Rule 707: AI evidence debate
- Lawyers are examining a proposed Federal Rule of Evidence 707 to address admissibility of machine‑generated evidence. - The proposed rule would add new considerations for federal practitioners handling AI outputs in courtrooms. - If adopted, organisations will need better logging, provenance, and audit trails for machine‑produced decisions. (reuters.com)
Federal judges’ rulemakers have published a proposed Rule 707 that would make some machine-generated evidence clear a Rule 702-style reliability screen before it comes into evidence. (uscourts.gov) The proposal covers machine output offered without a human expert when that output would count as expert testimony if a person had said it. The text says a court “may admit” it only if it satisfies Rule 702’s requirements. (uscourts.gov) The Judicial Conference’s Standing Committee approved the proposal for public comment on June 10, 2025, alongside other evidence-rule changes. A public hearing on new Rule 707 was held on January 29, 2026, in Washington. (uscourts.gov 1) (uscourts.gov 2) In plain terms, the fight is about whether a machine’s answer can come into court the way a lab result or valuation model might. Rule 702 is the gatekeeping rule judges already use to test whether expert evidence rests on sufficient facts and reliable methods. (uscourts.gov) The committee report says it is trying to solve one of two artificial-intelligence evidence problems now on its agenda: machine-generated evidence that functions like expert opinion. A separate problem—deepfakes and other fake audiovisual evidence—is being handled through possible changes to authentication rules instead. (uscourts.gov) (thesedonaconference.org) That distinction matters for lawyers using software to estimate value, classify images, predict outcomes, or flag patterns in large datasets. If the output is doing the work of an expert, the proposal would push judges to ask how the system was built, tested, and validated before jurors hear it. (uscourts.gov) Commenters have pressed for more detail on what parties must disclose to support or challenge that showing. Written testimony submitted for the January 2026 hearing said the draft does not spell out what information about a model must be produced, when it must be produced, or how Rule 707 should fit with civil discovery rules. (uscourts.gov) Others have raised the opposite concern: that the rule may be too narrow if it applies only when no expert witness is called. The May 2026 agenda book says the reporter flagged an article arguing Rule 707 should reach machine-generated evidence even when an expert sponsor is on the stand. (uscourts.gov) The Evidence Rules Advisory Committee was still refining the proposal in May 2026, including possible note language on “explainability,” the plain-English idea that a party should be able to show how a system reached its answer. The committee materials say that issue came up specifically in the Rule 707 discussion. (uscourts.gov) For companies and agencies that rely on automated scoring or model-driven decisions, the practical record is becoming as important as the output itself. The hearing materials and committee papers point toward the same courtroom question: not just what the machine said, but what logs, testing records, and provenance can prove about how it got there. (uscourts.gov 1) (uscourts.gov 2)