Proposed Federal Rule of Evidence 707: How Courts Would Screen Machine-Generated Evidence Offered Without an Expert

Proposed Federal Rule of Evidence 707 would bar a federal court from admitting machine-generated evidence offered without an expert witness unless that output satisfies the reliability requirements of Rule 702. It is a proposal, not law: published for public comment in August 2025, it drew 59 written comments, and on May 7, 2026 the Advisory Committee on Evidence Rules declined to advance it.

The proposal matters to legal and compliance teams less for what it says than for what it assumes — that software increasingly delivers conclusions that would, if a person said them from the witness stand, be treated as expert opinion and screened accordingly. The gap the drafters identified is a real one in federal practice. Whether Rule 707 is the right patch is now an open question inside the rulemaking process itself.

What would Rule 707 actually require?

The published proposal ties machine output to the existing expert-testimony gate. As quoted in the Advisory Committee on Evidence Rules' report dated December 1, 2025: "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."

Rule 702 is the federal standard for expert testimony. In its current form, amended effective December 1, 2023, it requires the proponent to establish that the specialized knowledge will help the trier of fact, that the testimony rests on sufficient facts or data, that it is the product of reliable principles and methods, and that those principles and methods were reliably applied to the facts of the case.

Rule 707 does not create a new reliability test. It redirects an existing one. The drafters' stated concern, per that December 2025 report, is that where "machine or software output is presented without the accompaniment of a human expert," Rule 702 "is not obviously applicable" even though the reliability stakes are the same. The report frames the conclusion bluntly: it cannot be that a proponent evades Rule 702's reliability requirements by offering machine output directly.

Two consequences follow from that drafting choice. The burden sits with the proponent, as it does under Rule 702. And the screening happens at admissibility, before a jury ever weighs the output.

Which evidence would the rule reach?

The proposal's reach is defined by a counterfactual, not by a technology list: would this output be subject to Rule 702 if a witness said it? That formulation is deliberately technology-neutral, and it is also the source of most of the objection the proposal attracted.

Nothing in the published text names generative models, risk-scoring engines, or forensic software. The category is functional. Output that performs the analytical work of an expert — an estimate, a classification, a probabilistic match — falls inside; output that merely records or transmits does not perform that work.

That boundary was contested throughout the comment period. Materials in the Committee's May 2026 agenda book, reporting the Committee's November 5, 2025 discussion, record open questions about whether the rule should apply when an expert does testify, whether an exemption for simple scientific instruments belongs in the text, whether independent validation and research-licensing access should bear on admissibility, and whether proprietary government software would be disadvantaged.

Those are not stylistic quibbles. Each one moves the line between a thermometer reading and a model output, and practitioners handling e-discovery analytics, transaction-monitoring alerts, or automated document review sit on both sides of it.

Where does the proposal stand as of August 2026?

Rule 707 is stalled at the advisory-committee stage, deliberately. The Committee met on May 7, 2026 in Washington, D.C., and its report dated May 17, 2026 states that "the Committee decided not to propose that the revised Rule 707 be released for a new period of public comment at this time," preferring instead to have the proposed rule "vetted at the Committee's Fall meeting by technology experts and others in the field of AI and law."

The federal rulemaking sequence explains why that is a meaningful pause rather than a defeat. The judiciary's own guidance notes that proposed amendments generally take about three years to move through advisory committee review, Judicial Conference consideration, and Supreme Court and congressional review before taking effect.

The comparison is the clearest available measure of how far apart the two proposals ended up. Both were published on the same day under the same comment deadline; only one emerged from the comment period intact.

What did commenters object to?

The comment record was split rather than hostile. The May 2026 report characterizes the 59 written comments as mixed: "only 3 comments expressed unqualified support; 27 supported the rule subject to revisions; and 27 opposed it." Roughly nine in ten commenters who engaged with the text wanted it changed or dropped.

Testimony was taken at a virtual hearing on January 15, 2026, running from 10:00 a.m. to 11:30 a.m. Eastern, with witnesses including Thomas Allman, Mary D'Agostino of Hancock Estabrook, Alex Dahl of Lawyers for Civil Justice, Jeannine Kenney of Hausfeld, Robert Levy of Exxon Mobil, and Joseph Zaki of Loko AI. A second Evidence Rules hearing was scheduled for January 29, 2026.

Written objections ran along recognizable lines. A comment filed February 16, 2026 in docket USC-RULES-EV-2025-0034 by four submitters from the Brooklyn Law Incubator and Policy Clinic argued that "the rule is premature given AI technology's rapid evolution and our still-developing understanding of machine learning systems," and further contended that undefined terminology could sweep in routine electronic evidence and that existing rules already supply adequate safeguards.

The Committee's response was structural rather than defensive: its May 2026 report describes three alternative versions of the rule prepared for the meeting, with one identified as the leading candidate should republication follow. Republication would restart the comment clock.

What this means in practice

Nothing in Rule 707 binds any court today, and the proposal governs federal proceedings only; state evidence codes are separate instruments on their own timetables. Still, the record supports several operational observations for legal-ops and compliance functions.

1. Reliability documentation is the durable asset. Every version of the proposal on the table routes machine output through Rule 702's factors, so validation records, error rates, version histories, and the identity of the person who can speak to the method retain value regardless of which text, if any, is adopted.
2. Vendor assurances are not validation. Under the proposal's logic, a supplier's claim about its own model's accuracy is the supplier's claim; the proponent carries the admissibility burden. Procurement terms that secure access to validation data and to a qualified witness are the practical hedge.
3. The fall 2026 meeting is the next real checkpoint. Because the Committee expressly deferred to that meeting for expert vetting, teams tracking this proposal can watch one agenda rather than a diffuse news cycle.
4. The timeline is long even if the rule revives. With republication, a further comment period, and the standard multi-year approval sequence, an effective date is not near. Systems being procured now will be litigated under existing rules first.

This article is information, not legal advice. Admissibility questions turn on jurisdiction, procedural posture, and the specific record; organizations should consult qualified counsel about their own facts.