By Bill Hartzer · Published · How this site is researched
Generated faces, voices and advocates
Most reported AI matters in litigation arise from written filings: a brief with a hallucinated citation (a reference to a case that does not exist), a declaration with an invented quotation. This part of the record covers something different. It collects matters in which a system appeared, spoke, or was consulted inside the proceeding itself. A generated face addressed an appellate panel. A generated voice gave a victim impact statement. Video exhibits of a witness turned out to be synthetic. A chatbot was to feed arguments to a defendant through an earpiece. A prospective juror asked ChatGPT about the case the juror had been called to decide. A trial transcript bore the marks of automated transcription.
The entries are grouped into 4 subtopics: AI avatars and generated video, victim and witness statements, AI in place of counsel, and courtroom tools. Each entry states what the participant did and what the tribunal did, with the primary document where one could be found.
Presenters, statements and machine assistance
Three patterns recur. The first is the generated presenter: software that renders a face and voice from a script. In the Dewald matter before the New York Appellate Division, First Department, a self-represented appellant used one to deliver his oral argument. In the Arizona Supreme Court, the court itself uses 2 named avatars to announce decisions. In the Pelkey sentencing in Maricopa County, a family used a generated likeness of the victim to speak words they had written for him. The technology is similar in all 3. What differs is who wrote the words, who approved them, and whether the audience was told.
The second pattern is substitution: a system offered in place of a human participant. DoNotPay's 2023 plan would have had a chatbot compose arguments in real time for a traffic defendant. The juror in State v. Boles in Paulding County, Georgia, substituted a chatbot's answer for the court's instruction to rely only on the evidence.
The third is the system embedded in the court's own processes, often without announcement. The Indiana Court of Appeals in Williams v. State inferred from the pattern of errors that AI had helped prepare a trial transcript. Fabricated evidence, as in Mendones v. Cushman & Wakefield, is the adversarial form of the same problem: a record that looks authentic and is not.
Matters in this topic
6 matters, most recent first within each group. 4 were checked against the primary document, and the others rest on reporting and say so.
AI avatars and generated video
- Dewald AI avatar argument (N.Y. App. Div., 1st Dept.)Primary document readA self-represented appellant presented his oral argument through an AI-generated avatar without telling the court it was synthetic. The presiding justice stopped the video and had him argue in person; the appeal was later decided on the merits.
- Arizona Supreme Court AI news reportersPrimary document readThe Arizona Supreme Court introduced 2 AI-generated video presenters to announce its rulings and court news. The justices write and approve each script; the software generates the presenter and voice.
Victim and witness statements
- State v. Horcasitas (Pelkey AI victim impact statement)Reported; document not locatedThe family of Christopher Pelkey played an AI-generated video of him delivering a victim impact statement at the sentencing of the man who shot him. Judge Todd Lang imposed 10.5 years and commented favorably on the video; defense counsel filed a notice of appeal.
AI in place of counsel
- DoNotPay "robot lawyer" and FTC orderPrimary document readDoNotPay planned to have a chatbot coach a traffic defendant in court through an earpiece in 2023 and abandoned the plan after warnings from state bars. In 2025 the FTC finalized an order requiring $193,000 and barring unsupported claims that its service could replace a lawyer.
Courtroom tools
- Williams v. State (Ind. Ct. App.) AI-assisted transcriptPrimary document readIn a July 2026 decision, the Indiana Court of Appeals noted that the trial transcript contained misattributions and typos suggesting generative AI had assisted its preparation. It affirmed the conviction and stated that users of such systems must proofread the result.
- State v. Boles juror ChatGPT research (Paulding County, Ga.)Reported; document not locatedDuring jury selection in a Paulding County, Georgia murder trial, a prospective juror asked ChatGPT about the case. Judge Dean Bucci released the jury pool, postponed the trial and moved against the juror for contempt.
Stopped videos, permitted statements and pending appeals
The responses in this part of the record have been immediate and practical more often than written. In Dewald, the presiding justice stopped the video and had the appellant argue in person; the later decision on the merits does not mention the avatar. In Pelkey, the judge received the video, commented on it from the bench, and imposed the maximum sentence; defense counsel filed a notice of appeal questioning the reliance placed on it. In Boles, the judge released the jury pool and postponed the trial. In Williams, the appellate court affirmed and added a footnote reminding users of such systems to proofread the output.
Written sanctions appear where the conduct was the submission of false material. The Alameda County court in Mendones compared the questioned videos with authentic footage of the witness, examined the metadata, imposed terminating sanctions and dismissed the action. The DoNotPay matter reached a regulator rather than a court: the Federal Trade Commission finalized a consent order in 2025 requiring $193,000 in monetary relief and barring unsupported claims that the service could replace a professional.
Disclosure as the recurring obligation
Disclosure, Clause 8 of the Expert Record Standard, runs through nearly every entry. The Arizona Supreme Court announced its avatars as AI-generated and named them. The Dewald avatar appeared under a permission granted for an ordinary video, and the presiding justice's objection was that she had not been told. Defense counsel in Pelkey objected that he had no advance notice. A tribunal can decide how to treat generated material only if it knows it is generated before it sees it.
Checkability, Clause 4, is the second thread. The Arizona videos can be checked against a written text approved by every justice. A dead man's generated voice cannot be checked against anything he said. A transcript can be checked against the recording, and in Williams the court's footnote describes what happens when it is not. Experts who work from transcripts, recordings or video should treat each as a document that may have passed through a system, and should confirm anything their opinion depends on against the original, as Clause 3 requires.
Non-delegation, Clause 2, applies wherever a system would speak for a participant. A professional who appears before a tribunal, whether as counsel or as a witness, does so personally. Generated delivery of that professional's words puts a system between the professional and the court.
Open questions on sentencing and courtroom tools
The record does not yet contain an appellate decision on whether an AI-generated victim impact statement may be received at sentencing. The Pelkey appeal was filed in May 2025, and no decision on it was located as of September 2026. The record contains no written order addressing the Dewald avatar, and no reported sanction against him. It does not show what the Boles juror asked or learned, and sources differ on whether a contempt finding was made. It does not establish which tool, if any, produced the errors in the Williams transcript; one commentator has argued that conventional speech-to-text software is a more likely source than generative AI.
The record also does not show how often generated material reaches courtrooms without being noticed. Each entry here is a matter in which someone noticed. The absence of reported incidents in a jurisdiction is not evidence that none occurred.
Verifying a viral courtroom clip before relying on it
Reports in this area travel quickly and are often built on a single video clip or a single article. Before relying on one, establish 4 things. First, identify the tribunal and the case number, and look for the written order; several matters here have none, and that absence should be stated. Second, identify who produced the material and who wrote its content, since a generated presenter reading a human-approved script raises different questions from a system composing words in real time. Third, determine whether the tribunal was told in advance. Fourth, separate what the tribunal did from what commentators have said about it.
Where sources conflict, as they do in Boles on the contempt step and in Williams on the decision date, the conflict belongs in the record rather than resolved by choosing the more dramatic version. Each entry in this section notes what could not be confirmed and the date it was last checked.
Primary documents behind this topic
The orders, opinions and published texts that entries in this topic were checked against. Read them before relying on any summary, including the ones here.
- Williams v. State (Ind. Ct. App.) AI-assisted transcript
Rashaan Williams v. State of Indiana, No. 25A-CR-3041 (Ind. Ct. App. July 23, 2026) (mem.) - Dewald AI avatar argument (N.Y. App. Div., 1st Dept.)
Dewald v Massachusetts Mut. Ins. Co., 2025 NY Slip Op 02267 (1st Dept Apr. 17, 2025) - Arizona Supreme Court AI news reporters
News Release: Arizona Supreme Court Introduces AI-Generated Court News Reporters (Mar. 11, 2025) - DoNotPay "robot lawyer" and FTC order
FTC Complaint, In the Matter of DoNotPay, Inc., Docket C-4812 (File No. 232-3042)
Questions about this topic
Has any court allowed an AI avatar to argue a case?
No matter in this record shows a court knowingly permitting an AI avatar to present argument. In the Dewald matter the New York Appellate Division had granted permission for a prerecorded video, and the presiding justice stopped it once the appellant confirmed the presenter was generated. The Arizona Supreme Court uses avatars only to announce its own decisions from scripts the justices approve.
Can an AI-generated victim impact statement be played at sentencing?
One Arizona trial court received such a statement in the Pelkey sentencing in May 2025, and the defense appealed. No appellate decision on the question was located as of September 2026. Defense counsel in that case agreed that victims have a right to be heard at sentencing, but the form of the statement and the notice given to the defense remain open questions in this record.
What happens when a party submits AI-generated video as evidence?
In Mendones v. Cushman & Wakefield the court found 2 witness videos were deepfakes, imposed terminating sanctions and dismissed the case. The court reached that finding by comparing the videos with authentic footage and examining their metadata. It declined to refer the matter to prosecutors and later denied reconsideration.
Does the Expert Record Standard apply to jurors, judges or court staff?
The Standard is written for professionals who produce expert work, so it does not bind jurors or court staff. The matters involving them are recorded because they show where untested system output can enter a proceeding. An expert who relies on a transcript, a court-produced summary or any document that may have passed through a system should confirm what the opinion depends on against the original source.
Should an expert disclose using AI to prepare a demonstrative or video?
Yes, where rules, orders or the engagement require it, and the matters here show why disclosure should come before the material is shown. The objections in Dewald and Pelkey were directed in part at the absence of advance notice. Disclosing the tool, the source text and the approval of the content lets the tribunal and the other side check the material before it is used.