Second, if the challenger meets that showing, the proponent may have the item admitted only by demonstrating that it is, in the Committee's words, "more likely than not authentic."

That second step is the meaningful shift. It moves the proponent from the ordinary prima facie standard to a preponderance standard — the Committee explained that the proponent "must prove authenticity under a higher evidentiary standard than the prima facie standard ordinarily applied under Rule 901." The design is conservative: it raises the bar only after a challenger has put real fabrication evidence on the table, so routine exhibits are not swept into a mini-trial on authenticity. It is a calibrated response, not a wholesale rewrite. Importantly, the Committee held Rule 901(c) back for future consideration rather than sending it out for public comment, a signal that the drafters see the need as contingent rather than urgent.

How does proposed Rule 707 fit alongside Rule 901?

Rule 707 addresses a different slice of the problem: not whether an item is a fake, but whether machine-generated output is reliable enough to be trusted like expert analysis. On August 16, 2025, the Committee on Rules of Practice and Procedure published a revised Rule 707 for public comment, a period that ran through February 16, 2026. The draft would channel certain AI output through the same reliability screen that governs human experts under Rule 702.

The logic is straightforward. A proponent should not be able to evade the reliability requirements of Rule 702 by offering raw machine output where the same conclusion, if voiced by a human expert, would face a Daubert-style challenge. Under the proposal, AI output offered for its truth would need to rest on sufficient facts or data, flow from reliable principles and methods, and reflect a reliable application of those methods to the facts. The practical consequence for litigators is a coming fight over discovery: how a model was built, what prompts and inputs produced the output, and how far a party may probe an opponent's use of AI without colliding with work-product protection. Rule 707 is aimed at the black box, not at the deepfake — it is a reliability rule, not an authentication rule — and counsel should not confuse the two.

How should counsel authenticate — or challenge — AI-era evidence?

Until the rules settle, the winning posture is preparation on both sides of the exhibit. Before you offer audio, video, or a document that a sophisticated opponent might attack as synthetic, and before you challenge one, work through a short set of diagnostic questions:

1. 01Can you establish the chain of custody and the native file, including metadata, hash values, and the capture device, rather than a re-encoded copy?
2. 02Do you have a witness with personal knowledge who can testify to what the recording or document depicts and when it was made?
3. 03If you are the challenger, what affirmative evidence of fabrication can you put forward — provenance gaps, generation artifacts, or forensic analysis — beyond the bare label "deepfake"?
4. 04If you are the proponent, can you meet a preponderance showing of authenticity if the court demands one, or are you relying on the low prima facie bar alone?
5. 05For any AI-derived output, can you explain the model, the inputs, and the methodology well enough to survive a Rule 702-style reliability inquiry?
6. 06Have you engaged a qualified forensic examiner early, while the digital evidence and its metadata are still intact?

These questions map directly onto the emerging framework: chain of custody and knowledgeable-witness testimony satisfy the traditional Rule 901 gate; affirmative fabrication proof and a preponderance rebuttal anticipate Rule 901(c); and methodological transparency anticipates Rule 707. A litigant who can answer all six is well positioned regardless of which proposals are ultimately adopted. As with any technical dispute, early intervention is key — it is faster and cheaper to secure forensic help before an authenticity fight erupts than to reconstruct provenance after the fact. Parties evaluating whether a neutral forensic examiner can help frame these questions may find our guidance for counsel and resources a useful starting point.

Conclusion

The trend is clear even if the rulemaking is not: synthetic media is arriving in litigation faster than the Federal Rules can be amended, and the drafters have chosen to prepare rather than to react. For now, the existing authentication framework — Rules 901 and 902, backstopped by Rule 403 and judicial gatekeeping — remains the operative law, and courts have found it adequate to the limited caseload of deepfake disputes so far. That may not hold, which is why Rule 901(c) sits ready on the shelf and Rule 707 has moved through public comment. The value of these tools, in the right circumstances, will depend on whether the bench and the bar develop the requisite technical and legal expertise to use them, and counsel who build authentication discipline into their evidence handling now will be ready whichever way the rules break. We should continue to track how authentication issues are managed by courts as the amendments move forward.

The content is intended for general informational purposes only and should not be construed as legal advice.

Editorial note — This briefing was drafted by an AI system from editor-selected sources and published under the editorial standards set out in our newsroom. It carries no individual byline because no individual wrote it. It is general information, not legal advice.