Deepfake Evidence in Court: Rule 707 and 901(c) (2027)

As of November 5, 2026, no Federal Rule of Evidence addresses deepfakes or AI-generated evidence by name: draft Rule 901(c) has never been published for public comment, and proposed Rule 707 was published in August 2025 and then withdrawn for further study. Courts decide these disputes under the authentication rules they already have, starting with Federal Rule of Evidence 901(a) and its state counterparts.

Status as of November 5, 2026: Neither proposal is law, and neither is out for public comment. The Advisory Committee on Evidence Rules had an October 15, 2026 meeting on its calendar in Boston, and no minutes or report from that meeting has been published. Its next meeting is April 14, 2027.

This page tracks the rules on proof in court, and we update this URL in place. It does not cover whether making or sharing a deepfake is a crime; that is on our deepfake laws by state hub. For the history of the Committee's October agenda, see our dated news report.

Where Rule 901(c) and Rule 707 Stand

How Authentication Works Today: Federal Rule of Evidence 901

Federal Rule 901(a) applies to any item of evidence, with no separate track for AI. It provides: "To satisfy the requirement of authenticating or identifying an item of evidence, the proponent must produce evidence sufficient to support a finding that the item is what the proponent claims it is."

The rule's list of ways to meet that standard is "examples only" and "not a complete list." Four examples are especially relevant to video and audio:

• Rule 901(b)(1): testimony from a witness with knowledge "that an item is what it is claimed to be."
• Rule 901(b)(4): "appearance, contents, substance, internal patterns, or other distinctive characteristics of the item, taken together with all the circumstances."
• Rule 901(b)(5): an opinion identifying a voice, "based on hearing the voice at any time under circumstances that connect it with the alleged speaker."
• Rule 901(b)(9): evidence "describing a process or system and showing that it produces an accurate result."

Rule 902 lets some items authenticate themselves. That includes certified records from an electronic process under Rule 902(13) and certified data copied from a device under Rule 902(14), each with notice to the other side.

The Committee's May 2026 report calls Rule 901(a) "a low standard of admissibility." The standard does not ask a judge to rule out fabrication. Iowa's high court made the same point about its matching rule in October 2026: the State "didn't have to prove that the video probably was genuine to get it admitted," though "as a practical matter, it had to convince the jury beyond a reasonable doubt that the video was legitimate to obtain a conviction," because its case depended on the video.

Two other rules matter in a deepfake dispute. Rule 1003 admits a duplicate like an original "unless a genuine question is raised about the original's authenticity or the circumstances make it unfair to admit the duplicate." Rule 403 lets a court exclude relevant evidence when its value is substantially outweighed by dangers such as confusing the issues or misleading the jury.

What Draft Rule 901(c) Would Do

Draft Rule 901(c), titled "Potentially Fabricated Evidence Created By Generative Artificial Intelligence," would set up a two-step process. It is printed at pages 150-151 of the October 2026 agenda book.

1. The challenger goes first. A party claiming an item was fabricated by generative AI "must present evidence sufficient to support a finding of fabrication to warrant an inquiry by the court."
2. Then the burden rises. If the challenger makes that showing, "the item of evidence is admissible only if the proponent demonstrates to the court that it is more likely than not authentic."

The draft also applies to items offered under either Rule 901 or Rule 902. Unless the court orders otherwise, the challenger "must provide reasonable pretrial notice." It defines generative AI as a system creating "derived synthetic content, including images, video, audio, text, and other digital content."

The draft Committee Note draws two limits. First, "a broad claim of 'deepfake' is not enough." Second, the rule "does not apply if the preparation and issuance of a deepfakes is the gravamen of the claim or charge," so a prosecution for making the fake itself stays under the general rules.

The Note also says what would change. Once the challenger clears the first step, witness testimony under Rule 901(b)(1) "will not ordinarily prove more likely than not that the item is authentic."

Why the draft has not moved

The Committee's own words point in two directions. The agenda book says the Committee "has prepared, but not finally approved" the draft. The May 2026 report says "at least for now, an amendment to Rule 901 to address deepfakes is not warranted," because "few examples exist of courts having to address the possibility of deepfakes," and describes holding a draft ready "in abeyance."

The Committee's numbers come from a Federal Judicial Center survey summarized in the draft May minutes. There was a 45 percent response rate, and "Of the 931 judges who responded, only 15 reported having encountered a deepfake challenge." Of the rest, 82 percent said they "would require some initial showing before entertaining a claim of fabrication." A narrow majority of respondents without deepfake experience also said the rules should be amended to address deepfakes.

The Department of Justice opposed moving ahead. Its representative, Ms. Shapiro, said "the DOJ does not believe Rule 901(c) is ready for publication" and that the existing rules are adequate. That view is hers, not the Committee chair's. The chair, Judge Furman, said no vote on public comment was needed for now, and that considering Rules 707 and 901(c) together "would allow for potential publication of both proposed AI rules together."

One DOJ concern was child-exploitation cases, where the government offers AI-generated images it says are AI-generated. The Reporter answered that the draft would not cover images "when the government concedes that the images are fabricated."

The Reporter's October memo also suggests factors for the first step, as a possible addition to the Note. They are not rule text. The list includes missing or anomalous metadata, inconsistent timestamps, unexplained edits, absence of the native file, gaps in chain of custody, audio or synchronization glitches, known use of generative AI tools by the proponent, and independent evidence inconsistent with the item.

Proposed Rule 707 and Its Withdrawal for Further Study

Proposed Rule 707 was about reliability, not forgery. It addressed machine output offered with no expert to explain it. The published text read: "When machine-generated evidence is offered without an expert witness and would be subject to Rule 702 if testified to by a witness, the court may admit the evidence only if it satisfies the requirements of Rule 702(a)-(d)."

The Committee approved publication 8-1, with the Department of Justice voting against, and comments closed February 16, 2026. The draft May minutes record 59 written comments by the deadline: 3 in unqualified support, 27 supporting it with revisions, and 27 opposed.

The October agenda book then says the Committee "voted to withdraw the proposed rule as issued for public comment, and to submit a modified rule proposal for further study." The May report words it differently: the Committee "does not recommend action on the proposed Rule 707 at this time." Neither document says the rule was rejected, and it was not adopted.

The Reporter told the Standing Committee "this was the first time in his experience that a rule published for public comment did not go forward."

Committee members preferred a revised "Version 2," printed at pages 173-174 of the agenda book. It says: "Ordinarily, the proponent must provide an expert to explain how the system of artificial intelligence reliably produced the evidence. But in exceptional circumstances, the court may consider other proof." The Reporter cautions that the Committee has not approved all of that language. Ms. Shapiro said the DOJ "would rather have no Rule 707 at all."

What Happens Next for the Federal Rules

A federal evidence rule moves through several bodies. An advisory committee proposes it, it is published for comment, the Standing Committee and the Judicial Conference review it, and the Supreme Court adopts it "by order before May 1, to take effect no earlier than December 1 of the same year" unless Congress acts.

Judge Furman said in May that Rule 901(c) "would not go into effect until December 1, 2028" even if published in summer 2026. It was not published then. So a 2027 publication could not take effect before December 1, 2029 at the earliest. That date is our arithmetic from the process, not a court statement.

We make no forecast about whether either rule will be published. The dated events are in the 2027 section below.

The Liar's Dividend

The liar's dividend is the flip side of deepfakes: because fakes exist, real evidence can be doubted. The Iowa Supreme Court defined it this way in October 2026: "The term 'liar's dividend' refers to the possibility that public awareness of deepfakes will make it easier to raise doubts about legitimate evidence," citing an article by Bobby Chesney and Danielle Citron.

The draft Committee Note uses the term for a related risk, that "juries might be led to think that no evidence can be trusted." It says draft Rule 901(c) does not directly address that. It points instead to existing tools: Rule 403, and a court's authority over lawyer argument "made without foundation in the evidence."

In its Iowa case, the court wrote that "apart from trying to cash in on the so-called 'liar's dividend,' Amyda's counsel had little to work with."

What Courts Have Done: Matter of M.S. and State v. Amyda

Two state high courts ruled on these arguments in 2026. They applied state rules, not the federal ones, and they came out differently because their records differed.

New York: Matter of M.S.

A New York family court found child abuse based entirely on three videos. They came from a man who told an FBI agent he had hacked security cameras, and he never testified. The Court of Appeals held: "Because the foundation offered for the videos was insufficient, we hold that Family Court erred in admitting the videos into evidence and therefore reverse."

This is an authentication ruling, not a deepfake finding. The majority wrote that "The failure to authenticate evidence sufficiently does not mean the evidence was false." A dissent noted that "no one uttered the word 'deepfake' during the Family Court proceedings."

The deepfake point came in the reasoning. The majority said testimony that a video matched a real store's layout had been held insufficient in an older New York case, a reading one dissent disputed, and it added that "the increasing prevalence of 'deepfake' videos has only rendered the method of matching circumstantial details in a video to personal observations a more suspect form of authentication." The other dissent called the draft federal rule an "eminently logical proposal."

Iowa: State v. Amyda

In a sexual abuse trial, the State offered a video with no chain-of-custody evidence, and the defendant claimed it could be a deepfake. The victim testified that she recognized the room, her clothing and the defendant's tattooed hand. The Iowa Supreme Court upheld admission under Iowa Rules of Evidence 5.901 and 5.1002, writing: "There is no special codicil to the rules of evidence governing videos."

On the deepfake claim, the court said: "Amyda didn't raise a genuine question, only conjecture." It set the New York case aside on the facts: "here, by contrast, there are no identifiable circumstances to suggest that the video might be a deepfake." It also said Iowa's rule on duplicates, and federal Rule 1003, "already work this way" in a sense, like the draft's first step.

Other cases, as the agenda book describes them

We rely here on the Committee's agenda book, not on the orders themselves.

• Mendones v. Cushman & Wakefield (California Superior Court, Sept. 9, 2025). An article by a federal magistrate judge, quoted in the agenda book, says a California judge imposed terminating sanctions after finding that plaintiffs "intentionally submitted fabricated evidence," with certain exhibits found to be "deepfakes or materially altered."
• Washington v. Puloka (Washington trial court, 2024). A passage quoted in the agenda book says a court held AI-enhanced video and the related expert testimony inadmissible under Rule 702 because the tool was "largely untested and potentially unreliable." This case was about enhancement, not forgery.
• United States v. Belt, 181 F.4th 931 (8th Cir. 2026). The Reporter describes a ruling that unexplained gaps in surveillance video "went to weight and not admissibility." The Reporter says the case was not about deepfakes, though a judge dissenting from denial of en banc review urged courts to be more stringent on foundation.

State Rules and Bills on AI Evidence

States are not bound by a federal draft, and we did not survey every state's court rules, so this list is not complete. Judge Furman noted in May that issues "may start to appear more in state courts (in family court cases, for example)" and that the federal rules "are a model for the state courts."

• Louisiana Act 250 (HB 178, 2025), in force. Effective August 1, 2025, it amends the Code of Civil Procedure, not the Code of Evidence. Under art. 371(C), an attorney "shall exercise reasonable diligence to verify the authenticity of evidence before offering it to the court." Under art. 1551, a party with "reasonable suspicion" that the other side's exhibits are falsified, including by AI, "shall raise these concerns at the pretrial conference" or a pretrial admissibility hearing, where the court has ordered one. See our Louisiana deepfake laws page.
• Connecticut Practice Book sec. 4-9, in force. Connecticut's Superior Court judges adopted this rule on June 11, 2026, and it took effect on publication in the Connecticut Law Journal on June 23, 2026. Anyone who uses generative AI in creating or editing a court filing "shall independently verify all citations, legal authorities or evidence produced by generative AI," and failure may bring sanctions, including a nonsuit or default judgment. The rule's stated concern includes "inaccurate or fabricated evidence," and a comment printed in the October 2026 federal agenda book, at pages 135-136, says it "applies to presentation of evidence."
• California SB 11, not law. It would have directed the Judicial Council on AI evidence. Governor's veto October 13, 2025; veto sustained March 2, 2026.
• New York S9390, pending as of November 5, 2026. It was referred to the Senate Judiciary Committee on March 6, 2026. It would bar an inquiry on a bare AI claim unless the challenger shows enough "to support a reasonable inference" that the evidence may not be authentic. The offering party would then have to establish authenticity by a preponderance.
• New York A1338, pending as of November 5, 2026. It was referred to the Assembly Codes Committee on January 7, 2026. It would limit AI-created evidence to items substantially supported by independent admissible evidence, with the proponent establishing reliability and accuracy of the specific AI use.

AI Victim-Impact Video: State v. Horcasitas (Pelkey)

The Pelkey AI video no longer stands as a sentencing that simply allowed it. In State v. Horcasitas, No. 1 CA-CR 25-0191, a published opinion filed September 30, 2026, the Arizona Court of Appeals affirmed the manslaughter conviction and vacated the sentence, sending the case back for resentencing.

The court held: "while the embedded 'real' video footage of the victim was permissible, allowing the presentation of the AI video constituted fundamental error. And on this record, the judge's consideration of the AI video so prejudiced Horcasitas as to render the sentencing procedure fundamentally unfair."

Nobody objected at sentencing, which is why the court reviewed for fundamental error. The ruling concerns a sentencing, not evidence at trial, and it did not overturn the conviction. The court tied its finding of prejudice to "this record," and this page does not say how its reasoning would apply to a different AI victim statement.

As of November 5, 2026, we have not confirmed whether anyone has asked the Arizona Supreme Court to review the decision. More on Arizona law is on our Arizona deepfake laws page.

AI Chat Logs as Evidence

Chat logs are items of evidence like any other, so Rule 901(a) applies to them. The Reporter's memo says draft Rule 901(c) "covers any item of evidence, including text."

A separate question is whether a chat is protected from the other side. Law-firm summaries report that in February 2026 a federal judge in the Southern District of New York held that documents a criminal defendant created with a consumer AI chat tool were neither privileged nor attorney work product.

Published reports also describe a May 2025 order in the New York Times copyright suit against OpenAI requiring preservation of chat output logs, and an October 2025 order ending that duty going forward, with exceptions. Both descriptions come from those reports, not the orders themselves, and each ruling turned on its own case.

If a Deepfake Claim Comes Up in Your Case

This section describes what the rules provide. It is not advice on what to do in a particular case, and it covers federal court unless it says otherwise.

When an objection is raised. Federal Rule 103(a) lets a party claim error in admitting evidence if the party "timely objects or moves to strike" and "states the specific ground," and the error affects a substantial right. In Iowa, the defendant raised the issue before trial in a motion in limine and again when the video was offered.

Who decides. For authenticity under Rule 901(a), the lower Rule 104(b) standard applies: the judge asks whether a jury could reasonably find the item authentic, and the jury decides. The draft May minutes record a Committee member's point that draft Rule 901(c) would make the trial judge "the gatekeeper" once a challenger clears the first step. Rule 104(e) preserves a party's right to offer the jury evidence on weight or credibility, so a video that gets in can still be attacked at trial.

What no rule requires today. No Federal Rule of Evidence sets a pretrial notice duty or a threshold showing specific to a deepfake claim. Those are features of the draft. State practice differs: Louisiana's rule above is one example, and New York's court looked at the whole foundation in M.S., where a dissent noted that no one raised deepfakes in the family court and that the theory first came up on appeal.

If fabricated evidence is found. Courts have tools beyond exclusion. The terminating sanctions described in Mendones are one example. Louisiana's art. 371 treats an attorney's offering of evidence the attorney knew or should have known was false, without disclosure, as a violation, and the article provides for contempt. In the Committee's May discussion, two members said presenting deepfake evidence is a crime; that passage of the minutes names no statute.

What Changes in Federal Evidence Law in 2027

As of November 5, 2026, no deepfake evidence rule is scheduled to take effect in 2027. These are the dated events we found:

• December 29, 2026: registration deadline for the January Standing Committee meeting. Its agenda book had not been posted as of November 5. Last January's book carried the Evidence Committee's report and the draft minutes of its fall meeting, so this one may be the first official record of October 15 and would carry any request to publish.
• January 5, 2027: Standing Committee meeting, Washington, D.C.
• February 15, 2027: comments close on Rules 104 and 902(1). These are not AI rules.
• April 14, 2027: Advisory Committee on Evidence Rules, Washington, D.C. This is the Committee's next scheduled meeting.
• June 8-9, 2027: Standing Committee meeting, Dallas.
• November 3, 2027: Advisory Committee meeting, location to be set.

We have not confirmed any date for a ruling on any review of Horcasitas, or for resentencing. The New York bills have no scheduled vote that we found.

The dates come from the Rules Committees' meetings calendar, read November 5, 2026.

Deepfake Evidence Is Separate From Deepfake Crimes

Whether a video is admissible is a different question from whether making or sharing it is illegal. Federal law separately makes it a crime to knowingly publish certain intimate images of identifiable people, including AI forgeries indistinguishable from real images, when the elements of the TAKE IT DOWN Act, 47 U.S.C. 223(h), are met, and requires covered platforms to remove reported depictions within 48 hours of a valid request (47 U.S.C. 223a). Pending bills are on our federal deepfake bills tracker.

Where to Get Help

A court's self-help center can explain local procedure in your case, and many state and federal courts run one. Your state or local bar association's lawyer referral service can connect you with a lawyer who handles evidence disputes, and a local legal aid office may help if you qualify for its income rules.

If a fake image or video of you is the problem, rather than a dispute over evidence, start with the deepfake laws hub or the TAKE IT DOWN Act page.

This page is general legal information about US law as of November 5, 2026, not legal advice. A lawyer licensed in your state can advise on your situation.

Related Pages

• Deepfake laws by state
• Federal deepfake bills tracker
• TAKE IT DOWN Act
• AI-generated sexual images laws
• Deepfake fraud and impersonation laws
• New York deepfake laws
• California deepfake laws
• Iowa deepfake laws
• AI laws

Frequently Asked Questions

Is there a federal rule of evidence for deepfakes?

No. As of November 5, 2026, no Federal Rule of Evidence addresses deepfakes or AI-generated evidence by name. Draft Rule 901(c) has never been published for comment, and proposed Rule 707 was withdrawn for further study.

Can AI-generated evidence be used in court?

No federal evidence rule addresses it by name, so ordinary rules decide. In federal court the party offering an item must produce evidence sufficient to support a finding that it is what that party claims. One Washington trial court excluded AI-enhanced video, according to a passage quoted in the Committee's agenda book.

What is Rule 901(c)?

It is a working draft, not a rule. A party claiming AI fabrication would have to present evidence sufficient to support a finding of fabrication. If that party did, the item would be admissible only if the proponent showed it is more likely than not authentic. It has never been published for comment.

What is Rule 707?

Proposed Rule 707 would have applied the Rule 702 reliability standards to machine-generated evidence offered without an expert witness, where Rule 702 would apply if a witness testified to it. It concerned reliability, not deepfakes. It was published in August 2025 and withdrawn for further study, and a revised version is under study as of November 5, 2026.

Was Rule 707 rejected or adopted?

Neither word fits. The Committee's October 2026 agenda book says it voted to withdraw the published proposal and submit a modified version for further study. Its May 2026 report says it does not recommend action at this time.

Can someone just claim a real video is a deepfake?

A party can raise the argument, but it may not go far without support. In State v. Amyda, the Iowa Supreme Court held on October 9, 2026 that a bare deepfake suggestion raised no genuine question about the video. The draft federal note says a broad claim of 'deepfake' is not enough, but that draft is not law.

What is the liar's dividend?

The Iowa Supreme Court describes it as the possibility that public awareness of deepfakes will make it easier to raise doubts about legitimate evidence. The court cited an article by Bobby Chesney and Danielle Citron.

Did a court rule that videos were deepfakes in Matter of M.S.?

No. New York's Court of Appeals held on February 17, 2026 that the foundation for the videos was insufficient. It said a failure to authenticate does not mean the evidence was false, and a dissent noted that nobody raised deepfakes in the family court.

What happened to the Pelkey AI video?

On September 30, 2026, the Arizona Court of Appeals held that allowing the AI video at sentencing was fundamental error and that, on that record, the judge's consideration of it made the sentencing fundamentally unfair. It vacated the sentence for resentencing and affirmed the conviction. The ruling is about sentencing, not evidence at trial.

Do state courts have to follow the federal evidence rules on deepfakes?

No. State courts apply their own evidence rules, and an unpublished federal draft binds no court. The Committee's chair has called the federal rules a model for the state courts, and Iowa's high court says its rules were modeled on them.

When could a federal deepfake evidence rule take effect?

Not before December 1, 2029 at the earliest if published in 2027, by our arithmetic from the rulemaking process. The Committee's chair said in May 2026 that even a summer 2026 publication could not take effect before December 1, 2028, and no publication happened.