U.S. courts imposed at least $145,000 in sanctions for AI-generated citation errors during the first quarter of 2026. That figure, compiled by ComplexDiscovery, spans four major sanctions orders in three months: a $10,000 penalty in the New York Appellate Division in January, $12,000 in the District of Kansas in February, $30,000 in the Sixth Circuit in March, and roughly $110,000 in the District of Oregon later that month. A Nebraska attorney was suspended from practice entirely. A Department of Justice lawyer was fired. And Damien Charlotin’s AI hallucination tracker, the most comprehensive database of these incidents, now catalogues more than 1,350 cases worldwide, with the pace accelerating: on March 31 alone, seventeen U.S. court decisions noted suspected AI hallucinations in filings.
When I wrote about Sullivan & Cromwell’s AI-contaminated bankruptcy filing last month, the central question was whether the profession’s policy-and-training approach to AI risk could deliver the “reasonable assurance” that Model Rule 5.1(a) requires. S&C had mandatory training modules, tracked completion, an Office Manual instructing lawyers to “trust nothing and verify everything,” and a policy requiring independent verification of all AI-generated citations before filing. The system failed anyway. The Q1 sanctions data raises a different question, and the answer complicates the story in ways the profession should find uncomfortable.
What the cases show
The four sanctioned matters follow a pattern that is hard to miss once you line them up.
In Deutsche Bank National Trust Co. v. LeTennier, 2026 NY Slip Op 00040 (3d Dep’t Jan. 8, 2026), attorney Joshua Douglass filed five papers in a mortgage foreclosure appeal containing no fewer than 23 fabricated cases. He “reluctantly conceded during oral argument that he used AI in the preparation of his papers” and told the court he had checked his work, though the filings demonstrated otherwise. The Appellate Division imposed $7,500 in sanctions against Douglass and $2,500 against his client, holding that lawyers bear a “nondelegable duty to verify legal authorities.” No evidence of any AI use policy or training appeared in the record.
In Lexos Media IP, LLC v. Overstock.com, Inc., No. 22-2324-JAR (D. Kan. Feb. 2, 2026), Senior Judge Julie Robinson sanctioned five attorneys a total of $12,000 for patent infringement briefs containing nonexistent cases, fabricated quotations, and citations to real cases that held the opposite of what the brief claimed. Attorney Sandeep Seth admitted he used ChatGPT to generate citations under time pressure and did not verify them. No pre-existing firm AI policy was in evidence. Judge Robinson’s order required attorney Christopher Joe to implement and certify enhanced firm AI policies by February 28, 2026—meaning the court had to order into existence the verification infrastructure that should have preceded the filing.
In Whiting v. City of Athens, Nos. 24-5918/5919, 25-5424 (6th Cir. Mar. 13, 2026), a published panel opinion by Judge John Bush imposed $15,000 fines on each of two Tennessee attorneys—Van Irion and Russ Egli—for appellate briefs containing more than two dozen fake or misrepresented citations across three consolidated appeals. The court issued a show-cause order asking whether generative AI had been used and how cite-checking was performed. Neither attorney responded substantively; one accused the court of participating in a “vast conspiracy.” The court did not resolve the AI question but stated that citations, “however generated,” must be “personally read and verified” before filing. Both attorneys had prior disciplinary records for lack of candor—Irion had been suspended from the Eastern District of Tennessee for five years in August 2025 for lying to the court in the underlying case.
And in Couvrette v. Wisnovsky, No. 1:21-cv-00157-CL (D. Or., orders of Dec. 12, 2025, and Mar. 23, 2026), Magistrate Judge Mark Clarke imposed the largest aggregate AI-related sanction in U.S. legal history: roughly $110,000 in combined fines and attorney fee awards against San Diego attorney Stephen Brigandi, plus approximately $14,200 against Portland local counsel Tim Murphy. Across three briefs filed over five months in a family vineyard dispute, Brigandi submitted 23 fabricated citations and 8 false quotations. The fabrications escalated over time: two in January 2025, seven in April, sixteen in May—even after opposing counsel had flagged the earlier errors. The court found persuasive evidence that Brigandi’s client, a serial pro se litigant, had drafted the briefs using AI and that Brigandi had filed them without verification. An AI policy was created only after the ethics investigation began. Judge Clarke called the case “a notorious outlier in both degree and volume” within “the quickly expanding universe of cases involving sanctions for the misuse of artificial intelligence.”
Two additional cases round out the quarter. In Prososki v. Regan, No. S-25-0295 (Neb. Mar. 20, 2026), the Nebraska Supreme Court found that 57 of 63 citations in attorney W. Gregory Lake’s appellate brief in a divorce case were defective—20 were hallucinated, and 3 were entirely fabricated. The justices interrupted oral argument 37 seconds in. Lake initially denied using AI, then reversed course with an affidavit admitting it was a “grave error of judgment.” He was suspended from practice on April 16. And in Fivehouse v. Department of Defense, No. 2:25-cv-00041 (E.D.N.C.), an Assistant U.S. Attorney named Rudy Renfer filed a brief containing fabricated quotations and misrepresentations of case holdings—errors discovered by the pro se plaintiff, a retired Air Force colonel and former JAG lawyer. The DOJ fired Renfer before the court could impose more than a public reprimand.
The common denominator
When I posed the question in the S&C post—whether the profession’s policy-and-training model could deliver “reasonable assurance” under Rule 5.1(a)—I was examining the ceiling: a firm with rigorous policies, mandatory training, and tracked compliance whose system failed anyway. The Q1 sanctions data reveals the floor.
Not one of the sanctioned lawyers in these six cases had a functioning AI verification practice in place when the defective filings were submitted. Douglass conceded he had not checked his AI-generated work. Seth admitted he used ChatGPT without verifying the output. Brigandi created a firm AI policy only after the investigation began. Lake denied using AI until the evidence became incontestable. Irion and Egli refused to engage with the court’s questions about their cite-checking procedures at all. Renfer’s employer, the Department of Justice, apparently lacked specific AI use protocols for this attorney.
The profession’s dominant response to AI risk has been to adopt policies and training programs. The Q1 data might seem to vindicate that approach: these lawyers got sanctioned because they had no policies, the reasoning goes, so the answer is more and better policies. If the only cases producing sanctions were ones where lawyers had no verification practices at all, the policy model would look adequate. Courts would be catching outliers—the reckless, the careless, the dishonest—and the institutional infrastructure the rest of the profession is building would appear to be working.
That reading is too comfortable, and the S&C filing explains why. Sullivan & Cromwell had everything the Q1 sanctions lawyers lacked: mandatory training, tracked completion, written verification requirements, institutional culture, and the resources to implement all of it. The firm’s policies were, by any reasonable standard, more rigorous than what most firms have. And the policies did not prevent the filing of a motion with roughly 40 corrupted citations and quotations before a Chief Judge of the U.S. Bankruptcy Court. If the sanctioned lawyers represent the failure to try, S&C represents the failure of trying.
The profession faces both problems simultaneously, and neither the floor nor the ceiling offers much reassurance about the space where most lawyers work. Below the floor are lawyers who are not verifying AI output at all—and based on the trajectory Charlotin’s database documents, the number of those incidents is growing, not shrinking. At the ceiling are firms whose compliance infrastructure cannot reliably produce the behavior it prescribes. In between is the bulk of the profession: lawyers who have attended a CLE, read a firm memo, and understood in the abstract that they should verify AI-generated citations, but who work under time pressure, with tools that produce polished-looking output, in conditions that systematically undermine the vigilance the policy demands.
What courts are doing about it
The remedial orders in the Q1 cases are themselves revealing, because they expose what courts think the problem requires—and the gap between judicial remedies and institutional practice.
Monetary sanctions are the default: fines ranging from $250 to $110,000, plus attorney fee awards. These punish after the fact but do nothing to prevent the next incident. Judge Robinson in Lexos Media went further, ordering attorney Joe to implement and certify enhanced firm AI policies, complete with training and verification procedures under ABA guidelines, by a date certain. That order is unusual. Courts rarely dictate the content of a law firm’s internal compliance infrastructure. That a federal judge felt it necessary to order into existence the policies that the profession’s own governance should have required tells you something about the gap between the bar’s aspirations and the bar’s performance.
Bar referrals and disciplinary proceedings represent a sharper escalation. Judge Clarke in Couvrette directed the clerk to send his sanctions order to the Oregon State Bar. The Nebraska Supreme Court referred Lake to the Counsel for Discipline and then suspended him from practice—reportedly the first U.S. bar discipline action to suspend an attorney’s license entirely over AI-related filing errors. The Whiting panel flagged both attorneys for potential disciplinary review. In each case, the court concluded that monetary sanctions alone were insufficient to address what the conduct revealed about the lawyer’s fitness to practice.
The most notable remedial language, though, comes not from the sanctions amounts but from the verification frameworks courts are articulating. The LeTennier court described a “nondelegable duty to verify legal authorities”—language borrowed from malpractice doctrine, where nondelegable duties cannot be discharged by assigning the task to someone (or something) else. The Whiting panel held that citations, “however generated,” must be “personally read and verified.” The Nebraska Supreme Court grounded the obligation in Rule of Professional Conduct 3-501.1, framing verification as a component of baseline competence. These formulations converge on a principle the profession has been slow to articulate this clearly: the obligation to verify is the lawyer’s, it attaches to every citation in every filing, and it cannot be satisfied by relying on the tool that generated the citation to have gotten it right.
The escalation pattern
Three features of the Q1 data deserve attention beyond the individual cases.
First, the sanctions are getting larger. In 2023, Mata v. Avianca produced a $5,000 fine that became the most-discussed sanction in legal technology. In 2024, sanctions in AI citation cases typically ranged from $1,000 to $5,000. In Q1 2026 alone, a single case produced a $110,000 penalty, and the quarterly aggregate exceeded the total for several prior quarters combined. Courts have moved from treating AI citation errors as novel problems warranting measured responses to treating them as professional failures warranting substantial financial consequences.
Second, the consequences are expanding beyond money. Lake’s suspension, Renfer’s firing, Seth’s pro hac vice revocation and mandatory self-reporting to two state bars, Clarke’s referral to the Oregon State Bar—each represents a different mechanism for making AI verification failures carry career consequences. The profession is groping toward a calibration between the response to a first offense (a fine and a stern admonishment) and the response to a pattern of conduct (discipline, suspension, or termination), but the trajectory is clear.
Third, and most troubling, the cases show no indication that the problem is self-correcting. Three years have passed since Mata v. Avianca put every lawyer in the country on notice that AI-generated citations require verification. Every state and federal bar association has issued guidance. Dozens of courts have adopted standing orders or local rules requiring AI disclosure. CLE programs on AI use are now ubiquitous. And the rate of AI citation failures documented in court opinions is accelerating—seventeen in a single day in March 2026, per Professor Volokh’s count. The gap between the profession’s awareness of the problem and the profession’s ability to prevent it continues to widen.
Two problems, one profession
The comfortable version of this story sorts lawyers into two categories: the responsible majority who verify their AI-assisted work, and the reckless minority who get sanctioned. Under that framing, the Q1 sanctions data is encouraging—courts are catching the bad actors, imposing escalating penalties, and the system is working.
The less comfortable version, which the evidence supports better, is that the sanctioned lawyers are the visible fraction of a much larger population. Charlotin’s database documents cases where courts caught the errors. It does not document the filings containing AI-generated inaccuracies that no one caught—cases where opposing counsel did not check, where the judge did not notice, where the errors were close enough to right that they passed without scrutiny. The Q1 sanctions cases represent the visible portion of the problem—the filings where someone did the verification work the filing attorney did not.
As I argued in the S&C post, the profession’s current approach to AI governance optimizes for the wrong variable: the existence of policies rather than the conditions under which lawyers follow them. The Q1 data adds a grimmer layer. For a significant number of lawyers, the policies do not yet exist at all. For the lawyers who do have policies, the S&C experience suggests those policies will not reliably produce verification under the conditions in which lawyers use AI. And for the profession as a whole, the combination of escalating sanctions and accelerating incidents means that courts are punishing the problem more severely while the problem itself continues to grow.
The verification obligation is not new. Lawyers have always been required to confirm that the authorities they cite exist, say what they claim, and support the propositions for which they are cited. What AI has changed is the volume and subtlety of the errors that verification must catch, and the cognitive conditions under which lawyers must catch them. A fabricated case name is easy to detect with a Westlaw search. A real case cited for the right proposition with a corrupted volume number and rewritten quotation is not. The Q1 data shows a profession that has adopted the vocabulary of AI governance—policies, training, verification protocols—without closing the distance between the vocabulary and the practice. Three years after Mata v. Avianca, the distance is growing.
This post draws on sanctions data compiled by ComplexDiscovery and Damien Charlotin’s AI Hallucination Cases Database at HEC Paris Smart Law Hub. Court filings referenced include Couvrette v. Wisnovsky, No. 1:21-cv-00157-CL (D. Or.); Lexos Media IP, LLC v. Overstock.com, Inc., No. 22-2324-JAR (D. Kan.); Whiting v. City of Athens, Nos. 24-5918/5919 (6th Cir.); Deutsche Bank Natl. Trust Co. v. LeTennier, 2026 NY Slip Op 00040 (3d Dep’t); Prososki v. Regan, No. S-25-0295 (Neb.); and Fivehouse v. Department of Defense, No. 2:25-cv-00041 (E.D.N.C.). It extends the analysis of AI governance and Rule 5.1 obligations from an earlier post on the Sullivan & Cromwell filing.