Authenticating AI-Generated Evidence Under FRE 901 and 902: Do the Rules Still Work?

Generative AI can now produce images, audio, and video that neither a juror nor an expert can reliably tell apart from the real thing. That has reopened a basic question: are Federal Rules of Evidence 901 and 902 still up to the job, or do we need a new rule built for synthetic media? Here is where the debate stands and what litigators should do regardless of how it resolves.

The authentication threshold under Rule 901 has always been deliberately low. The proponent need only produce evidence sufficient to support a finding that the item is what they claim it is; the trier of fact decides the rest. That standard has absorbed a century of new media without much strain. The worry now is that generative AI breaks the assumption underneath it: that a convincing photograph, recording, or video is, more likely than not, a record of something that actually happened. When a realistic fake costs nothing to make and even experts struggle to flag it, a lenient gatekeeping standard starts to feel thin.

I get asked a version of this question in nearly every matter where audio or visual evidence is in play: do the existing rules still work, or are we waiting on a new one? The honest answer is that the rules are holding up better than the alarm suggests, but the practice around them has to change now whether or not the text ever does.

Why AI evidence is not just another doctored photo

Courts are no stranger to manipulated images, and one school of thought holds that deepfakes are simply the latest entry in a long line of fakery the system has always managed. There is real force to that argument. But three features set AI-generated content apart from the airbrushed photo or the spliced tape, and they are why the debate is live rather than settled.

• Fabrication, not just distortion. Traditional manipulation alters something real. Generative models can invent an event whole cloth, leaving no underlying original to compare against.
• Detection is genuinely hard. The same adversarial training that makes models realistic also trains them to evade the detection tools built to catch them. Lay observers cannot reliably tell, and even forensic analysis can be inconclusive.
• It is cheap and ubiquitous. Anyone with a laptop can produce a passable fake in minutes, which changes the volume and the calculus in a way that doctored darkroom prints never did.

The flip side of cheap, convincing fakes is what scholars have called the liar's dividend: as everyone learns deepfakes exist, it becomes easier to dismiss authentic evidence as fabricated. The threat to litigation is not only that fakes get admitted, but that genuine recordings get waved away with an unsupported cry of deepfake.

The proposed-rule debate

Several proposals have been put to the Advisory Committee on Evidence Rules to address this directly. Two are worth knowing because they frame the choices.

A burden-shifting trigger for challenged evidence

The leading proposals would add a new subsection to Rule 901 that kicks in only when a party makes a real showing that an item is more likely than not fabricated or altered by AI. A bare assertion would not be enough; the challenger has to put forward supporting evidence first. Once that threshold is met, the burden shifts to the proponent, either to show the evidence's probative value outweighs its prejudicial effect, or, in a stricter variant, to authenticate it under Rule 901(b) and supply additional proof of reliability. The point of the trigger is to filter out reflexive deepfake objections while giving courts a structured path for the serious ones.

Reliability, not just authenticity, and a bigger role for the judge

A related strand argues that authenticity alone no longer guarantees the evidence is genuine, so the proponent of challenged audiovisual material should have to demonstrate reliability as well, with the judge rather than the jury deciding the question. The reasoning is that jurors take what they see at face value more readily than judges do, so the gatekeeping function should expand. This overlaps in spirit with proposals to handle AI and other machine-generated output as a reliability problem akin to expert evidence under Rule 702, an approach sometimes floated as a new Rule 707 for machine-generated evidence.

Why the committee chose to wait, and why that is defensible

For now, the Advisory Committee has declined to amend Rule 901, choosing a wait-and-see posture until courts confront these disputes often enough to show where the existing rules actually fail. That restraint is easy to criticize as too slow, but it has good precedent. When emails, texts, and social media arrived, the committee studied special authentication rules and ultimately decided the existing 901(b) illustrations were flexible enough. Courts then handled hacked-account objections by refusing to credit blanket claims without supporting proof, exactly the move that defuses a blanket it's-a-deepfake objection today. Rulemaking is also slow by design, and a rule written to today's technology risks being obsolete before it takes effect.

The early case law, though thin, is encouraging on this point. Courts have refused to let well-known figures dodge their own recorded statements by invoking the mere possibility of deepfakes, have demanded full metadata and forensic detail when exhibits looked synthetic, and have weighed deepfake claims against the totality of the circumstances rather than in isolation. The common thread is judicial skepticism applied in both directions, which is what the existing framework asks for.

What litigators and courts should do now

Whether or not a new rule ever arrives, the work in front of practitioners is the same. Treat authentication of digital media as a forensic problem, not a foundational formality, and build the record early.

1. Preserve and demand the native files and full metadata, not exports or screenshots. Creation and modification data, device and application fingerprints, and file structure are where AI generation and later tampering usually surface.
2. Decide early whether you need a digital forensic examination, and budget for it. Proving an item is authentic or fabricated takes specialized analysis, and qualified experts are not abundant, so this is a cost to scope at intake rather than discover on the eve of trial.
3. Hold the line on baseless deepfake objections. An unsupported assertion that evidence is AI-generated should be treated like an unsupported claim of hacking: it requires a threshold showing before it triggers any inquiry, and frivolous versions implicate counsel's ethical duties.
4. Pin down AI and authentication issues in the ESI protocol. Agreeing up front on native production, metadata fields, and how challenges to synthetic media will be handled avoids a mid-case scramble when the stakes are highest.
5. Consider a neutral or special master where authentication is central. A court-appointed technical neutral can interpret competing forensic analyses, narrow the dispute, and often resolve it more efficiently than dueling experts alone.

There is also a resource dimension worth naming. Forensic authentication is expensive, and a regime that turns every contested image into an expert battle risks pricing out the party with the weaker checkbook. That is a fairness concern for courts and a strategic one for litigators, and it is one more reason to scope the forensic question deliberately rather than letting it expand on its own.

The bottom line

Rule 901 and Rule 902 were not written with synthetic media in mind, but they are more adaptable than the panic implies, and courts so far are applying them with the right instincts. A targeted amendment may yet prove necessary once the case law matures; if it does, the burden-shifting and reliability proposals are the place to start. In the meantime, the cases turn not on the text of the rule but on whether someone did the forensic work to back up, or knock down, the claim that a file is what it purports to be. If your matter hinges on the authenticity of an image, recording, or video, the time to engage that analysis is before the motion lands. You can start a scoping conversation through our home page or email the team directly to discuss a conflict check and approach.