Deepfake Evidence in Court: How the Rules Are Adapting

How Courts Currently Handle Deepfake Challenges

When deepfake evidence—a photograph, video, or audio recording fabricated or altered using generative artificial intelligence—is offered for trial, it must clear the same authentication hurdle as any other audiovisual evidence under Federal Rule of Evidence 901. That rule requires the proponent to produce evidence "sufficient to support a finding that the item is what the proponent claims it is."[11] For audiovisual recordings, courts apply Rule 104(b), which sets a threshold called "prima facie" proof: the evidence need only be such that a reasonable jury could find authenticity based on it, not that the judge itself must be convinced.[15]

This standard, developed for a world of physical forgery and intentional misrepresentation, becomes fragile when the evidence at issue could have been fabricated by software that is accessible, constantly improving, and difficult to detect. The Advisory Committee on Evidence Rules has recognized that generative AI—artificial intelligence capable of creating new content rather than simply classifying or analyzing existing content—presents a problem the rules did not contemplate.[1][2]

The Authentication Problem Deepfakes Create

Deepfakes are rapid to produce and increasingly difficult to distinguish from authentic recordings. Traditional authentication methods—asking a witness to vouch for having made the recording, comparing it to known authentic versions, or pointing to distinctive characteristics—all assume that human perception or specialized knowledge can identify fabrication.[11] But as generative AI algorithms improve, that assumption weakens.[1]

The Advisory Committee Reporter has observed that the existing Rule 104(b) standard could be viewed as insufficiently protective against deepfake evidence precisely because any member of the public could now fabricate a realistic recording.[1] A jury asked whether a video "looks real" may not be able to detect a high-quality deepfake, and traditional authentication testimony may no longer provide reliable assurance.

At the same time, some scholars have argued that the problem is not new. Professor Amy Wexler has contended that courts have long managed forgery challenges under existing authentication standards and that deepfakes present comparable concerns that the current rules can address.[8] This debate—whether to amend the rules or rely on existing doctrine—has shaped the Committee's deliberations.

Proposed Rule 901(c): The Burden-Shifting Framework

In response, the Advisory Committee has drafted a new Rule 901(c) that would create a two-step burden allocation for evidence challenged as a deepfake.[1][2]

How the Proposed Rule Would Work

Under the proposal, an opponent who challenges an item of evidence as fabricated by generative artificial intelligence would first bear the burden of presenting evidence sufficient to support a finding that the item was so fabricated—in other words, meeting the Rule 104(b) threshold.[1][2] This is the same low bar that ordinarily applies to authentication, but now it must be cleared by the challenger rather than the proponent.

If the opponent clears that threshold—establishing a reasonable basis to believe fabrication occurred—the burden would then shift to the proponent. But the proponent's burden becomes heavier: it must show, by a preponderance of the evidence (more likely than not), that the item is authentic.[1][2] This is a Rule 104(a) standard, which requires that the judge itself be persuaded, rather than simply that the jury could be persuaded.[15]

This framework replaces the ordinary Rule 901(b) standard for that particular piece of evidence. Once a deepfake challenge is raised and satisfied under the proposed Rule 901(c), the lower prima facie threshold no longer applies. The proponent must clear a higher bar.[2]

Scope and Definition

The Committee has refined the language over successive drafts. The current version covers claims that an item of evidence was fabricated by "generative artificial intelligence," not artificial intelligence generally.[2][10] This distinction is important: the rule targets evidence that is entirely fabricated or substantively altered by AI, not evidence that is an acknowledged AI product (such as a machine-generated report or prediction) offered for what it actually is.

The proposed rule also includes a notice provision, ensuring that deepfake challenges are raised and addressed before trial rather than springing them as a surprise.[2] And it incorporates a definition of "generative artificial intelligence" into the rule text itself, establishing a consistent meaning across the Evidence Rules.[2]

Distinguishing: Machine-Generated Evidence and Rule 707

Parallel to the deepfake amendment, the Advisory Committee has also drafted a separate proposed Rule 707 to address machine-generated evidence.[5] This addresses a distinct problem.

Deepfakes are evidence claimed to be inauthentic because they are fabrications difficult to detect. Machine-generated evidence, by contrast, is evidence that is acknowledged to be a product of artificial intelligence, such as a prediction from a machine-learning algorithm, output from a statistical model, or an automated report.[5] Such evidence raises different admissibility questions—typically governed by Rule 702, which sets standards for expert testimony—because the question is not whether the item is what it purports to be, but whether the process that created it is reliable enough to be worth hearing.[5]

Rule 707 would set standards for offering machine-generated evidence without a testifying expert.[5] The proposed deepfake amendment and Rule 707 thus address different evidentiary problems, though both arise from the same technological reality.

Provenance at Capture: The Practical Defense

While rule amendments work through the system, the Advisory Committee materials acknowledge a practical solution available under existing law: establishing the complete provenance—the origin, creation, and custody record—of the evidence at the moment it was captured.

Metadata embedded in a file at the time of recording, cryptographic hash values, server logs, camera certification data, time-stamped backup records, and file integrity certificates created before any opportunity for fabrication all serve the same purpose: they make tampering after the fact either technically infeasible or demonstrably detectable.[1][2][4] This evidence is often more resistant to deepfake challenge than human perception or even visual comparison, because it does not depend on detecting sophisticated fabrication—it creates a technical record that would have to be forged separately from the audiovisual content itself.

The Committee has recognized that watermark technologies designed to embed authentication metadata into media files face their own vulnerabilities, as generative AI techniques improve.[1] But a robust record of provenance—particularly one created at the time of capture and maintained in a reliable custody chain—remains the strongest practical defense against fabrication allegations.

Authentication under Rule 901(b) can be established by opinion evidence, comparison to known authentic records, or distinctive characteristics.[11] But when deepfake allegations are raised, provenance evidence created at capture provides a distinct foundation: it demonstrates not merely that a witness says the recording is real, but that the technical conditions under which it was made, stored, and transmitted create a record that would have to be forged alongside the content itself.

Current Status and Timeline

As of May 2026, the Committee had not approved the proposed Rule 901(c) for publication in the Federal Register.[2] The decision to defer publication reflected the Committee's desire to continue refining the language and allowing further expert input.[2][3]

In October 2026, the Committee announced plans to host a mini-conference bringing together AI experts, forensic practitioners, and legal scholars to assess the deepfake rule proposal before deciding whether to move it toward final publication.[3] This step reflects the Committee's recognition that rule-making in a rapidly evolving field requires sustained consultation with practitioners and technologists.

Until a rule amendment is published and the amendment process is complete, courts continue to handle deepfake challenges under existing standards. This means the burden remains on the proponent to authenticate audiovisual evidence, and the threshold is the prima facie standard of Rule 104(b). A court may consider all relevant evidence—technical provenance, metadata, cryptographic verification, expert analysis, witness testimony, and circumstantial evidence—in deciding whether a reasonable jury could conclude that the recording is authentic.

Why the Stakes Matter

The deepfake authentication problem is not a specialized concern of laboratories and forensic examiners. It reaches criminal prosecutions (where a fabricated confession or threat could alter guilt or innocence), civil litigation (where altered contracts or recordings could reshape liability), and administrative proceedings (where false evidence could overturn decisions about licensing, benefits, or employment).

The current rules assume that evidence has been made and exists in some stable form; the question is whether it is what it claims to be. Deepfakes complicate that assumption by making the creation and alteration of evidence technically trivial. The law's response—whether through rule amendment or through rigorous application of existing standards to technical provenance—must account for that shift. Until amended rules take effect, practitioners and courts must work within the framework that now exists: the authentication rules of 901 and 902, applied with full awareness of the new threats to the integrity of audiovisual evidence.