Proposed Rule 707: AI Evidence and the Daubert Standard

No. Proposed Federal Rule of Evidence 707 is not in effect, and it is not on track to take effect soon. The federal Advisory Committee on Evidence Rules published it for comment in August 2025, then withdrew the published version in May 2026 after 59 comments and two hearings. A revised draft, which would apply to evidence produced by "artificial intelligence" rather than "machine-generated evidence," is under study and has not been republished. On the ordinary rulemaking calendar, the earliest a republished Rule 707 could take effect is December 1, 2029.

The idea behind the proposal is simple to state. When a party offers the output of an AI system without an expert witness to sponsor it, and that output would be expert testimony if a person said it, the proponent would have to meet the reliability requirements of Rule 702 (the rule that codifies the Daubert standard). That is the sense in which articles say AI evidence "must meet Daubert." Today, no federal rule says so in terms. Courts reach AI evidence through Rule 702, Rule 901 and state reliability tests such as Frye, matter by matter, as the entries below record.

What the published text said

The version released for public comment on August 15, 2025 was two sentences long. It 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." The text and its Committee Note are in the preliminary draft of proposed amendments the Judicial Conference published that day.

The Committee Note explains the gap the proposal was written to close. When a testifying expert relies on software, the method and the expert's reliance on it are already tested under Rule 702. But when output is presented through "a witness who applied the program but knows little or nothing about its reliability," the Note says, "Rule 702 is not obviously applicable." It continues: "it cannot be that a proponent can evade the reliability requirements of Rule 702 by offering machine output directly."

The Note names the reliability problems it had in mind: "using the process for purposes that were not intended (function creep); analytical error or incompleteness; inaccuracy or bias built into the underlying data or formulas; and lack of interpretability of the machine's process." It tells courts to consider whether training data is "sufficiently representative" for the population in the case, and whether the process "has been validated in circumstances sufficiently similar to the case at hand." It also says that self-authentication under Rule 902(13) is not enough, because that rule "covers authenticity, but does not assure reliability."

The Committee's May 2025 report, which accompanied the draft, gave three examples of the output it meant: stock-trading analysis offered to prove causation, software comparing two works in a copyright case, and machine learning that scores the complexity of source code in a trade-secret case. The same report gave the reason the hearsay rule does not reach such output: "the solution to hearsay is cross-examination, and a machine cannot be cross-examined." The Committee voted 8-1 to publish on May 2, 2025, with the Department of Justice voting against, and the Standing Committee approved publication on June 10, 2025, again over the Department's objection.

How it ties AI output to Rule 702 and Daubert

Rule 702 governs expert testimony in federal court. As amended effective December 1, 2023, it allows an expert to testify only if the proponent demonstrates to the court that it is "more likely than not" that the expert's knowledge will help the trier of fact, the testimony is based on sufficient facts or data, it is the product of reliable principles and methods, and the opinion reflects a reliable application of those methods to the facts. Those four requirements are the 702(a) through (d) the proposal refers to.

The requirements trace to three Supreme Court decisions: Daubert v. Merrell Dow Pharmaceuticals (1993), which made the trial judge the gatekeeper of scientific reliability; General Electric Co. v. Joiner (1997), which allowed a court to exclude an opinion when there is "too great an analytical gap" between the data and the conclusion; and Kumho Tire Co. v. Carmichael (1999), which extended the gatekeeping duty to all expert testimony, not only science. When practitioners say that AI evidence "must meet Daubert," they mean that the proposal would hold unsponsored AI output to this same showing.

The proposal does not add a new test. It borrows the existing one and applies it to output that today can reach a jury without passing through it. The Advisory Committee's reporter, Professor Daniel Capra, wrote in the May 2026 agenda book that practitioners had widely misread the purpose: "This was a common misconception of many practitioners – that Rule 707 was designed to make it easier to admit AI evidence. In fact it is designed to plug a gap in the rules, that currently render AI easier to admit than comparable expert testimony."

What the public comment said

The comment period closed on February 16, 2026. According to the reporter's summary in the May 2026 agenda book, "Of the 59 comments addressing Rule 707," three expressed unqualified support, 27 supported it subject to revisions and 27 opposed it. Two remote hearings were held, on January 15 and January 29, 2026, and the summary records that 13 witnesses testified. Transcripts of both hearings are posted (January 15; January 29).

The criticisms fell into recognizable groups, as the summary lists them: the overbreadth of the term "machine-generated evidence"; the vagueness of the "simple scientific instruments" exception; a concern that the rule creates a pathway for admitting AI output without an expert instead of restricting it; arguments that the rule should not be tied to Rule 702 at all; arguments that machine output should never come in without an expert; cost; and prematurity.

The positions did not divide along the usual lines. The Department of Justice opposed the proposal as premature, said its key terms were vague, and objected that it had no notice provision. Lawyers for Civil Justice opposed it as too permissive and asked for a presumption that only an expert may introduce such evidence. The National Association of Criminal Defense Lawyers warned that it could encourage parties to remove experts from trials. The ACLU supported it but asked that an expert be required in every case. Judge Paul W. Grimm (ret.) and Professor Maura R. Grossman strongly supported it and suggested the term "artificial intelligence," which the revised draft later adopted. The Federal Magistrate Judges Association thought it more prudent to wait and see how AI material is offered in practice.

In June 2026 Professor Capra told the Standing Committee that this was "the first time in his experience that a rule published for public comment did not go forward," and that "most of the practicing lawyers were misunderstanding the rule," according to the draft minutes printed in the October 2026 agenda book.

The revised draft

The Advisory Committee met on May 7, 2026. Its report to the Standing Committee, dated May 17, 2026 and signed by Chair Jesse M. Furman, states: "The Committee does not recommend action on the proposed Rule 707 at this time." The report sets out a revised draft under a new title, "Evidence Produced by Artificial Intelligence and Presented at Trial Without an Expert." Its five parts:

(a) General Rules. "If evidence is a product of artificial intelligence and is offered without an expert witness, but would be subject to Rule 702 if testified to by a witness, the proponent must establish that the evidence: (1) will help the trier of fact; (2) is based on sufficient facts or data; (3) is the product of reliable principles and methods; and (4) reflects a reliable application of the principles and methods to the facts of the case."

(b) Establishing Admissibility. "Admissibility under this rule ordinarily requires the proponent to provide an expert to explain how the system of artificial intelligence reliably produced the evidence; but in exceptional circumstances, the court may rely upon other proof."

(c) Notice. "The evidence is admissible only if the proponent gives an adverse party reasonable notice of the intent to offer the evidence — so that the party has a fair opportunity to meet it."

(d) Judicial Notice. The draft does not apply to facts that may be judicially noticed under Rule 201, which replaces the simple-scientific-instruments exception. (e) Definition. "Artificial intelligence" means "a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations or decisions influencing real or virtual environments," a definition taken from the National Artificial Intelligence Initiative Act of 2020.

The revised Committee Note adds factors a court would weigh: what records the system keeps and deletes, "whether the output can be reproduced or audited," evaluation by entities independent of the developer, and whether the opponent has access to the system. On opaque systems it says: "If the process cannot be explained then the court should in most cases find that the proponent has not established more likely than not that the methodology is reliable." It also suggests a limiting instruction that such evidence "should not be assumed to be reliable simply because it was produced by a machine."

The report states that the revised version "would require re-publication were it to go forward," and that the Committee chose instead to have it vetted by technology experts at its fall meeting. A working draft dated September 15, 2026, printed in the October agenda book, made smaller changes, including notice to "all parties," and the reporter noted that the Committee had not approved all of its language.

The timeline, and the earliest date it could apply

Federal evidence rules move through a fixed sequence, set out on the judiciary’s rulemaking overview: an advisory committee drafts, the Standing Committee approves publication, the public comments for six months, both committees approve a final text, the Judicial Conference approves it (normally in September), the Supreme Court adopts it by May 1, and the rule takes effect on December 1 unless Congress acts. The judiciary says the process usually takes two to three years.

The October 15 meeting in Boston heard two panels. The Rule 707 panel included Peter Henderson of Princeton, Melissa K. Taylor of NIST, Josh Goldfoot of the Department of Justice and Federal Defender Nicole Owens. The deepfake panel included Hany Farid and speakers from OpenAI and Meta. Minutes of that meeting were not posted on uscourts.gov when this guide was checked.

The committee calendar lists the next Advisory Committee meeting for April 14, 2027 and the Standing Committee for June 8-9, 2027. If the Committee approved a revised draft for republication in April 2027 and the Standing Committee agreed in June, the draft could be published in August 2027. On the ordinary sequence that follows, the earliest effective date would be December 1, 2029. That date is this record's calculation from the published calendar, not a forecast the Committee has made, and it assumes nothing goes wrong along the way.

One related change is moving. The August 2026 package proposes an amendment to Rule 104, which governs preliminary questions of admissibility; it is open for comment until February 2027, with a hearing set for January 11, 2027. The reporter told the Standing Committee that the Rule 104 change is needed in part because commenters asked whether the preponderance standard had to be written into Rule 707 itself.

Why so many articles say it is already law

Much of what is published about Rule 707 describes a rule that has been adopted, or gives an effective date of December 1, 2027. Both claims trace to the first months after the June 2025 vote. A June 11, 2025 post by Nelson Mullins carried the headline "Federal Rule of Evidence 707 Approved by Judicial Conference" and called it a "newly adopted rule"; what had been approved was publication for comment. A page on a Vanderbilt law professor's publications site was titled "New Rule 707 – effective December 1, 2027," which was the projected date if nothing changed; it has since added a note that the rule was effectively withdrawn. AI Vortex corrected an earlier version that described the rule as approved with a December 2027 effective date.

Others have reported the change accurately, including Redgrave LLP in July 2026 and the Center for Democracy & Technology in August 2026. The test for any article is the date it was written. Anything published before May 2026 that describes Rule 707 as current or imminent was overtaken by the Committee's vote to withdraw the published text.

What courts do with AI evidence today

Without Rule 707, courts have applied the rules they have. The results are mixed, and each one is recorded on its own entry with its sources.

In State v. Puloka, a Washington trial court excluded cellphone video enhanced with an AI tool, applying the Frye general-acceptance test. In Matter of Weber, a New York surrogate's court rejected the damages analysis of an expert who had used Microsoft Copilot to cross-check his calculations and could not recall his prompts, and stated that AI-generated evidence should be the subject of a Frye hearing before it is admitted. In Ferlito v. Harbor Freight, a federal court allowed an expert who had used ChatGPT to confirm a conclusion he had already reached. In Kohls v. Ellison, an expert declaration containing citations a generative system had invented was excluded.

Authenticity disputes over alleged deepfakes are handled under Rule 901 and its state counterparts. Mendones v. Cushman & Wakefield ended in terminating sanctions over video exhibits the court found to be AI-generated; State v. Amyda and Adams v. Anderson record how appellate courts treated claims that evidence was fabricated. The October 2026 agenda book also cites United States v. Gogic (E.D.N.Y. 2025) and Linkugel v. City of Los Angeles (Cal. App. 2026) on AI-processed and AI-altered material.

Older disputes over algorithmic forensic tools show how the existing reliability tests apply to software. In United States v. Gissantaner, 990 F.3d 457 (6th Cir. 2021), the district court excluded STRmix probabilistic genotyping under Rule 702, and the Sixth Circuit reversed. In People v. Wakefield, 38 N.Y.3d 367 (2022), New York's highest court held TrueAllele generally accepted under Frye. In Ohio, two trial courts excluded reports from the Cybercheck tool as discovery sanctions, and the Ninth District reversed both in September 2024 (State v. Carr, 2024-Ohio-4471; State v. Rankin, 2024-Ohio-4565). None of these depended on a rule written for AI.

State rules on AI evidence

No state has adopted a counterpart to the federal proposal as a rule of evidence, as far as this record has found. Several have acted nearby.

Louisiana. Act 250 of 2025 (HB 178), effective August 1, 2025, amended the Code of Civil Procedure rather than the Code of Evidence. It provides that an attorney "shall exercise reasonable diligence to verify the authenticity of evidence before offering it to the court," and requires a party with a reasonable suspicion that an exhibit was generated or altered by AI to raise it before trial.

New York. Assembly Bill A1338 would bar AI-created evidence in criminal proceedings "unless the evidence is substantially supported by independent and admissible evidence and the proponent of the evidence establishes the reliability and accuracy of the specific use of the artificial intelligence." Senate Bill S9390, introduced March 6, 2026, would add a deepfake burden-shifting provision modeled on the federal draft Rule 901(c). Both were in committee when checked.

California. SB 11, which would have added Evidence Code provisions on screening writings for synthetic content, was vetoed on October 13, 2025. Arizona. A subcommittee of the state's evidence rules committee is studying Rule 707; its May 8, 2026 minutes record "substantial concerns about scope, terminology, and utility." Court rules on AI use in filings, such as Connecticut Practice Book Section 4-9 and New York's Part 161, govern how lawyers use AI, not when AI output is admissible.

What it means for an expert who uses AI

The proposal is aimed at AI output offered without an expert. An expert who relies on such output is already inside Rule 702, and the revised Committee Note's factors read like a cross-examination outline: what data went in, whether the process was validated for this use, whether the output can be reproduced, and whether anyone independent of the developer has tested it.

Those are the questions the Standard answers in advance. Clause 4, checkability before adoption, requires the expert to identify the thing against which an output would be checked before adopting it. Clause 5, reproducible analysis, requires that any count or selection be produced by a method that returns the same result when repeated. Clause 7 keeps a contemporaneous record of each system used and how its output was verified, and Clause 8 commits the expert to disclosing that record when asked, which is the notice the revised draft would require of a party offering AI output directly.

Whether or not Rule 707 is ever adopted, an expert who can answer those questions from a record made at the time is in a stronger position than one who reconstructs the answers in a deposition.

Matters this guide relies on

• Arbuckle v. Tanner, 2026 NSCA 62 August 12, 2026 · AI-generated and altered evidence Primary document read
• Connecticut Practice Book Section 4-9 (Generative AI) June 23, 2026 · Rules, orders and guidance Primary document read
• New York courts: interim AI policy and Part 161 June 1, 2026 · Rules, orders and guidance Primary document read
• Adams v. Anderson (Kentucky Court of Appeals) May 29, 2026 · AI-generated and altered evidence Primary document read
• Proposed Federal Rule of Evidence 707 May 7, 2026 · Rules, orders and guidance Primary document read
• State v. Amyda (deepfake claim without evidence) January 28, 2026 · AI-generated and altered evidence Primary document read
• Mendones v. Cushman & Wakefield September 9, 2025 · AI-generated and altered evidence Primary document read
• Ferlito v. Harbor Freight Tools USA April 23, 2025 · Expert witnesses and AI Primary document read
• Kohls v. Ellison January 10, 2025 · Expert witnesses and AI Primary document read
• Matter of Weber October 10, 2024 · Expert witnesses and AI Primary document read
• State of Washington v. Puloka March 29, 2024 · AI-generated and altered evidence Reported; document not located

Questions