The Expert RecordThe standard for expert witnesses using AI

AI in the Courtroom

Fabricated citations and facts in filings

Filings that put a nonexistent case, quotation or fact before a court, and what the court did about it.

25 mattersLast reviewed 2026-10-21

By · Published · Updated · How this site is researched

Filings that put nonexistent material before a court

This part of the record collects matters in which a filing put something before a tribunal that did not exist: a case, a quotation, a statutory passage or a fact from the trial record. In each matter recorded here, the fabricated material came from, or was attributed to, a generative AI system (software that produces text in response to a prompt). The most common form is the hallucinated citation, meaning a reference to a case that does not exist. Fabricated quotations attributed to real cases are close behind. The most serious form recorded so far is fabricated fact.

The volume of reported matters is large and still growing. The AI Hallucination Cases database maintained by the researcher Damien Charlotin listed 2,038 decisions when this page was checked on September 11, 2026, with its most recent entry dated September 9, 2026. That database counts decisions in which a court or tribunal addressed AI-generated false content in more than passing terms. It is one researcher's compilation rather than an official statistic, and it counts decisions rather than incidents, so the figure should be read as an indication of scale only. The database is useful for finding matters; each entry in this record has been checked against the order or against reporting.

Research database, drafting aid and record summary

The matters fall into a small number of patterns. The first is the chat system used as a research database. In Mata v. Avianca (2023) and Johnson v. Dunn (2025), a lawyer asked a chat system for supporting case law and inserted what came back. In both, the cases were not pulled from a reporter before filing.

The second is the rewriting pass. In Noland v. Land of the Free, counsel wrote drafts and then ran them through several AI platforms to "enhance" them. The California Court of Appeal found that 21 of 23 case quotations in the opening brief were fabricated and recorded that counsel did not read the enhanced briefs before filing.

The third is the summary of the record. In the Sandoval murder appeal in New Mexico, appellate counsel loaded a trial transcript into ChatGPT and relied on its summary. The resulting brief described witnesses and testimony that did not exist. This moves the problem from legal authority to the facts an appellate court reviews.

Other patterns recur. Research supplied by a client was adopted without checking in Al-Haroun v Qatar National Bank. A sampling check, where the first citations were verified and the rest assumed, was described in the Victorian murder case involving Rishi Nathwani KC. An emergency motion was drafted under time pressure in In re Prince Global Holdings. Repeated filings by a self-represented party after a warning led to sanctions in Obi v. Cook County. Prosecutors are not exempt: the Nevada County, California district attorney's office acknowledged AI-related errors in 4 felony cases and faces a sanctions inquiry.

Matters in this topic

25 matters, most recent first within each group. 25 were checked against the primary document, and the others rest on reporting and say so.

Civil litigation

  • October 7, 2026 · Florida (state)Kid International v. City National Bank of FloridaPrimary document readThe plaintiffs' initial brief cited a nonexistent case seven times, attributed fictional quotations to real decisions and misstated holdings; counsel admitted his firm used generative AI and relied on client-supplied research. The appellate court affirmed summary judgment and granted the defendants' motion for appellate attorney's fees as a sanction, with plaintiffs and counsel jointly and severally liable.
  • September 25, 2026 · U.S. federalAguilar v. The Crawford Group, Inc.Primary document readA California lawyer admitted pro hac vice in Massachusetts filed oppositions with fictitious and misattributed citations and said he believed the enterprise version of his AI software did not hallucinate. The court granted Rule 11 sanctions, ordered his firm to pay up to $10,000 in fees and revoked his pro hac vice admission.
  • September 24, 2026 · California (state)Cervantes v. Bianco (Supreme Court of California)Primary document readIn an original writ proceeding over ballots seized by the Riverside County Sheriff, the sheriff's preliminary briefing misquoted and misstated one of the court's own decisions. The firm apologized and withdrew the argument. The Supreme Court of California disapproved the improper use of AI tools and referred the matter to the State Bar.
  • September 22, 2026 · U.S. federalZhejiang Dingli Machinery Co. v. United StatesPrimary document readA Court of International Trade judge found that a brief's statements that the Commerce Department took part in a Canadian antidumping investigation cited no evidence and appeared to be the product of hallucination by a person or by AI. He ordered the plaintiff to substantiate them or show cause under Rule 11(b)(3), with declarations on who or what wrote the passages; no ruling on the order has been located.
  • September 14, 2026 · U.S. federalHenry v. Long Island UniversityPrimary document readAn employment plaintiff's opposition brief cited cases that do not exist and quotations not found in real decisions; counsel called the problems typographical and never filed promised corrections. Because the client had discharged counsel, the court imposed no fine but required two attorneys to file the opinion in every Eastern District of New York case they appear in for two years.
  • September 3, 2026 · District of Columbia (local)Douglas v. Deutsche Bank National Trust Co.Primary document readDeutsche Bank's brief as appellee cited four authorities that do not exist; the attorney who drafted it said she had used Google's generative AI search tool. The District of Columbia Court of Appeals struck the brief, rejected a proposal to redact the bad citations, and referred the matter to disciplinary counsel.
  • July 16, 2026 · U.S. federalDaghra v. Hinkley (government counsel)Primary document readThe Government's response to a show-cause order cited a Sixth Circuit case that does not exist. The court found the citation was likely produced by generative AI, declined to impose sanctions, and directed that future filings not include non-existent case law.
  • July 14, 2026 · U.S. federalIn re Rosslyn2016 (a legal research tool is still a tool)Primary document readCounsel's motions contained fabricated quotations and invented authority. He denied using ChatGPT and attributed the citations to a legal vendor's AI research feature. The court rejected the distinction, found contempt, ordered CLE and awarded $29,877 in fees.
  • June 3, 2026 · U.S. federalLnu v. Blanche (Ninth Circuit suspension)Primary document readTwo attorneys filed briefs containing fabricated cases and misattributed quotations, then characterized the errors as typographical and denied AI use at argument. The Ninth Circuit fined and suspended both and imposed a two-year certification requirement.
  • March 13, 2026 · U.S. federalWhiting v. City of AthensPrimary document readThe Sixth Circuit found more than two dozen fake citations and misrepresentations in the plaintiff's briefs across three consolidated appeals. After counsel refused to answer its show-cause questions, including whether they used generative AI, the court imposed full appellate fees, double costs, punitive fines and a disciplinary referral; it made no finding on AI use.
  • February 18, 2026 · U.S. federalFletcher v. Experian (Fifth Circuit sanction)Primary document readThe Fifth Circuit determined that counsel used artificial intelligence to draft a substantial portion of a reply brief without verifying it, and that she was not forthcoming in response to a show-cause order. It imposed a $2,500 sanction.
  • February 9, 2026 · United States (federal)Clinco v. CommissionerPrimary document readAn attorney's post-trial brief cited four cases in support of an argument; the court found three of them to be hallucinations generated by a large language model. It described the filing as a Rule 11(b) violation and imposed no sanction.
  • September 12, 2025 · California (state)Noland v. Land of the Free, L.P.Primary document readAn appellant's lawyer in a California employment appeal filed briefs whose case quotations were largely fabricated by AI tools. The Court of Appeal affirmed the judgment, sanctioned him $10,000 and referred him to the State Bar.
  • June 22, 2023 · U.S. federalMata v. Avianca, Inc.Primary document readPlaintiff's lawyers in a personal injury suit against Avianca filed a brief citing cases produced by ChatGPT that did not exist. Judge P. Kevin Castel imposed a $5,000 penalty and ordered letters to the client and to the judges named as authors of the fake opinions.

Criminal defense and prosecution

  • August 21, 2026 · New Mexico (state)State v. Sandoval (Aarons sanction)Primary document readAppellate counsel in a New Mexico murder appeal filed a brief prepared with ChatGPT that described nonexistent witnesses and testimony. The New Mexico Supreme Court held him in contempt, ordered $5,000, removed him and referred him to the Disciplinary Board.
  • January 14, 2026 · California (state)Kjoller v. Superior Court (Nevada County DA)Primary document readProsecutors in Nevada County, California filed briefs in felony cases that cited authority that does not exist. The California Supreme Court directed the Court of Appeal to issue an order to show cause on sanctions, and a judge was appointed in 2026 to investigate.

Large law firms

  • April 18, 2026 · U.S. federalIn re Prince Global Holdings (Sullivan & Cromwell)Primary document readSullivan & Cromwell filed an emergency motion in a Chapter 15 case that contained inaccurate citations and misquotations, including AI hallucinations. The firm reported the errors to Chief Judge Martin Glenn and filed a corrected motion. No court action has been reported.
  • July 23, 2025 · U.S. federalJohnson v. Dunn (Butler Snow)Primary document readLawyers at Butler Snow defending Alabama prison officials filed 2 motions containing 5 citations a partner obtained from ChatGPT that did not exist or did not say what was claimed. The court reprimanded 3 lawyers, disqualified them and referred them to the bar.

Self-represented litigants

  • April 9, 2026 · U.S. federalObi v. Cook CountyPrimary document readA self-represented plaintiff filed post-judgment briefs, which the court found were generated with AI, containing nonexistent cases and invented quotations after an earlier warning. Chief Judge Virginia M. Kendall struck the briefs and imposed a $5,000 Rule 11 sanction.
  • February 9, 2026 · U.S. federalAmarsingh v. Frontier Airlines (Tenth Circuit)Primary document readAn appellant appearing without counsel, but licensed to practice in Maryland, used ChatGPT to research and draft her appellate brief. Seven cited cases could not be located. The court affirmed, ordered $1,000 to the appellee, and referred her to Maryland.

Outside the United States

  • October 15, 2026 · IndiaDeepak Raheja v. Omkara Assets ReconstructionPrimary document readIn an insolvency appeal, the respondents objected that the appellants' compilation of judgments was replete with AI-generated nonexistent judgments, and the Advocate-on-Record blamed his clients. The Supreme Court of India proposed exemplary costs but deleted them after the first appellant died, ordering no costs.
  • September 3, 2026 · TanzaniaAssemblies of God Gospel Church v. Registrar of SocietiesPrimary document readAn advocate seeking an extension of time cited and quoted two Court of Appeal decisions that do not exist. The High Court of Tanzania found him guilty of professional misconduct, expunged the cases, suspended him for six months and ordered a published apology in English and Swahili.
  • August 25, 2026 · England and WalesSRA v Kumar (Solicitors Disciplinary Tribunal)Primary document readDefending disciplinary proceedings, a registered foreign lawyer filed an Answer and an email drafted with generative AI that cited nonexistent cases and quotations, and repeated the errors after being warned. The Solicitors Disciplinary Tribunal found the AI allegations proved and struck him off.
  • August 14, 2025 · Australia (Victoria)Victorian murder trial submissions (Nathwani KC)Primary document readDefense counsel in a Victorian murder case filed submissions containing fabricated quotations and nonexistent case citations generated by AI. Rishi Nathwani KC apologized. Justice James Elliott called the events unsatisfactory and ruled a day later than planned.
  • June 6, 2025 · England and WalesAyinde v Haringey and Al-Haroun v Qatar National BankPrimary document readIn 2 cases heard together, lawyers in England put fictitious case law before the courts. The Divisional Court declined contempt proceedings, referred a barrister and 2 solicitors to regulators and sent its judgment to the profession's governing bodies.

Sanctions, referrals and removal from the case

Responses vary by court, by the conduct and by how the lawyer responded once the problem surfaced. Monetary sanctions appear often: $5,000 in Mata, $10,000 in Noland, $5,000 against the self-represented plaintiff in Obi and $5,000 in the Sandoval appeal. Not every court has fined. In Johnson v. Dunn the court wrote that fines do not account for the harm and instead publicly reprimanded 3 lawyers, disqualified them from the case and referred them to the state bar. The New Mexico Supreme Court combined contempt, a payment, removal from the appeal and a disciplinary referral.

Referral to a licensing body is the most consistent response across jurisdictions. It appears in Noland, Johnson, Sandoval and both English cases heard in Ayinde and Al-Haroun. The English Divisional Court also set out its full range of powers, from public admonishment to referral to the police, and sent its judgment to the profession's governing bodies.

At the other end, some matters have produced no reported sanction. Sullivan & Cromwell self-reported the errors in its Prince Global motion, filed a corrected version and apologized; no court action had been reported when this page was checked. In the Victorian case the judge criticized the conduct on the record and no sanction was reported. Courts have also struck briefs, dismissed appeals and, in the Nevada County matter, ordered a formal show-cause proceeding with an appointed judge to take evidence.

Independent authority as the common thread

Almost every matter in this record engages Clause 3 of the Expert Record Standard, independent authority. Nothing enters the work on a system's authority. Every citation, quotation, figure and date is confirmed against an independent source before it is used. Clause 4, checkability before adoption, follows closely: the material in these matters could have been checked, and the check was not done.

The record also shows where other clauses come in. Summarizing trial testimony is a judgment about what the evidence showed, which engages Clause 2, non-delegation. Reconstructing after the fact which filings involved AI, as the Nevada County office had to do, is what Clause 7, the contemporaneous record, is meant to avoid. A prompt and complete account to the court, as in Prince Global, is the form of disclosure contemplated by Clause 8, though it does not substitute for verification.

The same obligations apply to an expert witness. A report that cites literature, standards or data carries the same risk as a brief that cites cases. The governing test is the practical check: a competent examiner in the same field, working only from the primary sources and with no AI system, should be able to reproduce every reference in the work.

Using AI has not itself been sanctioned

The record does not show that using AI in legal work is itself sanctionable. The court in Mata wrote that "there is nothing inherently improper about using a reliable artificial intelligence tool for assistance." The sanctions recorded here attach to filing unverified material and, in several matters, to how counsel responded when it was questioned.

The record does not show which tools are more or less reliable. Many orders do not identify the tool at all, and where one is named, the court usually records only what the lawyer said. The record also does not show how often fabricated material goes undetected. Every matter here was caught by a judge, a clerk or opposing counsel. Finally, the outcome of most disciplinary referrals is not yet public, so the professional consequences beyond the court's own order are often unknown.

Separating findings from explanations in the coverage

Start with the order if it can be found. Press coverage often compresses what the court found and what the lawyer said into one sentence. Separate the tribunal's findings from counsel's explanation, and note whether the court found bad faith, negligence or neither.

Identify what was fabricated. A nonexistent case, a false quotation from a real case and an invented fact from the record are different failures, and courts have treated them differently. Record the sanction exactly: to whom a payment is made, whether a referral was made or only threatened, and whether the lawyer was removed.

Check dates against the calendar. In the Sandoval matter, reports placed the court's action on different days, and at least one republication gave a weekday that does not match the date. Finally, record the posture. Many of these matters are reported at the sanctions stage, while the underlying case and any disciplinary proceeding continue.

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.

Questions about this topic

Is it sanctionable for a lawyer to use generative AI to draft a filing?

Not in itself, in the matters recorded here. The sanctions attach to filing citations, quotations or facts that the signer did not verify. The court in Mata v. Avianca wrote that there is nothing inherently improper about using a reliable AI tool for assistance. Local rules and standing orders may separately require disclosure of AI use.

What sanctions have courts imposed for fabricated citations?

The recorded responses range from criticism on the record to monetary penalties, public reprimand, disqualification, contempt and referral to licensing bodies. Payments of $5,000 to $10,000 appear in several matters in this record. The court in Johnson v. Dunn declined to fine and chose non-monetary sanctions instead.

Does self-representation excuse fabricated citations?

No, in the matters recorded here. In Obi v. Cook County the court stated that pro se status does not shelter a plaintiff from sanctions under Rule 11 and imposed $5,000 after an earlier warning. The court had extended leniency after an earlier set of fabrications in the same case.

Does a firm AI policy protect a lawyer from sanctions?

A policy has not protected the individual lawyer who failed to follow it. In Johnson v. Dunn the firm was released without sanction, but the lawyers who filed the unverified citations were reprimanded, disqualified and referred to the bar. In Prince Global, Sullivan & Cromwell told the court its policies were not followed.

How can a professional check AI-assisted work before filing?

By confirming every citation, quotation and factual statement against the primary source itself. That means opening each case in a reporter or official database, finding each quoted passage and checking each statement of fact against the transcript or exhibit. Agreement between two AI systems is not an independent check.

Top