AI-generated evidence and deepfakes: what Rule 707 would do

The short version

• Authentication means showing a court that a piece of evidence is really what you say it is. A deepfake is audio, video or an image that generative AI has made or altered to look real.
• There is no United States federal rule written for AI-generated or deepfake evidence. Courts use the authentication rule, Rule 901, and the expert-reliability rule, Rule 702.
• Proposed Rule 707 would make AI output offered without an expert meet the Rule 702 reliability test, closing the gap where an AI conclusion reaches a jury through a lay witness.
• Draft Rule 901(c) would set a two-step test for a deepfake challenge: the challenger shows enough to suggest fabrication, then the proponent proves the item is more likely than not authentic.
• Both are proposals, not law, as of 10 November 2026. Rule 707 was withdrawn after public comment and is back for study; Rule 901(c) is prepared but not approved and not published.
• Deepfake challenges are still rare in federal court. Of 931 judges who responded to a survey, only 15 had ever seen one.
• This guide is about United States evidence law. Deepfake evidence inside arbitration is in AI document review in arbitration; the overview is AI in dispute resolution.^1,2,3,5,6

What is the problem with AI-generated and deepfake evidence?

There are two problems, and the Advisory Committee on Evidence Rules, the body that drafts changes to the Federal Rules of Evidence, keeps them apart. The first is reliability: an AI tool can produce a result no one can explain or reproduce. The second is authenticity: a deepfake can make a fake photo, recording or video look genuine.³

Two plain definitions first. Authentication means producing enough evidence that an item is what its proponent, the party offering it, says it is. A deepfake is synthetic media, audio, video, text or images that generative AI, software that creates new content from patterns in existing data, has produced or altered to deceive.^1,3

Deepfakes carry a second problem the committee calls the “liar’s dividend”: once people know convincing fakes exist, a party may wave away real evidence as a deepfake. The committee’s materials note that a broad claim that nothing can be trusted can itself be kept out under Rule 403, which lets a judge exclude evidence that would mainly confuse or mislead the jury.³

The related failure mode, AI inventing cases and citations, is a hallucination, a confident but made-up output; the dated record of court sanctions for it is in the AI hallucination court sanctions record. AI document production and deepfakes inside arbitration, where the evidentiary rules are looser, are in AI document review in arbitration.³

How do you authenticate AI evidence under the rules today?

Under the current Federal Rules of Evidence there is no AI-specific rule, so a party authenticates AI material the ordinary way. Rule 901(a) sets a low bar: the proponent “must produce evidence sufficient to support a finding that the item is what the proponent claims it is.” That is a prima-facie showing, enough to pass at first sight, and the jury decides ultimate authenticity.¹

Rule 901(b)(9) is the usual route for machine output. It treats as sufficient evidence “describing a process or system and showing that it produces an accurate result.” Rule 702 handles expert reliability separately: the proponent must show, “more likely than not,” that an opinion rests on sufficient facts, reliable methods and “a reliable application of the principles and methods to the facts of the case.” That reliability review is often called Daubert, after the Supreme Court case behind Rule 702.¹

AI output is usually not hearsay. The hearsay rules cover out-of-court statements by people, and the committee’s materials explain that a machine’s output is not hearsay because a machine is not a person. That is why the committee’s focus is reliability and authentication, and why Rule 707 is aimed at reliability rather than the hearsay rules.³

As of the enacted text, no federal rule names AI, deepfakes or machine output offered without an expert. The two drafts below would fill that gap.¹

What would proposed Rule 707 do, and is it law yet?

No. Rule 707 is a proposal, not law, and as of 10 November 2026 it is not even out for public comment. The current comment cycle, open from 14 August 2026 to 15 February 2027, lists several Civil, Criminal and Evidence Rules, but the only Evidence Rules among them are 104 and 902, and the pending-amendments list carries no Rule 707. It is not pending before the Judicial Conference, the Supreme Court or Congress.^3,5,6

As published in August 2025, Rule 707, titled “Machine-Generated Evidence,” read: “When machine-generated evidence is offered without an expert witness and would be subject to Rule 702 if testified to by a witness, the court may admit the evidence only if it satisfies the requirements of Rule 702(a)-(d). This rule does not apply to the output of simple scientific instruments.” It applies the Rule 702 reliability review, the Daubert reliability test, to AI output a party offers through a lay witness instead of an expert; the “simple scientific instruments” carve-out was meant to keep calculators out.²

The Standing Committee, the body that must clear a proposal before it goes out for public comment, approved Rule 707 for public comment in 2025, and the comment period closed on 16 February 2026. There were two public hearings and 59 written comments by the deadline, of which 3 gave unqualified support, 27 supported it with revisions and 27 opposed it. At its 7 May 2026 meeting the Advisory Committee voted to withdraw the published version and send a modified proposal for further study, which it took up at its 15 October 2026 meeting. The September 2026 report of the Committee on Rules of Practice and Procedure lists Rule 707 as still under consideration.^3,4

The revised working draft, called Version Two, would ordinarily require an expert. Its admissibility clause reads: “Ordinarily, the proponent must provide an expert to explain how the system of artificial intelligence reliably produced the evidence. But in exceptional circumstances, the court may consider other proof to find that the output satisfies the requirements of (a)(2)-(4).” It also drops the “simple scientific instruments” exception, covers evidence that is “a product of artificial intelligence,” adds a notice requirement and defines artificial intelligence. Our AI in law guide places Rule 707 in the wider courts picture.³

What would draft Rule 901(c) do about deepfakes?

Rule 901(c) is at an earlier stage: prepared, but not finally approved, and not published for comment. It would set a two-step process when a party claims an item was fabricated, in whole or in part, by generative AI.³

Step one puts a burden on the challenger: “present evidence sufficient to support a finding of fabrication to warrant an inquiry by the court.” A bare cry of “deepfake” is not enough. Step two shifts to the proponent: if that showing is made, the item is admissible only if the proponent shows it is “more likely than not authentic.” That is a Rule 104(a) preponderance standard, higher than the ordinary prima-facie authentication bar. A member of the committee called the draft’s burden-shifting its principal innovation. The rule would cover items offered under Rule 901 or Rule 902 and require pre-trial notice.³

Not everyone agrees it is ready. The United States Department of Justice “does not believe Rule 901(c) is ready for publication,” its representative told the committee, while still giving input on the draft, including on how pre-trial notice would work in criminal cases.³

How often do deepfake challenges actually arise in court?

Rarely, so far. The Federal Judicial Center, the courts’ research agency, surveyed federal trial judges. Of the 931 who responded, only 15 reported ever meeting a deepfake challenge. Among those who had not, 82% said they would want some initial showing before taking a fabrication claim seriously, which is the mindset step one of Rule 901(c) would put into a rule. A narrow majority of those without experience also thought the rules should be updated for deepfakes.³

How have courts handled AI-generated and challenged evidence so far?

A handful of decisions show courts applying the existing rules, not any new one.^3,7,8

In Matter of Weber a New York Surrogate’s Court, which handles wills, estates and trusts, tested Microsoft Copilot on three court computers and got three different figures for the same question. It held that a court “cannot blindly accept as accurate, calculations which are performed by artificial intelligence,” and that counsel has an “affirmative duty to disclose the use of artificial intelligence,” with a Frye hearing before AI-generated evidence comes in. Frye is New York’s test for whether a scientific method is generally accepted in its field. This is one trial court under New York law, not a deepfake case.⁸

In United States v. Belt the Eighth Circuit applied Rule 901(a)’s low bar to home-surveillance video, holding that a party “need only prove a rational basis” for its claim about the video, that the government “is not required to exclude all possibilities of tampering,” and that the unexplained gaps the defence relied on go to the weight of the evidence, not its admissibility. The panel opinion does not mention deepfakes. A separate dissent from the denial of rehearing en banc, that is, the full court declining to reconsider the panel’s decision, as quoted in the agenda book, used the case to argue that as AI video improves, “courts should be more stringent in considering foundation.”^3,7

The agenda book also describes two trial-court matters whose orders are not on an official source, so we state only what the book says. It cites Mendones v. Cushman & Wakefield as “one of the first times a trial court was confronted with deepfake video evidence.” And it summarises Washington v. Puloka, where a court applied Rule 702 to exclude an AI-enhanced video as “largely untested and potentially unreliable.”³

Deepfake handling inside arbitration, where the rules of evidence are looser, is in AI document review in arbitration. Arbitration rulings that turn on AI are collected in AI arbitration case law.

Is this only a United States problem?

No. The rules here are United States materials, but other courts face the same questions. Judicial guidance in England and Wales already warns judges about deepfakes, AI-faked images, audio or video (p. 7). We route the detail out rather than repeat it: the EU’s horizontal regime is in the EU AI Act and dispute resolution, United States state statutes are in US state AI laws: Colorado and Texas, and institutional AI guidance is in the SVAMC AI Guidelines and AI in ADR.⁹

What is still open?

• Whether the 15 October 2026 meeting changed either draft. The agenda book is a pre-meeting book, so it does not record what was decided there.
• Whether Rule 707 is republished for public comment, and whether Rule 901(c) is published at all.
• The earliest either rule could take effect. Under the Rules Enabling Act, the law that governs how the federal rules of evidence are made, a proposal must clear public comment, the Advisory Committee, the Standing Committee, the Judicial Conference, the Supreme Court and a review period in Congress before it takes effect. As of 10 November 2026 neither is on that path, and we found no official effective date, so treat any projected date as unconfirmed.^3,5,6

What does this mean for anyone using AI evidence?

In our view, the drafts tell one story even before any of them is law. A person stays responsible: AI output does not authenticate itself, and someone has to stand behind it. Rule 707’s expert, Rule 901(c)’s inquiry and a court’s own reliability test under Rule 702 all ask the same question of whoever offers the evidence. Anything an AI tool contributes to the evidence or the outcome has to be capable of being checked, and an AI result has to be verified before anyone leans on it. A figure no one can explain or reproduce, as in Matter of Weber, is a figure a court can refuse. The rules are running behind the technology. So the safest reading, as of 10 November 2026, is that the duties already in force, authentication and expert reliability, already govern AI evidence. The proposals would mostly make those duties explicit, though draft Rule 901(c) would also raise the bar a challenged item must clear, and Rule 707 would close the gap where AI output reaches a jury through a lay witness.^1,2,3,8