In August 2025, a proposed federal rule was published for public comment. It was designed to stop machine output from entering courtrooms without anyone having to prove the machine was trustworthy. By May 2026, the committee that wrote it had declined to send it forward. The rule’s own fate now mirrors the problem it was written to solve: something presented as authoritative, quietly withdrawn, and still shaping what happens in American courtrooms.
The Gap the Rule Was Built To Fill
Federal Rule of Evidence 707 is not law. It was published for comment in August 2025 to make AI and other machine output offered without an expert meet the Rule 702 reliability test. In May 2026 the Advisory Committee declined to send it forward, rewrote it around “artificial intelligence” and sent it back for study. It has no effective date, and it cannot take effect on 1 December 2026.
The problem the draft addressed is a specific kind of deception: not a machine lying, but a machine’s output arriving in court with the appearance of expertise while nobody has tested whether that expertise exists. When an expert relies on software to reach an opinion, the judge tests that reliance under Rule 702, the expert testimony rule. But AI output can also come in without any expert: directly, through a self-authenticating certificate under Rule 902(13), or through a lay witness who “pressed the button” and knows little about how the system works. The committee’s report gives facial recognition evidence offered this way as an example. [1].
The version published for public comment in August 2025 read, in full: “Rule 707. Machine-Generated Evidence. 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.” In practice, that would mean the proponent has to show the output helps the jury, rests on sufficient facts or data, comes from reliable principles and methods, and reflects a reliable application of those methods to the case. The committee’s draft note listed the reliability problems it had in mind: using a system for purposes it was not built for, analytical errors, bias in training data and the lack of interpretability of how the machine reached its result. These are the same concerns courts already weigh for forensic methods, as the record on fingerprint accuracy shows.
A Comment File Split Down the Middle
The minutes record that the comments were split: 3 offered unqualified support, 27 supported the rule with changes and 27 opposed it. [1] Critics said “machine-generated evidence” was too broad, that the “simple scientific instruments” exception was vague, that the rule might create a route to admit AI output without an expert, and that it was premature Groups including the ACLU and the New York City Bar Association said the problem was already arising, especially in criminal cases.
That split is itself a signal about how hard it is to draw a line between a tool that assists judgment and a tool that substitutes for it. The opposition did not argue that AI output was reliable. It argued that the rule’s wording could not separate the two cases cleanly. The deception at stake is structural: a system that produces fluent, confident answers looks the same whether it was validated for the task or merely pressed into service for it.
The Committee Says No, Then Rewrites
The committee’s May 2026 report states plainly that it “does not recommend action on the proposed Rule 707 at this time.” [1] The October 2026 agenda book describes the May decision as a vote “to withdraw the proposed rule as issued for public comment, and to submit a modified rule proposal for further study.” [1] When the chair reported to the Standing Committee in June, the committee’s reporter, Professor Daniel Capra, said this was the first time in his experience that a rule published for comment had not gone forward. [1] The October 2026 agenda book describes the May decision as a vote “to withdraw the proposed rule as issued for public comment, and to submit a modified rule proposal for further study.” That procedural fact deserves weight. A rule that had been drafted, published, commented on and debated was pulled back — not because the underlying risk had evaporated, but because the committee judged the instrument too blunt. The technology’s appearance of reliability may have outrun the legal system’s ability to certify it.
What the Redraft Changes
The redraft the committee preferred, known in the minutes as Version 2, is titled “Evidence Produced by Artificial Intelligence and Presented at Trial Without an Expert.” It differs from the published text in several respects.
The draft note says courts would usually look at whether training data is representative, whether the system has been validated in similar conditions, whether its output can be reproduced or audited, whether independent evaluators have tested it, and whether the opponent has enough information about how it works. The revised text also presumes that an expert is needed and requires notice. If a process cannot be explained, the note says, the court should in most cases find reliability not established. It also suggests a limiting instruction telling jurors not to assume machine output is reliable simply because a machine produced it.
That last instruction is the most direct acknowledgment of the deception problem. [1]. Jurors, like everyone else, read fluency as competence. A system that answers in complete sentences and confident tones invites trust it may not have earned. Telling jurors to discount that instinct is an attempt to correct, at the point of deliberation, an impression formed long before.
The Department of Justice’s Reservation
The Department of Justice raised continuing concerns about how the rule would work in practice, according to the minutes, and favoured considering it alongside the deepfake draft. [1]. The committee also flagged that the civil and criminal rules might need disclosure changes if Rule 707 goes ahead.
Those concerns point at a different face of the same problem. A reliability rule assumes the evidence is genuine but possibly unreliable. A deepfake rule assumes the evidence may be fabricated entirely. Both involve a gap between what an exhibit appears to show and what it actually is, and the committee has chosen to develop the two tracks in parallel rather than let one outrun the other.
The Calendar Nobody Can Compress
Not on 1 December 2026, and not on 1 December 2027. Under the Rules Enabling Act, 28 U.S.C. §2074, the Supreme Court must send a new rule to Congress by 1 May, and it takes effect no earlier than 1 December of that year. The US Courts guide to the rulemaking process describes the sequence: advisory committee, public comment, Standing Committee, Judicial Conference, Supreme Court.
To take effect in December 2027, Rule 707 would have needed Judicial Conference approval in September 2026. It did not get that. The committee has also said the revised version must be republished. On the usual annual cycle, where proposals are published in August and receive final approval the following year, a revised rule published in August 2027 could not take effect before 1 December 2029. That is an estimate based on the standard timetable, not a date the committee has set. It is also possible the committee drops the idea altogether.
What Judges
Do in the Meantime Until then, federal judges handle AI evidence under the existing rules: Rule 702 when an expert relies on AI, Rule 901 for authentication, Rule 403 for unfair prejudice and their general case management powers. Lawyers who use AI to prepare filings face separate risks, covered in the guide to AI hallucination sanctions.
The practical effect is that the reliability question is answered case by case, by judges applying general standards to systems whose inner workings they may not be able to inspect. The machine’s claim to authority is tested, if at all, by whatever the parties happen to bring forward. That is precisely the condition the draft rule was meant to end.
Deepfakes: A Different Kind of Doubt
Deepfakes raise a different problem: not whether real AI output is reliable, but whether a video, image or recording is genuine at all. Authentication under Rule 901(a) has a low bar: evidence “sufficient to support a finding” that the item is what its proponent claims.
The committee has decided that, for now, an amendment to Rule 901 is not warranted. Its May 2026 report says few courts have yet faced deepfake claims, and a Federal Judicial Center survey found only 15 judges reporting a deepfake issue. [1]. Instead, it has a working draft of a new Rule 901(c) ready to use if needed. Under that draft, a party claiming an item was fabricated by generative AI must first produce evidence sufficient to support a finding of fabrication. If it does, the proponent must show the court the item is “more likely than not authentic.” The draft also requires reasonable pretrial notice of a fabrication claim.
The low threshold matters. [1]. Rule 901(a) was written for a world where authentication meant showing an item came from where it claimed to come from. Generative systems have made the appearance of authenticity cheap to produce and hard to disprove, so a standard designed to filter out obvious forgeries now has to contend with forgeries that leave no obvious trace.
Two Foreign Approaches, Neither a Rule 707
Neither England and Wales nor the European Union has a direct equivalent of Rule 707, but both are moving in related directions.
In England and Wales, criminal courts presume that a computer was operating correctly unless there is evidence to the contrary. [1]. That common law presumption replaced section 69 of the Police and Criminal Evidence Act 1984, which was repealed by the Youth Justice and Criminal Evidence Act 1999. After the Post Office Horizon scandal, the Ministry of Justice opened a call for evidence on software-generated evidence on 21 January 2025, which closed on 15 April 2025. It proposed limiting any reform to evidence generated by software, expressly including AI and algorithms. When checked, the page said a summary of responses would be published in due course.
The Horizon scandal is the clearest recent illustration of the deception at the heart of this debate. A system that appeared to produce accurate records was trusted for years, and the presumption that computers work correctly gave that appearance legal force. The call for evidence is an attempt to ask whether that presumption should survive contact with software whose behavior no one can fully explain.
In the European Union, the AI Act regulates AI systems rather than courtroom admissibility, which stays with national procedural law. [1]. Annex III lists as high-risk AI systems used by law enforcement “to evaluate the reliability of evidence” in criminal investigations and prosecutions, and systems used to assist a judicial authority “in researching and interpreting facts and the law.” The timing of those obligations has shifted, as explained in the guide to the EU AI Act after the Digital Omnibus.
A Checklist for the Courtroom
Even without a new rule, the committee’s drafts are a useful checklist for challenging or defending AI output. Ask for the training data description, validation studies, error rates, audit logs and whether independent testers have had access. Raise reliability early, before trial, and ask for a limiting instruction if the output comes in. If you rely on AI tools for review or analysis, keep records of what the tool did and how it was checked. The duties in ABA Formal Opinion 512 apply to any AI a lawyer uses on a matter. [1].
Each of those requests targets the same gap: the distance between what a system claims to do and what can be shown that it does. A validation study answers whether the system was tested on the kind of case at hand. An audit log answers whether the output can be traced. A limiting instruction answers whether the jury will mistake fluency for accuracy. None of them closes the gap entirely, but each narrows it.
Questions the Record Answers
Is Rule 707 in force? No. Rule 707 was published for comment in 2025, but the Advisory Committee did not recommend final approval in May 2026. It has not been approved by the Standing Committee, the Judicial Conference or the Supreme Court, so it has no effective date.
Could it take effect on 1 December 2026? No. A rule taking effect on 1 December 2026 needed Supreme Court approval by 1 May 2026. Rule 707 never reached that stage, and the committee has said a revised version would need a new public comment period.
What would the rule have required? As published, it requires machine-generated evidence offered without an expert, which would be expert testimony if a person gave it, to meet the Rule 702 reliability standards. The revised draft also presumes an expert is needed and requires notice.
Does it address deepfakes? No. Rule 707 is about the reliability of genuine AI output. Deepfakes raise authenticity questions, which the committee is addressing through a separate working draft of Rule 901(c) that it is holding in reserve for now.
How do courts handle AI evidence now? Under the existing rules. Judges apply Rule 702 when an expert relies on AI, Rule 901 for authentication, Rule 403 to exclude unfairly prejudicial evidence, and their general case management powers before trial.
Does England and Wales have an equivalent? No. The Ministry of Justice consulted in 2025 on reforming that presumption for software-generated evidence, including AI, but no new rule had been made when checked.
The Variable That Decides It
The rulemaking record leaves one open question that will determine whether the gap between machine claims and machine performance is ever closed in federal court: whether the committee’s revised draft survives its next public comment period. [1]. The May 2026 vote was not a rejection of the problem, and the Version 2 text shows the committee still believes a rule is needed. But the same objections that stalled the first draft — breadth, vagueness, timing — have not been resolved, and the republishing requirement means the earliest realistic effective date sits years away.
In the interval, the deception persists in its ordinary form. A system produces an answer. A party offers it. A judge decides, with whatever information the parties supply, whether the answer deserves weight. The proposed rule would have made that decision systematic. [1]. Its withdrawal leaves it, for now, to the judgment of individual courts — which is exactly where the problem was before anyone drafted Rule 707 at all.
Sources
Mentioned organisations (context, not sources)
- ACLU — Organisation (homepage)
- New York City Bar Association — Organisation (homepage)
- Department of Justice — Organisation (homepage)
- Federal Judicial Center — Organisation (homepage)
- Ministry of Justice — Organisation
- European Union — Organisation (homepage)
- American Bar Association — Organisation (homepage)