State v. Coleman

The record

• CourtCourt of Appeals of Ohio, Eleventh Appellate District, Ashtabula County
• Citation2026-Ohio-965, reported at 280 N.E.3d 1042. Case No. 2024-A-0040; trial court No. 2022 CR 00416
• Decided20 March 2026, “Decision on Sanctions and Judgment Entry”. Judgment: “Sanctioning Attorney for Misconduct”
• SanctionedWilliam B. Norman, appellate counsel for the defendant, named in the caption as Respondent
• MovantThe State of Ohio, by the Ashtabula County Prosecutor
• ToolChatGPT, named: “The fabricated quotations were generated by ChatGPT, a publicly available generative artificial-intelligence platform” (¶ 24)
• What was falseNot case citations. Trial transcript quotations. “The statements were entirely fabricated. They were not paraphrases, approximations, or loose characterizations of actual statements; they were wholesale inventions that bore no resemblance to the record” (¶ 23)
• HowA paralegal uploaded case materials to ChatGPT and used its output in the application to reopen, filed 19 May 2025 with counsel’s affidavit swearing to its truth (¶¶ 2, 24). Counsel “did not independently verify the quotations attributed to the prosecutor against the trial transcript before filing” (¶ 28)
• After the flagWhen the State identified the fabrications and moved for sanctions, counsel “did not withdraw the application, amend the filing, or notify the court of the errors” (¶ 30), and then appealed the denial to the Supreme Court of Ohio without disclosing the infirmity (¶ 34)
• FindingConduct “contemptuous” and in bad faith sufficient to warrant sanctions under the court’s inherent authority (¶ 87). Fraud on the court, candor, competence, supervision of nonlawyers and misrepresentation rules all engaged (¶¶ 55 to 63)
• AuthorityThe court’s inherent authority, with R.C. 2705.02 contempt power for enforcement
• The dissentJudge Eklund concurred in the $2,000 sanction and the disciplinary referral and dissented from the six other measures as beyond the court’s authority (¶¶ 136, 152, 153)

The eight sanctions, from the judgment entry

• A$2,000, with credit for the $2,000 already paid to the prosecutor’s office under the parties’ settlement. The court held the private settlement did not displace its own power: “The court’s authority to levy a monetary sanction exists independent of, and is not extinguished by, the parties’ private agreement” (¶ 91). The amount is compensatory, measured against 32.1 documented hours of the prosecutor’s staff time and benchmarked against Wadsworth, Mata and HoosierVac
• BReferral to the Office of Disciplinary Counsel of the Supreme Court of Ohio, which the court records as mandatory under Prof.Cond.R. 8.3(a) (¶ 93)
• CService on every court for two years. A copy of the judgment to the presiding or administrative judge of every court in which counsel files a new appearance, within 14 days, with proof filed within seven (¶ 96)
• DA signed certification on every filing in any Ohio court for two years, that (a) every citation “has been independently verified by the attorney of record as being real, accurately quoted, and still valid”; (b) every quotation “attributed to any person, transcript, or document” has been verified against the original source; (c) every factual representation has evidentiary support; (d) any use of a generative AI tool is disclosed and all AI-generated content independently verified; and (e) the attorney “has personally reviewed the entire filing” (¶ 98)
• ESix hours of CLE within 150 days, in addition to the regular biennial requirement, on the ethical use of AI, AI-generated content, supervision of nonlawyer assistants using AI tools, and candor and the duty to verify (¶ 100)
• FThe application to reopen struck from the record, so that the fabricated content “is formally repudiated” (¶ 103)
• GRemoval as counsel for the defendant, with a plain-language written explanation to the client of the misconduct, his right to new counsel, and the public defender’s contact details, within 14 days (¶¶ 105, 110)
• HWritten apologies within 30 days to the trial prosecutor, the trial judge, and prior trial and appellate counsel, each acknowledging that the statements attributed to them “were false, were generated by an artificial intelligence tool, and were filed without verification” (¶ 114 and the judgment entry)

Failure to comply with any provision “may result in contempt proceedings pursuant to R.C. 2705.02”, with fines, imprisonment or both. Counsel must keep a record of compliance and produce it on request.

What the court found

The fabricated quotations were the substantive core of the application. They ascribed “specific, inflammatory statements to an identifiable individual”, the trial prosecutor, and by implication accused the trial judge of letting them pass and prior counsel of failing to object (¶ 25). “These were not victimless fabrications.” Had they been credited, they could have reopened a final appellate judgment in a murder case.

“The submission of fabricated transcript quotations to an appellate court is precisely such conduct. It is not merely negligent; it is an assault on the integrity of the appellate process, which depends entirely upon the fidelity of the record and the candor of counsel.” (¶ 54)

The court rejected the argument that the system “worked as intended” because the fabrications were caught: “Neither this court nor movant should serve as a vetting body for another’s lack of diligence” (¶ 76). On supervision, it held that a directive to staff not to use certain AI tools is not enough: “A policy that is neither monitored nor enforced is no policy at all”, and “The standard is meaningful review, not blind ratification” (¶ 62).

Three findings weighed on bad faith. The remedial AI policy counsel produced after the sanctions motion “bore the hallmarks of having itself been generated by an AI platform”, with unfilled placeholders such as “[Insert Date]” (¶ 31). Two months after the sanctions hearing, a motion counsel filed in another county contained a citation that could not be located and a ChatGPT prompt left in the text: “Would you like me to draft the next argument section…” (¶ 32). And counsel pursued an appeal to the Supreme Court of Ohio from the denial of the application without telling that court the application rested on fabricated quotations; the Supreme Court declined jurisdiction on 14 October 2025 (¶ 34).

On the certification requirement, the court wrote that it “imposes no burden beyond what every attorney should already be doing as a matter of professional obligation” (¶ 99), and described the certification and notification measures as a “direct and proportionate consequence of respondent’s demonstrated inability to ensure the accuracy of his filings without external accountability” (¶ 88).

Judge Eklund agreed with the fine and the referral and dissented from the rest, writing that the remaining measures “exceed the ‘authority’ we invoke to impose them” and that an appellate court should not adjudicate culpability under the Rules of Professional Conduct (¶¶ 137, 153). The majority answered the dissent at length (¶¶ 120 to 131). The scope of an appellate court’s inherent authority in Ohio is therefore contested inside the decision itself.

Why it matters

This is the most comprehensive AI sanctions order on this record, and the one that comes closest to writing down what a verification record would have to contain. Sanction D is a five-part certification: real, accurately quoted and still valid citations; quotations checked against the original source; factual representations with evidentiary support; AI use disclosed and AI content independently verified; and the attorney’s personal review of the whole filing. For two years, every filing this lawyer makes in Ohio carries that signed statement.

The decision also answers two questions that recur across the record. A private settlement with the other side does not end the court’s interest, because the court’s authority “is not extinguished by” the parties’ agreement. And a written AI policy is not supervision; the court looked for evidence that the policy was monitored and enforced, found an AI-generated policy with the brackets still in it, and drew its conclusion.

What was fabricated is the other reason this decision travels. Every earlier matter on this record involves case law. Here the tool invented testimony from a trial record, attributed it to a named prosecutor, and the invention went to court under an affidavit. The Missouri Court of Appeals quoted this decision on 8 September 2026 for exactly that point, that generative AI is prone to “generating false testimony from a legal record”.

Sources

• State v. Coleman, 2026-Ohio-965, Decision on Sanctions and Judgment Entry, 20 March 2026, 65 pages, read in full at the Supreme Court of Ohio’s reporter of opinions
• Supreme Court of Ohio, reporter of opinions, where Ohio appellate decisions are published

Related

• The full sanctions record
• State ex rel. Schnell v. Richardson, the Missouri Court of Appeals quoting this decision on 8 September 2026
• Wadsworth v. Walmart and Mata v. Avianca, the two matters the court benchmarked the amount against
• When the evidence is AI, the separate record of AI-made evidential material
• Verification is becoming a duty. Evidence of verification is the missing layer.

Verified 9 September 2026 against the court’s decision and judgment entry, read in full. Information, not legal advice.