Appeal from the United States District Court for the Southern District of Texas USDC No. 4:24-CV-370

Before Elrod, Chief Judge , and Smith and Wilson, Circuit Judges . Jennifer Walker Elrod, Chief Judge :

Pending before the court is an issue that has become central to the ongoing discussions of the relationship between law and technology: the use of artificial intelligence ('AI') in the drafting of legal documents. On December 18, 2025, the court issued an order to show cause as to why Appellant's counsel should not be sanctioned for including, in a brief, quotations, citations, and assertions that were not supported by the underlying case law.

Page 2Having considered counsel’s responses to the show-cause order, we have determined that counsel used artificial intelligence to draft a substantial portion, if not all, of her reply brief and then failed to verify the accuracy of the content generated. We have also determined that she was not forthcoming in her response to the show-cause order. For those reasons, IT IS ORDERED that Heather Hersh pay to the clerk of court within 30 days a sanction of $2,500.

The first high-profile incident of AI-fabricated case citations in the federal courts occurred in the Southern District of New York. Larry Neumeister, Lawyers blame ChatGPT for tricking them into citing bogus case law , AP (June 8, 2023), https://apnews.com/article/artificial-intelligencechatgpt-courts-e15023d7e6fdf4f099aa122437dbb59b. In that case, the plaintiff's AI-generated brief cited seven nonexistent cases. Mata v. Avianca, Inc. , 678 F. Supp. 3d 443, 449–51 (S.D.N.Y. 2023). For example, the brief cited a case called Varghese v. China Southern Airlines Co. , which did not exist, and the federal reporter citation led to a different case altogether. Id. at 151. The brief listed our own distinguished colleague, Judge Patrick Higginbotham, as a member of the Varghese panel, even though that case was entirely fictional. Id. at 453 & n.7.¹

Fabrications of this sort have been dubbed 'hallucinations.' See Snell v. United Specialty Ins. Co. , 102 F.4th 1208, 1230 (11th Cir. 2024) (Newsom, J., concurring) (“Put simply, [a generative AI program] ‘hallucinates’ when, in response to a user’s query, it generates facts that, well, just aren’t true—Page 3 or at least not quite true.'). The hallucination problem has no end in sight, as AI’s tendency to fabricate results arises from the training and structures of AI programs. See Why language models hallucinate , OpenAI (Sep. 5, 2025), https://openai.com/index/why-language-models-hallucinate. As time goes on, these hallucinations grow “all the more insidious and harder to guard against,' as the models are more sophisticated and appear more truthful. Jane Bambauer, Negligent AI Speech: Some Thoughts About Duty , 3 J. Free Speech L. 343, 356 (2023). This problem now often manifests as false quotes or statements of law attributed to real cases, rather than the more easily recognizable fake cases.

Following the Mata case and other reports of AI hallucinations in court filings, our court appointed a three-judge AI Subcommittee to study the issue in Spring 2024. The Subcommittee developed a proposed rule. The rule would have required counsel and pro se litigants to certify either: (a) that no generative AI program was used to prepare any submitted document; or (b) if an AI program was used, that a human checked the AIgenerated text for accuracy. Notice of Proposed Amendment to 5th Cir. R. 32.3 (Jan. 4, 2024), https://www.ca5.uscourts.gov/docs/default-source/defaultdocument-library/public-commentlocal-rule-32-3-and-form-6.

We published the proposed rule on our court website and asked for public comment. Thirteen comments were submitted by judges, professors, and practitioners. Many of the comments reasoned that such a rule was unnecessary because existing rules such as Federal Rule of Civil Procedure 11 and Federal Rule of Appellate Procedure 46(b)(1)(B) already impose an obligation on counsel to submit accurate information to courts.

Ultimately, the Subcommittee recommended to the court that we decline to adopt the proposed rule, and we followed the Subcommittee's recommendation. We instead issued a notice on our court website reminding Page 4counsel of their obligation to review all filings for accuracy. Court Decision on Proposed Rule , https://www.ca5.uscourts.gov/docs/default-source/defaultdocument-library/court-decision-on-proposed-rule.pdf?sfvrsn=5967c92d_2 (last visited Feb. 18, 2026). In doing so, we concluded that existing rules were sufficient to deter misconduct related to generative AI use, without the need for a rule specific to generative AI. Federal Rule of Appellate Procedure 46(c), for instance, allows a court of appeals, after notice and an opportunity to show cause, to “discipline an attorney who practices before it for conduct unbecoming a member of the bar or for failure to comply with any court rule.” We also have “inherent power to impose sanctions for abuse of the judicial process.' Anderson v. Wells Fargo Bank, N.A. , 953 F.3d 311, 315 (5th Cir. 2020). Other courts have noted that submitting a brief riddled with fabricated quotations and assertions is such an abuse. See Park v. Kim , 91 F.4th 610, 615 (2d Cir. 2024) (referring attorney to the Second Circuit's Grievance Panel under 2d Cir. R. 46.2 for failing to make inquiry into the validity of her argument).

Regrettably, despite numerous news stories, CLE presentations, scholarly articles, and judicial entreaties, AI-hallucinated case citations have increasingly become an even greater problem in our courts, and the problem shows no sign of abating. Damien Charlotin, a French lawyer and data scientist, maintains a database that tracks court orders related to AIhallucinated content. AI Hallucination Cases , Damien Charlotin, https://www.damiencharlotin.com/hallucinations (last visited Feb. 18, 2026). As of the date of this order, Charlotin has identified 239 cases of hallucination by lawyers in the United States. Id. 2 Within our circuit, district courts have shouldered the burden of addressing AI hallucinations in court Page 5filings.³ It is a problem that is getting worse-not better. If it were ever an excuse to plead ignorance of the risks of using generative AI to draft a brief without verifying its output, it is certainly no longer so. To ethically use generative AI in the practice of law—which we do not dispute can be helpful if done properly and carefully-a lawyer must 'ensure that the legal propositions and authority generated are trustworthy.' ByoPlanet Int’l, LLC v. Johansson , 792 F. Supp. 3d 1341, 1347 (S.D. Fla. 2025). Failure to do so “abdicate[s] one’s duty, waste[s] legal resources, and lower[s] the public’s respect for the legal profession and judicial proceedings.' Id.

A brief recitation of the facts underlying this case is helpful. This is an appeal of a sanctions award against Shawn Jaffer, plaintiff’s counsel in the district court, and his law firm, Jaffer & Associates. Jaffer sued two defendants for violations of the Fair Credit Reporting Act. The gist of the Page 6case was that Robert Fletcher, Jaffer’s client, was a victim of identity theft and someone else had opened an automobile finance account in his name.

The district court instead determined that “Mr. Jaffer had not done even a minimal investigation of Fletcher’s claims before filing a suit seeking damages that were barred by law, or based on false factual allegations.” 4 As a sanction, the district court ordered Jaffer and his law firm to pay defendant Bridgecrest (a lender) about $20,000 in attorneys’ fees under Federal Rule of Civil Procedure 11 and defendant Experian (a credit reporting agency) about $13,000 in attorneys’ fees under 28 U.S.C. § 1927. We vacated that sanctions order, holding that Jaffer needed a greater opportunity to defend his pre-suit investigation and that the early discovery of the false allegations brought Jaffer’s conduct outside the strictures of 28 U.S.C. § 1927, which only applies when an attorney 'multiplies the proceedings.' Fletcher v. Experian Info. Sols., Inc. , No. 25-20086, 2026 WL 37428, at *3-5 (5th Cir. Jan. 6, 2026).

But a problem remained. Heather Hersh, counsel for plaintiff and a member of the Jaffer & Associates law firm, had filed a reply brief on appeal containing numerous inaccurate citations, quotations, and statements of fact.⁵ We issued a show-cause order, enumerating 16 instances of fabricated quotations and 5 additional serious misrepresentations of law or fact. We directed Hersh, the only attorney to sign the brief, to “explain whether and Page 7how she verified the accuracy of the propositions in her brief.” We strongly suspected that Hersh had used AI to draft a substantial portion—if not the entirety—of her brief. The suspect quotations that we enumerated were:

Citation Quotation In re Deepwater Horizon , 857 F.3d 246, 250 (5th Cir. 2017). The Fifth Circuit has warned that courts must not “act as advocates” when imposing sanctions. Fox v. Vice , 563 U.S. 826, 834-36 (2011). “The question is whether the case was frivolous when filed, not whether it later proved to be so.' Cooter & Gell v. Hartmarx Corp. , 496 U.S. 384, 393 (1990). 'Rule 11 sanctions are to be imposed only for arguments that are frivolous when made, not for those that later prove unsuccessful.” In re Deepwater Horizon , 857 F.3d 246, 250 (5th Cir. 2017). “Due process requires notice and an opportunity to be heard before sanctions are imposed.” Bridgecrest’s brief Bridgecrest concedes its records were produced after the complaint but insists that they “do not retroactively render the claim frivolous.” Vaughan v. Lewisville Indep. Sch. Dist. , 62 F.4th 199 (5th Cir. 2023). That argument fails under Vaughan , which requires courts to “segregate fees incurred defending frivolous claims from those incurred defending non-frivolous ones.”

Page 8Fox v. Vice , 563 U.S. 826, 836 (2011). It also violates Fox , 563 U.S. at 836, which limits recovery to fees “that would not have been incurred but for the frivolous claim.” Fox v. Vice , 563 U.S. 826, 836 (2011). 'Defendants may receive only the portion of their fees that they would not have incurred but for the frivolous claim.” Vaughan v. Lewisville Indep. Sch. Dist. , 62 F.4th 199, 207 (5th Cir. 2023). “The district court must segregate fees incurred defending frivolous claims from those incurred defending non-frivolous ones.” Fox v. Vice , 563 U.S. 826, 836 (2011). “The court must determine which fees were incurred because of the frivolous claim.” Cooter & Gell v. Hartmarx Corp. , 496 U.S. 384, 397 (1990). 'Sanctions must not be used as mere fee-shifting devices.” Bryant v. Mil. Dep't of Miss. , 597 F.3d 678, 682 (5th Cir. 2010). “Sanctions under Rule 11 are collateral to the merits and may be imposed after dismissal.” In re Case , 937 F.2d 1014, 1023 (5th Cir. 1991). 'The court must carefully tie the award to the sanctionable conduct.” Mendez v. County of San Bernardino , 540 F.3d 1109, 1131 (9th Cir. 2008). 'A sanctions order must be supported by a specific finding of bad faith and a clear articulation of the factual basis.”

Page 9Thomas v. Cap. Sec. Servs., Inc. , 836 F.2d 866, 873–74 (5th Cir. 1988) (en banc). 'The district court must carefully articulate the evidence and reasoning supporting its decision to impose sanctions.” Donaldson v. Clark , 819 F.2d 1551, 1556 (11th Cir. 1987). 'Discretion is not unbounded. It must be exercised within the framework of the applicable rules and precedents.”

We also alerted Hersh to the following inaccurate citations and assertions:

•  The brief cites Edwards v. General Motors Corp. , 153 F.3d 242, 246 (5th Cir. 1998), for the proposition that 'sanctions imposed sua sponte without notice or hearing implicate due process and are reviewed de novo.' Edwards does not mention de novo review, sua sponte sanctions, or lack of notice or a hearing.
•  The brief cites In re Deepwater Horizon Inc. , 857 F.3d 246, 250 (5th Cir. 2017), for the proposition that the 'Fifth Circuit has rejected attempts to substitute informal notice for procedural compliance.' In re Deepwater Horizon has no apparent relation to this proposition.
•  The brief asserts that “no formal Rule 11 motion was ever served.” Bridgecrest served its Rule 11 motion on Appellant at least 21 days before filing its Rule 11 motion.
•  The brief cites United States v. Miller , 953 F.3d 1095, 1101 (9th Cir. 2020), for the proposition that “government-issued fraud documentation triggers evidentiary presumption.' We cannot discern this proposition from Miller .
•  The brief asserts that Lewis v. Brown & Root, Inc. , 711 F.2d 1287, 1292 (5th Cir. 1983), reversed 28 U.S.C. § 1927 sanctions where the record

Page 10lacked evidence of vexatious intent. Lewis affirmed a sanctions award issued under § 1927.

Hersh's response was disappointing. She asserted that she had “relied on publicly available versions of the cases, which [she] believed were accurate.' Believing that response to be incredible on its face, the court directed Hersh to answer additional questions. Hersh answered these questions, noting that she “endeavored to answer each question directly and transparently.” We address each question and Hersh’s response in turn.

First, when asked when she became aware of the inaccuracies in the brief, Hersh replied that it was at the time of our show-cause order. We do not at this time have reason to doubt that Hersh first became aware of the inaccuracies when we pointed them out. Of course, if Hersh had discovered the inaccuracies in her reply brief in the three months between its submission and our show-cause order, she should have immediately alerted the court and sought leave to file a correct brief.

Second, when asked what 'publicly available versions of the cases' had led to the inaccuracies described above, Hersh named several wellknown legal databases. The court does not find it credible that these sources produced the hallucinated quotations that appeared in Hersh’s brief. The first source, Google Scholar, does not contain summaries of legal cases that could contain inaccurate quotes.⁶ While the second, third, and fourth sources—CourtListener, Justia, and FindLaw-do provide legal case summaries, we have reviewed the summaries for each case cited in Hersh’s reply brief, and the problematic quotations and propositions are nowhere to Page 11be found.⁷ The remaining two sources, Casetext, which has been spun off into Thomas Reuter's CoCounsel, 8 and vLex, are both generative AI products.

To the extent these products generated the inaccurate summaries, Hersh’s response is misleading in several respects. For one, in her initial response, Hersh stated that she relied on “publicly available versions of the cases” and that “certain paraphrased summaries did not match the reporter

Justia: See, e.g. , Fox v. Vice , Justia, https:// supreme.justia.com/cases/federal/us/ 563/826 (last visited Feb. 18, 2026).

FindLaw: See, e.g. , Fox v. Vice , FindLaw, https:// caselaw.findlaw.com/summary/ opinion/us-supreme-court/2011/06/06/255340.html (last visited Feb. 18, 2026).

Page 12text.' Neither of these descriptions fairly characterizes an AI-generated sentence. In addition, if these products generated the hallucinated quotations, they are not fairly characterized as “free online case repositories,” the use of which phrase suggested that these quotations came from summaries posted that could lead others astray. Further, Hersh did not even mention AI in her first response to the court.

Third , when asked specifically about whether she used AI and how she verified the accuracy of the case citations, Hersch finally admitted to the use of AI, though she neglected to do so in her initial show-cause response. We do not find it credible that Hersh used AI solely to 'help organize and structure [her] arguments and to break up overly long paragraphs.” Even when asked directly, Hersh did not explain the steps she took to verify the factual assertions in her brief, so the court concludes that she took none.

Fourth , when asked about inaccurate assertions of record facts, Hersh provided no explanation of how 'publicly available sources' could have caused record facts to be inaccurate. She had two such factual misstatements in her reply brief.⁹

Fifth , when asked how the paraphrased summaries caused the inaccurate citations and quotations, Hersh continued to attribute the inaccurate quotations to 'paraphrased summaries,' which she claims to have 'mistakenly believed . . . reflected the actual language of the cases.' However, Hersh did not point this court to a single one of those summaries. We have put tremendous effort into attempting to find where these summaries could be located and have been unable to find them. To the extent Page 13Hersh is referring to a case summary generated by AI as a 'paraphrased summar[y],” her response is evasive, misleading, and sanctionable.

In sum, the court finds that Hersh used artificial generative intelligence to draft a substantial portion—if not all—of her reply brief and failed to check the brief for accuracy. It is also likely that she used artificial generative intelligence in her response to the show-cause order. Had Hersh accepted responsibility and been more forthcoming, it is likely that the court would have imposed lesser sanctions. However, when confronted with a serious ethical misstep, Hersh misled, evaded, and violated her duties as an officer of this court.

Modern generative AI may be a new technology, but the same sanctions rules apply, and the rules we have are well equipped to handle these types of cases.¹⁰ First , Rule 46(c) allows us to discipline an attorney who practices before us for 'conduct unbecoming a member of the bar or for failure to comply with any court rule.' Discipline under Rule 46(c) may include monetary sanctions. In re Violation of Rule 28(d) , 635 F.3d 1352, 1360 (Fed. Cir. 2011); see 16AA Wright & Miller’s Federal Practice and Procedure § 3992.2 (5th ed. 2025).

Page 14Conduct “unbecoming a member of the bar” is broad and includes making frivolous arguments and misrepresenting facts or law. See, e.g. , United States v. Martinez-Martinez , 1999 WL 1330642, at *1 (5th Cir. Dec. 15, 1999) (show-cause order under Rule 46(c) for misrepresenting “a critical fact relating to jurisdiction”); Waldon v. Wal-Mart Stores, Inc., Store No. 1655 , 943 F.3d 818, 825 (7th Cir. 2019) (noting that “conduct unbecoming a member of the bar' includes 'deliberately misleading the court or displaying egregious misjudgment”).

The conduct at issue in this case is certainly “unbecoming a member of the bar.” Fed. R. App. P. 46(c). As discussed above, Hersh failed to check her own brief before submitting it, leading her to repeatedly misrepresent the law to the court. Cf. Fed. R. Civ. P. 11(b) and (c) advisory committee’s note to 1993 amendment ('The rule . . . require[s] litigants to 'stop-and-think' before initially making legal or factual contentions . . . . A litigant’s obligations . . . include reaffirming to the court and advocating positions contained in those pleadings and motions after learning that they cease to have any merit.'); Tex. Disciplinary Rules Prof'l Conduct R. 3.01 ('A lawyer shall not . . . assert or controvert an issue [in a proceeding], unless the lawyer reasonably believes that there is a basis for doing so that is not frivolous.”); Id. 3.03(a)(1) (“A lawyer shall not knowingly . . . make a false statement of material fact or law to a tribunal'); Id. 8.04(a)(3) (A lawyer shall not “engage in conduct involving dishonesty, fraud, deceit[,] or misrepresentation”).

Second , we have the “inherent power to impose sanctions for abuse of the judicial process.' Anderson , 953 F.3d at 315; see also Amarsingh v. Frontier Airlines, Inc. , 2026 WL 352016 at *6 (10th Cir. Feb. 9, 2026) (noting a court’s inherent authority to sanction in the context of a brief with AI-hallucinated citations). Submitting a brief riddled with fabricated quotations and assertions is such an abuse. See Park , 91 F.4th at 615 (noting that an 'attempt Page 15to persuade a court or oppose an adversary by relying on fake opinions is an abuse of the adversary system'). Hersh's misleading the court as to the source of her errors further justifies sanctions under our inherent powers. See Ben E. Keith Co. v. Dining All., Inc. , 80 F.4th 695, 703 (5th Cir. 2023) (affirming inherent-power sanctions where attorney discovered misrepresentation and failed to correct it).

We have recognized, in reviewing a district court’s sanctions order, that “an admonition by the court may be an appropriate sanction, in instances where the attorney’s sanctionable conduct was not intentional or malicious, where it constituted a first offense, and where the attorney had already recognized and apologized for his actions.' Jenkins v. Methodist Hosp. of Dallas, Inc. , 478 F.3d 255, 265 (5th Cir. 2007) (reviewing Rule 11 sanctions). These factors do not counsel against sanctions here.

IT IS ORDERED that Heather Hersh shall pay $2,500 in sanctions to the United States Court of Appeals for the Fifth Circuit within 30 days of this order.