Proposed Federal Rule of Evidence 707 Would Set New Standard for Machine-Generated Evidence

For more than a century, the federal rules of evidence have asked a simple question when a party offers a conclusion drawn from data: who is vouching for it, and can that person’s methodology withstand scrutiny? Artificial intelligence has quietly complicated that question. When a machine learning system flags a fraudulent transaction, matches a face to a database, or interprets a medical scan, no witness necessarily takes the stand to explain how the conclusion was reached. The Advisory Committee on Evidence Rules, the body within the federal judiciary responsible for studying and proposing changes to the Federal Rules of Evidence, has spent the past two years wrestling with that gap. The result is proposed Federal Rule of Evidence 707, titled Machine-Generated Evidence, which would require that machine output offered without a sponsoring expert witness satisfy the same reliability standard that already governs expert testimony.¹ The rule has proven unusually contentious for a proposal this narrow on its face, and it remains a work in progress rather than settled law. But its trajectory over the past year offers litigators a useful preview of how federal courts are likely to handle AI-derived proof, whatever final form the rule takes.

The Gap Rule 707 Is Meant to Close

Federal Rule of Evidence 702 governs the admissibility of expert testimony, requiring that it assist the trier of fact, rest on sufficient facts or data, reflect reliable principles and methods, and represent a reliable application of those principles and methods to the facts of the case.² Those requirements trace back to the Supreme Court’s decision in Daubert v. Merrell Dow Pharmaceuticals, Inc.,³ which charged trial judges with gatekeeping scientific and technical testimony before it reaches a jury. The Advisory Committee’s concern is that a party can sidestep all of that scrutiny simply by not calling an expert. If a diagnostic algorithm concludes that a physician missed a diagnosis, and the proponent offers that conclusion directly, without a human expert testifying about how the software reached it, Rule 702 by its terms does not apply, even though the algorithm’s output is functionally indistinguishable from expert opinion testimony.⁴ The Advisory Committee has cited that scenario, an AI diagnostic tool’s conclusion in a medical malpractice case, as the paradigm example of the loophole Rule 707 is meant to close.

The problem is not hypothetical. Trial courts have already struggled, on an ad hoc basis, to figure out which evidentiary framework applies to algorithmically processed evidence. In the Kyle Rittenhouse trial, a Wisconsin judge balked at admitting a video enlarged using an iPad’s pinch-to-zoom function, uncertain whether the underlying algorithm generated new pixels rather than merely magnifying existing ones, and the judge candidly admitted he understood little about how the software worked. A Washington state court later excluded AI-enhanced cellphone video of a shooting after applying a patchwork of the Frye general-acceptance test alongside Rules 702, 401, and 403, with the enhancement tool’s opaque methods undermining the showing required under each.⁵ Absent a uniform standard, courts have been left to improvise, an approach that risks inconsistent outcomes as AI tools proliferate across forensic, medical, financial, and investigative contexts.

What the Proposed Rule Would Require

As published for public comment, proposed Rule 707 states: “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.”⁶ In practice, that means a party offering machine output, whether an AI system’s prediction, a proprietary algorithm’s risk score, or software-generated forensic analysis, without an accompanying expert witness must still show that the evidence assists the jury, rests on sufficient facts or data, is the product of reliable principles and methods, and reflects a reliable application of those principles and methods to the case at hand. The rule expressly carves out simple scientific instruments, so a digital thermometer or scale reading would not trigger the heightened standard.

Two limitations on the rule’s scope matter a great deal to litigators evaluating its practical effect. First, Rule 707 applies only where the proponent acknowledges that the evidence is machine-generated; it says nothing about evidence whose AI origins are disputed or concealed. Second, the rule addresses reliability, not authenticity. A party who denies that a video or document was manipulated by AI at all is not affected by Rule 707, because the rule presupposes that everyone agrees a machine produced the evidence and only disputes whether that machine’s output can be trusted. Deepfakes and other authenticity disputes remain governed, for now, by the existing framework under Rule 901.

Why the Committee Bypassed Rule 901

The Advisory Committee did not arrive at Rule 707 as its first choice. At its November 2024 meeting, the Committee considered a detailed proposal from former U.S. District Judge Paul Grimm and Dr. Maura Grossman to amend Rule 901, the authentication rule, directly.⁷ That proposal would have added a burden-shifting mechanism to Rule 901(c): once a party challenging the authenticity of digital evidence showed that a reasonable jury could find it was altered or fabricated by artificial intelligence, the burden would shift to the proponent to prove the evidence more likely than not authentic. A companion amendment to Rule 901(b) would have replaced the existing requirement that process-or-system evidence be shown “accurate” with a requirement that it produce a “valid and reliable result,” with additional disclosure obligations, training data and methodology, when the proponent acknowledged AI involvement.

The Committee ultimately declined to amend Rule 901. Several members concluded that importing reliability criteria into an authentication rule was, in the words used during the Committee’s deliberations, “mixing apples and oranges,” since authentication asks only whether an item is what it purports to be, not whether its underlying methodology is sound. Others pointed to the still-limited number of reported cases involving contested AI evidence and favored a wait-and-see posture, preserving Rule 901’s existing flexibility rather than rewriting it preemptively.⁸ Instead of amending authentication doctrine, the Committee redirected its efforts toward a freestanding reliability rule, which became Rule 707. That choice is why Rule 707, whatever its final form, will not resolve the deepfake and disputed-authenticity problem that motivated the original Rule 901 proposal; that issue remains under separate study.

The Rulemaking Timeline and Current Status

Amending a Federal Rule of Evidence is a multi-year process under the Rules Enabling Act. An advisory committee must first approve a proposal for publication, then the Standing Committee on Rules of Practice and Procedure must approve publication for public comment, then the advisory committee reviews the comments and decides whether to revise or advance the proposal, then the Standing Committee, the Judicial Conference, and ultimately the Supreme Court must each approve the rule before it goes to Congress, which has the opportunity to reject, modify, or defer it before it takes effect.⁹ Rule 707 has moved through only the early stages of that process, and its most recent turn was backward rather than forward.

The Advisory Committee voted 8-1 on May 2, 2025, to seek publication of Rule 707, with the Department of Justice’s representative dissenting on the ground that Rule 702 already reaches AI-generated evidence and that the new rule addressed a speculative future problem rather than a present one.¹⁰ The Standing Committee approved publication for public comment on June 10, 2025, and the proposal was open for comment from August 15, 2025 through February 16, 2026, with public hearings held in January 2026.¹¹ The Committee received 59 written comments by the formal February 16, 2026 deadline—3 in unqualified support, 27 supporting the rule subject to revisions, and 27 opposed—though earlier Committee materials had cited a preliminary count of more than 70.¹² The Department of Justice told the Committee at its November 2025 meeting that the agency had “greater overall concerns” with the rule as drafted; after reviewing the full comment record at its May 2026 meeting, the Advisory Committee itself chose not to advance the rule toward final approval. Instead, the Committee opted to narrow the rule’s scope and solicit further technical input at a mini-conference scheduled for its fall 2026 meeting on October 15.¹³ The Standing Committee, meeting June 3-4, 2026, agreed not to recommend action on Rule 707 at that time, sending it back for revision alongside the still-unresolved Rule 901 deepfake question.¹⁴ As of this writing, Rule 707 has not been adopted, no effective date has been set, and the rule remains under active revision.

A Contested Comment Record

The public comment period exposed sharply divided views on the rule’s basic design. The American Association for Justice, the national plaintiffs’ bar association, filed a formal comment on the final day of the comment period urging the Committee to pause and redraft, arguing that “machine-generated evidence” as originally defined was so broad it would sweep in routinely admitted digital evidence such as geolocation data, surveillance footage, and electronic health records, forcing litigants to satisfy Daubert-style reliability showings for information courts have long admitted without expert sponsorship.¹⁵ The Association recommended narrowing the definition to focus specifically on machine learning and AI systems and exempting commonly used, judicially noticeable technologies.

Other commentators argued the rule did not go far enough. Because Rule 707 applies only to evidence the proponent concedes is machine-generated, critics noted it does nothing for the harder case: a party who disputes that evidence is AI-generated at all, which is precisely the deepfake scenario the Committee’s earlier Rule 901 proposal had targeted.¹⁶ The New York City Bar Association took a more constructive view, supporting adoption of a machine-generated evidence rule but recommending that it not be adopted unless paired with corresponding amendments to Federal Rule of Civil Procedure 26 and Federal Rule of Criminal Procedure 16 to build in explicit pretrial notice and disclosure obligations.¹⁷ Still other commenters, including the Center for Democracy and Technology, flagged that the original carve-out for “simple scientific instruments” was too vague to administer consistently and that borrowing Rule 702’s criteria, developed to evaluate human expert qualifications and methodology, was an imperfect fit for evaluating an AI system, whose reliability depends on training data, testing protocols, and independent validation rather than credentials or professional experience.¹⁸

A recurring theme across the comment record was cost and access. Satisfying Rule 707’s reliability showing will often require technical experts capable of analyzing proprietary AI systems, an undertaking that is expensive and that disproportionately burdens under-resourced litigants and criminal defendants facing government-deployed AI tools.¹⁹ Commentators have predicted that, if adopted, Rule 707 will generate substantial satellite litigation: discovery fights over access to opposing parties’ AI systems, trade secret and confidentiality disputes, in limine motions, and pretrial proffer hearings that one commentator suggested might come to be known as “Rule 707 hearings,” much as claim-construction proceedings became known as Markman hearings.²⁰

The Committee’s Narrower Revised Draft

The revisions the Advisory Committee is now developing respond directly to several of these criticisms. The revised draft applies only to evidence generated by “artificial intelligence” rather than all “machine-generated” evidence, a narrower category intended to capture systems whose outputs depend on predictions or inferences from data while excluding routine digital records, such as faxes, emails, and cellphone extractions, that raise none of the reliability concerns the rule was designed to address.²¹ Because the AI-specific definition does most of that work, the vague “simple scientific instruments” carve-out is being dropped from the revised draft. The accompanying Committee Note has also been expanded to direct judges toward additional reliability questions: whether an AI system’s conclusions can be reproduced or audited, whether the system has been evaluated by parties independent of its developer, and whether the opposing party has received enough information to test the system itself.

One structural feature has not changed. The revised rule continues to treat sponsoring expert testimony as the ordinary, but not mandatory, path to admissibility, permitting a court in “exceptional circumstances” to admit AI-generated evidence on other proof of reliability. Critics remain skeptical that questions about training data representativeness, validation methodology, and adversarial testability can meaningfully be answered without expert testimony, and the Committee’s fall 2026 mini-conference is expected to revisit that issue along with the rule’s ultimate scope.²²

Practical Considerations for Litigation Practices

First, litigators should not wait for Rule 707’s final text before treating unsponsored machine output as a live evidentiary issue. Existing Rule 702 and its state-law analogues already give opposing counsel a basis to challenge AI-derived conclusions offered without expert support, and courts confronting these questions today, as the Rittenhouse and Puloka examples illustrate, are already applying some version of the reliability scrutiny Rule 707 would formalize. Waiting for the rule to take effect before building an evidentiary strategy around AI-derived proof would leave a litigator behind the current state of the law, not ahead of it.

Second, parties relying on AI tools, whether in forensic analysis, damages modeling, medical diagnostics, or e-discovery, should begin documenting those systems’ training data, validation testing, and error rates now, regardless of whether Rule 707 is ultimately adopted. That documentation is the raw material any reliability showing will require under Rule 702 today or under Rule 707 in whatever final form it takes, and it is far easier to assemble contemporaneously than to reconstruct after a challenge is filed.

Third, litigators should anticipate that offering or opposing AI-derived evidence will increasingly trigger discovery disputes over access to the underlying system, along with confidentiality and trade secret objections from developers and deploying parties. Building a discovery and protective-order strategy for these disputes before they arise, rather than improvising once a motion to compel is filed, will save time and expense regardless of which evidentiary standard ultimately governs.

Fourth, practices with a significant litigation footprint should track the Advisory Committee’s fall 2026 mini-conference and any further revisions to Rule 707, since the rule remains genuinely unsettled and open to additional public input. Firms and clients with a stake in how AI evidence is treated at trial retain a meaningful opportunity to weigh in before the rule reaches final form.

Fifth, litigators should not assume that a stalled federal rule means the issue can wait. States including Louisiana, New York, and California have advanced their own AI-evidence measures on separate timelines, and because many state evidence codes track the Federal Rules of Evidence, state-level developments may move faster than, and ultimately influence, the federal rulemaking process itself.

Whatever becomes of Rule 707 in its current form, the debate it has generated reflects a broader and durable trend: courts and rulemakers are moving, deliberately if unevenly, toward treating AI-derived conclusions with the same skepticism traditionally reserved for expert opinion testimony. Litigators who build that expectation into their evidentiary practice now will be better positioned than those who wait for a final rule number to make it official.