Authenticating AI and Deepfake Evidence Under FRE 901

Rule 901(a) poses one deliberately modest question: has the proponent offered enough to support a finding that an item is what it is claimed to be? The threshold is meant to be low. It screens out the plainly spurious and sends everything else to the finder of fact, who decides authenticity in the end. The court is a gatekeeper, not the final arbiter of whether a recording is genuine; it asks only whether a reasonable juror could so find.

Generative AI leaves that question intact. What it transforms is the expense and technical depth of the answer. A Slack thread, a voicemail attributed to an executive, footage of a server room at 2 a.m.: each can now be fabricated convincingly by anyone with a laptop and a few hours. Because the bar to admission is low, a competent fake may clear it, and the contest migrates from admissibility to weight. In a mediation that migration is decisive, because the parties, not a jury, are pricing the probability that an exhibit is exactly what it purports to be.

The federal Advisory Committee on Evidence Rules studied the problem over several years and, on its most recent review, declined to amend Rule 901. Its logic repays attention. Courts absorbed email, text messages, and social media without bespoke authentication rules, and a rule drafted against today's tools risks obsolescence before it takes effect. The Committee chose to let case law develop rather than legislate at a moving target.

Three traits of generative media put pressure on the conventional toolkit. The output now routinely exceeds what a lay observer, and often an expert, can detect by eye, because the models are trained adversarially to defeat detection. The technology is cheap and pervasive, so the supply of plausible fakes is effectively unbounded. And manipulation is frequently partial rather than wholesale, with synthetic elements grafted into otherwise authentic footage, which can quietly distort even a witness's own memory of the event.

A second-order problem should concern any neutral more than the first. As awareness of deepfakes spreads, a party can weaponize that awareness against real evidence, branding a genuine recording a fabrication. Bobby Chesney and Danielle Citron called this the liar's dividend: the better fakes become, the more plausible it is to dismiss the authentic as fake. Courts have begun to see the maneuver. In the litigation over the fatal Tesla crash involving Walter Huang, counsel floated the possibility that a years-old recording of Elon Musk was itself a deepfake; the court declined to let a public figure's prominence convert his recorded words into something presumptively suspect.

The counterexample is just as instructive. In Mendones v. Cushman & Wakefield, an Alameda County court grew suspicious of exhibits offered on summary judgment, noted looping video and the absence of facial expression, ordered production of complete file metadata, concluded the metadata had likely been altered after the fact, and imposed terminating sanctions. The court reached the right result because someone, in effect, performed the work of a careful forensic examiner.

A neutral begins with no presumption that an exhibit is real or counterfeit. The task is to build, or to probe, the foundation Rule 901 contemplates, using the methods courts already trust. The illustrations in Rule 901(b) are neither exhaustive nor mutually exclusive; they are meant to be combined: a witness with personal knowledge under (b)(1), comparison against a known authentic specimen under (b)(3), distinctive characteristics under (b)(4), voice identification under (b)(5), and proof that a process or system yields an accurate result under (b)(9).

In practice, the assessment of digital media turns on convergence. Metadata is read for internal consistency: capture device, lens, shutter values, creation and modification timestamps, file type, and the custody trail from device to production. No single field decides the matter, since metadata can be forged, but anachronisms and contradictions are telling. The media itself is examined for the signatures of synthesis: motion that does not obey physics, lighting at odds with the scene, looping artifacts, an incoherent compression history. Where they exist, provenance signals such as cryptographic hashes and content credentials either corroborate or undercut the file's account. And the technical record is checked against the human one: does the system that supposedly produced the footage actually behave the way the custodian describes?

The measure is not certainty. Conclusive proof of authenticity has never been a condition of admission. The neutral's role is to tell the parties candidly how strong the foundation is and where it is exposed, so each can value the claim with eyes open.

Two leading proposals would amend Rule 901 for synthetic evidence, and a neutral should know both, because parties will argue from them whether or not either is ever adopted. Paul Grimm and Maura Grossman would tighten Rule 901(b)(9) to require a valid and reliable result, with an added showing for conceded AI output, and would create a new Rule 901(c): once a challenger demonstrates that an item is more likely than not fabricated or altered, the proponent must show that its probative value outweighs its prejudicial effect, or it is excluded.

Rebecca Delfino's proposal also adds a Rule 901(c) but on a different design. Once a challenger makes a threshold showing of AI manipulation, the proponent must authenticate under 901(b) and supply additional proof of reliability, with the judge deciding admissibility under Rule 104(a). Delfino would shift the authenticity finding to the court and instruct jurors not to discount evidence merely because deepfakes exist in the world.

Both frameworks share an instinct already visible in the case law: a bare allegation will not do. Courts long ago stopped crediting reflexive my-account-was-hacked defenses without more, and the same discipline should govern it's-a-deepfake claims. Adopted or not, the burden-shifting logic gives a mediator a ready architecture for deciding who must come forward, and with what.

Authenticating synthetic evidence is costly. Qualified digital forensic examiners are scarce, their work can be expensive, and electronic discovery already consumes a large share of the litigation budget. That cost falls unevenly. A party who cannot fund an examination may have to surrender otherwise meritorious evidence, which is an access-to-justice problem as much as an evidentiary one.

Mediation can absorb pressure a courtroom cannot. A neutral with forensic fluency can frame the authenticity dispute at the outset, scope a proportionate inquiry rather than a scorched-earth one, and help the parties agree on what metadata, hashes, or custodial testimony would actually move the needle. Where the question is genuinely contested and outcome-determinative, the parties can stipulate to a single jointly retained examiner, sharing the cost and avoiding a duel of hired experts. That path is faster, cheaper, and often more accurate than litigating the point to verdict.

It is also why courts and parties increasingly turn to technical neutrals and special masters where authentication looms large. A neutral who already understands how generative models work needs no tutorial, which compresses a months-long adversarial fight into a focused inquiry. In a breach dispute, where the exhibits are digital by definition and the stakes turn on who knew what and when, that capacity is not a luxury. It is the difference between resolving the matter on the facts and resolving it on who could afford to prove them.