Federal Rule of Evidence 707: The AI Evidence Rule That Isn't Law Yet

Proposed Federal Rule of Evidence 707 would require AI-generated evidence offered without an expert witness to satisfy Rule 702's reliability requirements. It is not law. As of September 2026 the Advisory Committee on Evidence Rules has revised the proposal, declined to advance it, and sent it for further study, and no version of Rule 707 can take effect before December 1, 2029.

That last point matters most. Nearly all coverage of Rule 707 was written during the 2025 comment period, and some of it reads as though the rule exists.

This article is for general information and is not legal advice. Rulemaking status changes; confirm the current posture on uscourts.gov before relying on anything here.

Where proposed Rule 707 actually stands

The Rules Enabling Act process runs through five stages: Advisory Committee, Standing Committee, Judicial Conference, Supreme Court, Congress. Rule 707 has not cleared the first.

The Advisory Committee's May 17, 2026 report is unambiguous: "The Committee does not recommend action on the proposed Rule 707 at this time." It has instead revised the rule and will study it further, alongside deepfakes, at its next meeting. Of the 59 comments, three gave unqualified support, 27 supported it subject to revisions, and 27 opposed it.

So Rule 707 has never reached the Judicial Conference, the Supreme Court, or Congress, and appears nowhere on the judiciary's pending amendments page. Two traps: Evidence Rules 104 and 902 are out for comment through February 15, 2027, and neither is an AI proposal; and several June 2025 client alerts say Rule 707 was "approved," when what was approved was publication for comment.

The earliest possible effective date

The Committee has said its revised draft "would require re-publication were it to go forward." The August 2026 cycle passed without it, making August 2027 the earliest realistic republication and September 2028 the soonest Judicial Conference approval. Under the rulemaking timetable, the Supreme Court promulgates approved rules before May 1, effective no earlier than December 1 that year. The earliest an AI evidence rule binds a federal courtroom is December 1, 2029, if one survives.

The gap the rule was written to close

Expert opinion runs a gauntlet. Under Rule 702, amended effective December 1, 2023, the proponent must show it is more likely than not that the opinion will help the trier of fact, rests on sufficient facts or data, and reliably applies reliable methods — all before the jury hears a word.

A machine output offered without a sponsoring expert can walk past it. As the Committee Note puts it, "the technician who enters a question and prints out the answer might have no expertise on the validity of the output." The exhibit arrives with nobody to cross-examine about training data, error rates, or validation.

What the proposed rule would do

The text published for comment, reproduced in the Advisory Committee's December 1, 2025 report, was one sentence plus a carve-out:

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.

That last sentence was the thermometer clause, "intended to give trial courts sufficient latitude to avoid unnecessary litigation over the output from simple scientific instruments that are relied upon in everyday life." Nobody wants a Daubert hearing on a bathroom scale.

Commenters attacked "machine-generated" as sweeping in faxes, emails, and cellphone extractions, and called that exception vague. The May 2026 redraft narrows the trigger to "artificial intelligence," deletes the carve-out, and adds pretrial notice. Its subsection (b) says admissibility "ordinarily requires the proponent to provide an expert," with other proof allowed only "in exceptional circumstances"; subsection (d) exempts facts subject to judicial notice under Rule 201, doing the thermometer work without the vague language.

Rule 901, and the separate deepfake draft

Keep these apart. Rule 707 goes to reliability; Rule 901 goes to authenticity, at a much lower bar. Per the Committee Note, clearing Rule 707 would automatically satisfy Rule 901(b)(9), but clearing 901(b)(9) "does not suffice for admissibility." Rule 902(13) self-authentication likewise reaches authenticity, not reliability.

Deepfakes are a separate track. The Committee has a working draft of a new Rule 901(c) requiring an opponent alleging fabrication to make a threshold showing before the court inquires, then shifting to the proponent to prove authenticity by a preponderance. It has never been published for comment. In May 2026 the Committee found such an amendment "is not warranted" for now, citing a Federal Judicial Center survey in which only fifteen judges reported dealing with deepfakes.

What this looks like in a small-firm case

Rule 707 gets discussed as a big-firm problem. The exhibits that would trigger it show up in ordinary cases:

• Accident reconstruction — a simulation sponsored by an adjuster who cannot explain the physics engine.
• Damages models — a life-care plan or lost-earnings projection from an algorithmic platform, offered through an economist who did not build it.
• Medical-record summarization — an AI chronology offered as a demonstrative that quietly makes causation claims.
• Algorithmic scoring — pretrial risk assessments, tenant-screening scores, insurance fraud scores, credit models.
• Video enhancement — footage upscaled by a model that may have generated detail rather than recovered it.
• Voice and image analysis — speaker identification or facial recognition offered by an officer who ran a query.

In each, the tool does what an expert would do, with nobody on the stand to say why it should be trusted.

What to do now, while there is no rule

You argue Rules 702, 901, 403, and 104. That works only if you build the record early.

1. Discover the tool, not the output. Ask for the vendor, product, and exact version; the training data and the population it represents; validation studies and who ran them; published error rates; and whether anyone independent of the developer evaluated it. The Committee Note lists nearly this checklist as the anticipated Rule 707 analysis, which makes it a defensible framework today.

2. Use Rule 26 and Rule 16. Raise AI-derived evidence at the Rule 26(f) conference and get it into the scheduling order. If the output comes through a retained expert, Rule 26(a)(2)(B) already reaches the tool through the facts or data considered. If it comes in without one, ask for a disclosure deadline on the software's provenance.

3. Preserve the objection. "You cannot cross-examine a machine" is easy to waive by silence. Depose the sponsoring witness on what they actually know about the system and preserve your objections cleanly on the record. Then move in limine — the motion in limine is where this fight belongs, before the jury sees a simulation it cannot unsee. If the output comes in, ask for the limiting instruction the Committee Note suggests: that machine-generated evidence "is subject to error and ... should not be assumed to be reliable simply because it was produced by a machine."

4. Read the local standing orders. Many judges have entered their own AI orders, and those bind you today as Rule 707 does not.

AI evidence and AI citations are different problems

Rule 707 concerns AI output offered as evidence by a party — an exhibit, a score, a reconstruction. The control is judicial gatekeeping, and it is years away. Hallucinated citations concern AI output offered as argument by a lawyer, which is not an evidence problem at all. It is policed by Rule 11 and standing orders, and courts are sanctioning it now.

So you can be sanctioned this month for citing a case that does not exist, while the rule governing the AI exhibit across the table remains a draft. Our guide to AI legal citation hallucinations covers how fabricated cases get caught, and our look at how accurate AI legal research actually is explains why chatbots produce them. Tools built for legal research retrieve citations rather than generating them and flag what cannot be verified, which is why verifying every citation before filing is a workflow step, not a personality trait.

The case against Rule 707

The opposition is serious, and at the moment it is winning. Critics argue the existing rules already do this work: Rule 702 reaches AI relied on by a testifying expert, Rule 901(b)(9) requires a showing that a process produces an accurate result, and Rule 403 excludes unfairly prejudicial output.

The overbreadth objection did the most damage. "Machine-generated evidence" plausibly covered geolocation data, surveillance footage, electronic health records, and routine business software — categories courts admit thousands of times a day. A rule putting a Daubert hearing in front of each is not gatekeeping; it is gridlock. A further objection cuts the other way: by specifying when AI evidence may come in without an expert, Rule 707 could be read as a pathway to admission rather than a barrier. The Reporter concluded the criticisms do not warrant abandoning the rule. The Committee declined to advance it anyway.

Frequently asked questions

Is Federal Rule of Evidence 707 in effect?

No. As of September 2026 it is a proposal, not law. Published for comment in August 2025, the comment period closed February 16, 2026, and in May 2026 the Advisory Committee revised it and declined to advance it. It has never reached the Judicial Conference, the Supreme Court, or Congress.

What would proposed Rule 707 require?

The version published for comment provided that 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 it only if it satisfies Rule 702(a)-(d): helpfulness, sufficient facts or data, reliable methods, and reliable application.

When could Rule 707 take effect?

Not before December 1, 2029, and only if it survives. The Advisory Committee has said the revised rule would require republication, and the August 2026 cycle did not include it. The earliest realistic republication is August 2027, producing a December 1, 2029 effective date under the standard Rules Enabling Act timetable.

Does Rule 707 govern AI-hallucinated citations in a brief?

No. Rule 707 would govern AI output offered as evidence by a party. Fabricated citations in a brief are an advocacy problem policed by Rule 11, inherent authority, and local standing orders, not by the Rules of Evidence. You can be sanctioned for a hallucinated case today while Rule 707 remains years away.

Before you file

Rule 707 will not help with the AI exhibit in your case this year. Rule 702, Rule 901, and a well-built discovery record will. Start asking about training data and validation now, and the motion in limine writes itself later.

The other half is the part you control. Before a draft goes out, run it through the Hallucination Shield — free, no signup, and each citation it finds (up to 25 per run) is checked for existence and support.