AI evidence scrutiny rises

U.S. courts are moving to police artificial intelligence evidence more tightly as judges and lawyers confront fabricated citations and opaque software output. (uscourts.gov) The Judicial Conference’s rules process published a proposed Federal Rule of Evidence 707 on June 10, 2025, and took public comments through February 16, 2026. The draft says that when machine-generated evidence is offered without an expert witness, and the output would count as expert testimony if a person gave it, a judge may admit it only if it satisfies Rule 702. (uscourts.gov, uscourts.gov) The committee’s December 1, 2025 report says members had spent three years studying whether existing evidence rules were enough for artificial intelligence. It identified two problems: machine-generated output that looks like expert opinion, and deepfake audio or video that is hard to detect. (uscourts.gov) This is a fight over a basic courtroom question: when software produces an answer, who vouches for it. Proposed Rule 707 tries to stop parties from slipping in a machine’s conclusion through a witness who used the tool but cannot explain how reliable it is. (uscourts.gov) The debate sharpened this week after Sullivan & Cromwell apologized to a federal bankruptcy judge for a filing that contained about 40 inaccurate citations and other errors generated by artificial intelligence. Reuters reported the firm said the mistakes appeared in an emergency motion and were flagged after filing. (reuters.com) At a January 29, 2026 hearing on Rule 707, committee chair Jesse Furman said the panel was taking testimony on the proposed new rule while the comment period was still open. Witnesses included lawyers from Hausfeld, Exxon Mobil, Lawyers for Civil Justice and Loko AI, showing how broadly the issue now cuts across plaintiffs’ firms, defense groups and legal technology companies. (uscourts.gov, uscourts.gov) Some commenters urged caution. Testimony submitted for the January 15, 2026 hearing argued that Rule 707 was premature, questioned whether generative artificial intelligence would really be offered routinely without expert testimony, and said any new rule should focus more directly on “black-box” systems whose reasoning cannot be inspected. (uscourts.gov) The committee is still working on the proposal. A May 7, 2026 agenda book lists a reporter’s memorandum on “Machine Learning and Proposed Rule 707,” which means the drafting fight is continuing after the public hearings closed. (uscourts.gov) For companies building artificial intelligence systems, the legal standard is starting to look less like a product feature list and more like a chain-of-custody file. Courts are asking, in effect, for records that show what model produced an output, what data and prompts went in, and whether another person can reproduce the result before it reaches a jury. (uscourts.gov, reuters.com) The next step is not a ban on artificial intelligence in court. It is a higher burden to show that a machine’s answer is authentic, reliable and traceable before a judge lets it in. (uscourts.gov, reuters.com)