Authenticating AI-touched evidence: do we need a new rule?

The Federal Rules of Evidence set a deliberately low bar for getting a photo, video, or audio file in front of a jury. Under Rule 901(a), the proponent need only “produce evidence sufficient to support a finding that the item is what the proponent claims it is.”^[1] That is a screening standard, not a proof standard: the judge decides whether a reasonable juror could find the item genuine, and the jury decides whether it actually is. For a century that lenient gate worked, because fabricating convincing audiovisual evidence was hard, expensive, and usually detectable.

Generative AI removes all three frictions at once. Realistic fakes are now cheap, fast, and — to a layperson and increasingly to experts — hard to distinguish from the real thing. The question consuming the Advisory Committee on Evidence Rules is whether a rule written for the analog era needs surgery, or whether judges already have the tools.

How 901 and 902 actually work today

Rule 901(b) supplies a non-exhaustive menu of ways to clear the bar. The provision most often invoked for machine output is 901(b)(9): “Evidence describing a process or system and showing that it produces an accurate result.”^[2] This is the workhorse for surveillance footage — a custodian testifies that the camera system runs as described and is reliable, and the video comes in. Other subsections do the rest of the heavy lifting: a witness with personal knowledge under 901(b)(1), comparison by an expert or the trier of fact under 901(b)(3), or “distinctive characteristics” under 901(b)(4).

Rule 902 goes further, making certain records self-authenticating so no live witness is needed. Two 2017 additions matter here: 902(13), covering “certified records generated by an electronic process or system,” and 902(14), covering “certified data copied from an electronic device, storage medium, or file,” authenticated by digital identification such as hash comparison.^[3] Notice the shared premise running through 901(b)(9), 902(13), and 902(14): a system that “produces an accurate result.” That phrase assumes the system is faithfully recording reality. A generative model is engineered to do the opposite — to synthesize output that only looks like a recording of reality.

The deepfake problem, and the liar's dividend

The danger runs in two directions. The first is the obvious one: fabricated evidence that slips through a permissive gate. Because the 901(a) threshold is so low, a competent fake can satisfy it. The second is subtler and, so far, more common in real dockets — the “liar's dividend,” a term popularized by law professors Bobby Chesney and Danielle Citron for the way pervasive awareness of deepfakes lets bad actors dismiss genuine evidence as fake.^[4]

Courts are already seeing both. In Huang v. Tesla, lawyers for Elon Musk argued that an authentic, years-old recorded statement might be a deepfake; the judge rejected the gambit as “deeply troubling,” refusing to let public figures hide behind the mere possibility of fakery.^[5] Running the other way, in Mendones v. Cushman & Wakefield (Alameda County Superior Court, 2025), the court suspected that summary-judgment exhibits were AI-generated — flagging looping video and absent facial expressions — ordered the plaintiffs to produce full metadata, and ultimately imposed terminating sanctions after concluding the videos were deepfakes.^[6] The lesson of Mendones is encouraging: a careful judge, asking for the right technical record, caught the fake.

The proposals on the table

Two reform tracks dominate the debate. The first, advanced by retired U.S. District Judge Paul W. Grimm and Dr. Maura R. Grossman, would amend 901(b)(9) so that a process or system must show it produces a “valid and reliable” result — not merely an “accurate” one — and would add a new 901(c) creating a burden-shifting test: once a challenger shows it is more likely than not that electronic evidence was fabricated or altered, the item is admissible only if its probative value outweighs its prejudicial effect.^[7] A parallel proposal from Professor Rebecca Delfino would likewise add a 901(c) but require the proponent to authenticate under 901(b) and prove reliability, relocating the call from the jury to the judge.^[8]

The second track moved faster, but on different terrain. In June 2025 the Advisory Committee approved publication of an entirely new Rule 707, “Machine-Generated Evidence,” which went out for public comment from August 15, 2025 through February 16, 2026.^[9] Rule 707 targets a distinct problem from deepfakes: machine output offered without a human expert — say, an algorithm's conclusion presented directly to the jury. It would subject that output to the same reliability scrutiny that Rule 702 applies to expert testimony.^[10] The Department of Justice, the lone dissenting vote, argued Rule 702 already reaches such evidence. Notably, the Committee declined for now to advance a deepfake-specific 901(c), concluding existing tools suffice given how few deepfake disputes have actually reached the courts — while keeping the draft on the shelf should that change.^[7]

The argument for patience

There is a real case for the Committee's wait-and-see posture. In 2014 the same body considered, and rejected, special rules for authenticating emails, texts, and social-media posts — and in hindsight that restraint looks correct, because courts adapted 901's existing categories without much trouble.^[11] Critics such as Riana Pfefferkorn note that courts are “no stranger to doctored photographs” and have absorbed every prior wave of fakery without the sky falling.^[12] Rulemaking is slow — typically three years — and a rule drafted to today's model architecture could be obsolete before it takes effect.

What practitioners should do now

Whether or not the rules change, the practice has already changed. Three moves pay off regardless of how the rulemaking lands:

1. Build the authentication record at creation, not at trial. Capture and preserve hashes, signed provenance (C2PA-style Content Credentials), and full metadata. A 902(14) digital-identification certificate is worth far more than a witness's after-the-fact memory. See /provenance.
2. Treat “it's a deepfake” as a claim that needs proof. Courts have rejected bare hacking and fakery assertions for years; a challenge should be backed by metadata, expert analysis, or visible artifacts — exactly the record the Mendones court demanded.
3. Budget for the expert, and for the access-to-justice gap. Forensic authentication of contested audiovisual evidence is expensive and qualified experts are scarce, which risks pricing out under-resourced litigants. Plan for it early.

The honest answer to “do we need a new rule?” is: maybe, but not yet for deepfakes, and possibly soon for machine output. Either way, the durable defense is the same one provenance has always provided — evidence that can prove what it is, rather than evidence we are merely asked to believe.