Forging a convincing audio recording used to take a studio, an expert, and money. Now it takes a laptop and a few minutes of someone's voice, as Joel demonstrates to his guest by playing a cloned version of the guest's own voice. "Now," says Judge Paul Grimm, "we've democratized fraud." The flip side is the "liar's dividend": once everyone knows fakes are easy, a witness confronted with a real recording can simply say it is fake, and a jury may believe her.
In this TalksOnLaw interview, recorded in December 2024, Joel Cohen speaks with the Honorable Paul W. Grimm, a former United States District Judge for the District of Maryland, then Director of the Bolch Judicial Institute at Duke Law School, and the country's most influential judicial voice on electronic evidence, about how the Federal Rules of Evidence handle AI-generated and AI-altered evidence, where they fail, and what he proposes to do about it.
Why the Current Rules Are Not Enough
Grimm explains the mechanics: authentication under Rule 901(a) requires only evidence "sufficient to support a finding," which under Rule 104(b) sends a disputed recording to the jury on a bare showing; Rule 403 balancing is discretionary; the process-or-system provision of 901(b)(9) and self-authentication under 902(13) were written for machines that do not learn; and "hearsay is completely inapplicable because hearsay requires a human declarant." The result is that a jury of lay people, which cannot distinguish a deepfake, is asked to decide whether one exists, after watching a video it cannot unsee.
Algorithms Already in the Courtroom
He describes AI that judges already rely on, from the COMPAS risk-assessment tool used in sentencing and bail (and the Wisconsin case that upheld it despite a secret algorithm) to the Washington state court that excluded AI-enhanced video in Puloka, and he insists on the upside: with guardrails, generative tools can draft a discrimination complaint in minutes and close part of the access-to-justice gap, work he is pursuing with Duke's RAILS initiative. He explains why an AI system's error rate, not its brand, is the reliability question a court should be asking.
A Proposed Fix
Grimm lays out the rule he and Professor Maura Grossman have proposed: when a party makes a showing that a jury could reasonably find an item of evidence was fabricated or altered by AI, the judge, not the jury, should decide admissibility under a balancing test that weighs probative value against the risk of deception. He also points to Maryland's twenty-year-old rule requiring pretrial disclosure of computer-generated evidence as a model, and closes with advice to lawyers on handling AI evidence before a rule arrives.
What to Know Now
The rulemaking Grimm describes has moved, but not to a finish. The Judicial Conference's Advisory Committee on Evidence Rules published proposed Rule 707, which would subject machine-generated evidence to Rule 702-style reliability requirements, for public comment in August 2025; after the comment period closed in February 2026, the Advisory Committee (May 2026) and the Standing Committee (June 2026) sent the proposal back for revision, so the earliest effective date is now December 2028. The Grimm–Grossman proposal for a Rule 901(c) on potentially fabricated evidence was tabled in May 2026, with the committee concluding that existing rules could handle the problem for now. States have moved faster: Louisiana's Act 250 (2025) requires pretrial notice of AI-generated evidence, Illinois has adopted a judicial AI policy, and in Mendones v. Cushman & Wakefield (Cal. Super. Ct. Sept. 2025) a court dismissed a case with prejudice after finding that the plaintiffs' video evidence was AI-fabricated. The problem has grown in parallel with hallucinated citations: by mid-2026 a running tally counted more than 1,600 court decisions worldwide addressing fake AI-generated authorities, more than 500 of them in the United States in 2025 alone. A Federal Judicial Center survey found 55 percent of judges favor a rule amendment. Judge Grimm stepped down as Director of the Bolch Judicial Institute at the end of 2025; his article with Professor Grossman setting out the proposed rule appears in the Columbia Science & Technology Law Review (2025).
