By Bill Hartzer · Published · How this site is researched
Tribunals have treated the two directions differently. Where evidence was shown to be fabricated, as in Mendones v. Cushman & Wakefield and Adams v. Anderson, the consequences were dismissal with prejudice and a contempt sentence. Where a party only asserted that genuine evidence might be fake, as in State v. Amyda and Burnley v. Valentin, tribunals asked for an evidentiary predicate and, finding none, admitted the evidence. A proponent who cannot authenticate a recording can still lose it.
Deepfake claims in the record, by claim and response
A deepfake is synthetic audio, video or imagery made to look like a real person or event. The matters below fall into three groups: material offered as genuine that proved fabricated, genuine material attacked as fabricated, and AI-processed or generated media offered openly. Several rest on reporting rather than a located order, and each linked entry says which.
| Matter | Tribunal and date | The claim | What the tribunal did |
|---|---|---|---|
| Mendones v. Cushman & Wakefield | Alameda County Superior Court, Sept. 9, 2025 | Self-represented plaintiffs filed video "testimony" and images the court suspected were generated or altered | Found two videos AI-generated; dismissed the case with prejudice as a terminating sanction |
| Adams v. Anderson | Kentucky Court of Appeals, May 29, 2026 | A father alleged a prescription-bottle photo was altered or AI-generated | Mother admitted lying and served 45 days for contempt; affirmed in an unpublished opinion with no finding that AI was used |
| Maryland v. Darien (Pikesville High) | Baltimore County Circuit Court, Apr. 28, 2025 | Audio presented as a principal's racist remarks | After forensic analysis found it synthetic, the defendant entered an Alford plea; 4 months in jail |
| Florida v. Schinault | Pinellas County court, Dec. 1, 2025 | A complainant gave police a generated image of an intruder | No contest plea to false reporting; probation and a fine |
| State v. Amyda | Iowa Court of Appeals, Jan. 28, 2026; Iowa Supreme Court, Oct. 9, 2026 | The defendant argued a video of the charged abuse may have been a deepfake | Admission affirmed twice; no identifiable circumstances suggesting a deepfake |
| Matter of M.S. | New York Court of Appeals, Feb. 17, 2026 | Videos recovered from a hacker's computer offered to prove abuse | Held not properly authenticated, 4-3; petitions dismissed; no finding of fabrication |
| Burnley v. Valentin | E.D. Va., Mar. 18, 2026 | A party said a recorded call was an AI clone of his voice | After a hearing, recording found authentic; settlement enforced |
| Huang v. Tesla | Santa Clara County Superior Court, Apr. 27, 2023 (tentative) | Tesla declined to admit recordings of Elon Musk, citing deepfakes of public figures | Limited 3-hour deposition of Mr. Musk ordered; case later settled |
| State v. Puloka | King County Superior Court, Mar. 29, 2024 | The defense offered a bystander video enhanced with Topaz Labs software | Enhanced version excluded after a Frye hearing; trial on the original footage |
| State v. Horcasitas (Pelkey) | Arizona Court of Appeals, Sept. 30, 2026 | A generated likeness of a homicide victim delivered a statement at sentencing | Sentence vacated; the video held not reliable and its use fundamental error |
Exhibits that proved to be fabricated, and how it was shown
In Mendones, Judge Victoria Kolakowski compared two video exhibits with an authentic video of the same witness already in the record and found the voice, cadence, gestures and expression markedly different, concluding that the exhibits were products of generative AI rather than recordings of the witness. It also noted a figure apparently inserted into a doorbell-camera photograph, altered message screenshots and metadata anomalies, and observed that the device the plaintiff named could not run the features she described. Lesser sanctions were considered and found insufficient. The California Supreme Court denied review in the writ proceeding in January 2026.
Adams v. Anderson was resolved by records. The father showed that the urgent-care center and the pharmacy had no records for the date the photograph implied. The mother then conceded the testimony and documents were false or altered, pleaded guilty to criminal contempt and, as the opinion records, "admitted to lying to the court." The Kentucky Court of Appeals affirmed in an opinion designated not to be published. It made no finding about how the documents were made.
Two criminal matters outside the courtroom show the same method. In the Pikesville High case, police relied on named forensic analysts and the FBI before announcing that the principal's supposed remarks were not authentic, more than three months after the clip spread. In St. Petersburg, a detective recognized a complainant's suspect image from an online trend, and investigators found it in a deleted folder on her own device, created days before the reported crime. Both accounts rest largely on news reporting and police records rather than a court opinion.
Genuine evidence called fake: the predicate tribunals required
Authentication is the threshold showing that an item is what its proponent claims. The liar's dividend is the advantage a party gains when public awareness of deepfakes makes genuine evidence easier to doubt. The Iowa Supreme Court used the term in Amyda, observing that "apart from trying to cash in on the so-called 'liar's dividend,' Amyda's counsel had little to work with."
The Court of Appeals had described the claim as "speculative claims that the video was a 'deepfake' without any evidentiary" support. On further review, the Supreme Court affirmed unanimously in an opinion by Justice Mansfield. It acknowledged that "We are not oblivious to the risks of video evidence," but concluded that "here, by contrast, there are no identifiable circumstances to suggest that the video might be a deepfake," and found that the best-evidence objection raised only conjecture. Review was for abuse of discretion, and the decision is an affirmance on that record.
Burnley shows what a predicate looked like when the proponent supplied one. Magistrate Judge Mark R. Colombell heard the recording at an evidentiary hearing, received two sworn declarations on how it was recorded and preserved, and accepted identification by a witness familiar with the voice under Federal Rule of Evidence 901(b)(5). The challenger offered no expert and declined to testify under oath that the voice was not his. In Huang, Judge Evette D. Pennypacker called Tesla's position troubling and sent the question to the one person who could answer it, ordering a limited deposition.
Matter of M.S. runs the other way, and it is the necessary counterweight. New York's highest court held, 4-3, that videos recovered from a hacker's computer were not authenticated by matching details to the family's home, a method the majority said the prevalence of deepfakes had made more suspect. It made no finding that the videos were fake. The dissents warned of new hurdles for video evidence. The Iowa Supreme Court contrasted M.S. directly before concluding that nothing identifiable in Amyda suggested a deepfake.
Enhanced and generated media offered openly
Enhancement is different in kind from fabrication. In Puloka, the defense ran a short, blurry bystander video through Topaz Labs software. After a Frye hearing (an inquiry into general acceptance in the relevant scientific community), the King County court treated forensic video analysts as the relevant community and excluded the enhanced clip, leaving the original available to both sides. The order is unpublished, and the account rests on press and practitioner reporting.
Generated media offered openly has also met resistance. In State v. Horcasitas, a victim's family presented a generated likeness of Christopher Pelkey reading a statement they had written. The Arizona Court of Appeals vacated the sentence, holding that the video "does not reflect actual events" and that "No disclaimer could cure that error." No mandate had issued as of October 8, 2026. In the Dewald appeal, a presiding justice stopped an undisclosed avatar presenting oral argument. And in a French appeal, the Cour d'appel de Rouen found causation in professional evidence and said expressly that a party's Microsoft Copilot analysis was not among it.
Proposed Rule 707, draft Rule 901(c) and the October 2026 meeting
In US federal courts, Federal Rule of Evidence 702 sets reliability requirements for expert testimony. Proposed Rule 707 would apply those requirements to machine-generated evidence offered without a sponsoring witness. It was published for comment from August 2025 to February 2026. In May 2026 the Advisory Committee on Evidence Rules declined to recommend it, revised the draft to refer to artificial intelligence, deleted the exception for simple scientific instruments and added a notice provision. The same report stated that a Rule 901 amendment on deepfakes was, at least for now, not warranted.
The agenda book for the October 15, 2026 meeting records that the committee had voted to withdraw the published rule and submit a modified proposal for study, and that it "has prepared, but not finally approved, a new Rule 901(c)." As drafted, an opponent would first offer evidence sufficient to support a finding that an item is a deepfake; the proponent would then have to show the item is more likely than not authentic. The Reporter's memorandum discusses Adams and quotes an Eighth Circuit dissent describing Mendones as among the first times a trial court confronted deepfake video evidence.
At the Boston meeting, the committee heard expert panels, including speakers from OpenAI and Meta. According to press reports, the chair, Judge Jesse M. Furman, appointed a subcommittee on deepfakes. No vote to publish either rule was reported, and the May draft minutes record the Justice Department's view that Rule 901(c) was not ready. As of the record's October 21, 2026 check, neither rule had been approved, and no federal evidence rule specific to AI output or deepfakes was in effect. Judicial guidance elsewhere, such as the England and Wales guidance, tells judges to be aware of the challenges deepfake technology poses; it is guidance, not a rule of court.
What an examiner asked to authenticate should do first
The resolved matters share a pattern: none turned on a detection tool's verdict. Each was settled by ordinary provenance, the record of where an item came from and what should corroborate it. An examiner asked to authenticate, or to challenge, a recording or image can start in the same place.
- Device file history. Creation dates, file locations and deletion records on the source device answered the question in St. Petersburg.
- Records that should corroborate. In Adams, the provider and pharmacy records that should have existed did not.
- Authenticated exemplars and metadata. Mendones compared the exhibits with a known video of the same witness and read the metadata.
- Custody from the source. Burnley rested on declarations about recording and preservation; M.S. failed where no one could speak to the recording's creation.
- The original beside any enhancement. Puloka kept the unprocessed footage before the jury.
Clause 4, checkability before adoption, asks the examiner to name, before adopting any output, the specific thing it would be checked against. A detector's confidence score with nothing behind it fails that test, because the opposing party will ask how it was produced. Clause 5, reproducible analysis, requires a method another examiner could repeat from the file and its provenance. Clause 3, independent authority, and Clause 7, contemporaneous record, complete the picture: confirm each step against a source outside the item, and record it as it is taken. Kohls v. Ellison is the cautionary footnote. A declaration defending a state deepfake statute was excluded after GPT-4o supplied citations to two articles that do not exist.
Matters this guide relies on
- Adams v. Anderson (Kentucky Court of Appeals) Primary document read
- Proposed Federal Rule of Evidence 707 Primary document read
- Burnley v. Valentin Primary document read
- Matter of M.S. (M.H.) Primary document read
- Cour d'appel de Rouen, RG 25/01677 (Copilot analysis) Primary document read
- State v. Amyda (deepfake claim without evidence) Primary document read
- Florida v. Schinault (ChatGPT image false report) Primary document read
- AI Judicial Guidance, England and Wales (October 2025) Primary document read
- Mendones v. Cushman & Wakefield Primary document read
- State v. Horcasitas (Pelkey AI victim impact statement) Primary document read
- Maryland v. Darien (Pikesville High deepfake audio) Reported; document not located
- Dewald AI avatar argument (N.Y. App. Div., 1st Dept.) Primary document read
- Kohls v. Ellison Primary document read
- State of Washington v. Puloka Reported; document not located
- Huang v. Tesla Primary document read
Questions
What is the liar's dividend?
It is the advantage a party gains when public awareness of deepfakes makes it easier to cast doubt on genuine evidence. The Iowa Supreme Court defined it in a footnote in State v. Amyda and observed that defense counsel, apart from trying to cash in on it, had little to work with. The video in that case was held properly authenticated.
Does claiming a video is a deepfake shift the burden to the party offering it?
In the matters recorded here, a bare claim did not. Amyda and Burnley required something more than the claim, and the recordings were accepted. The draft federal Rule 901(c) would formalize a two-step approach, with the opponent first showing enough to support a deepfake finding, but it has not been approved. Matter of M.S. shows the proponent still carries the ordinary authentication burden.
Is there a federal rule of evidence on deepfakes?
Not as of October 21, 2026. The Advisory Committee withdrew published Rule 707 for further study, prepared but did not approve a Rule 901(c), and at its October 15 meeting heard expert panels; press reports describe a subcommittee being appointed. No vote to publish either rule was reported.
Can an AI detection tool settle whether a recording is genuine?
None of the resolved matters in this record turned on a detector's verdict. They were settled by comparison with authenticated exemplars, metadata, device file history, custody declarations and records that should have corroborated the item. Under the Standard, a tool's score is adoptable only if the examiner can name what it would be checked against.
About this guide
This guide is drawn from the entries it links, each of which lists its own sources and says whether its primary document was read. It describes what tribunals did and does not state the law of any jurisdiction. Last checked October 21, 2026. If a matter here has moved, or one is missing, write to [email protected].