Reasonable or Overreach? Rethinking Sanctions for AI Hallucinations in Legal Filings

[EDRM Editor’s Note: EDRM is proud to publish the advocacy and analysis of the Hon. Ralph Artigliere (ret.) and Professor William F. Hamilton. The opinions and positions are those of the Hon. Ralph Artigliere (ret.) and Professor William F. Hamilton, and not those of their institutions, law schools, universities, or clients. © Ralph Artigliere and William F. Hamilton 2025.]

Abstract

This article examines the problem of fabricated or hallucinated citations in legal briefs, focusing on the recent federal decision in Johnson v. Dunn as a case study. That decision highlights the risks when lawyers rely on generative AI without adequate verification. This article analyzes the unique challenges posed by AI hallucinations, emphasizing the importance of distinguishing between reckless misconduct and honest mistakes, and explores whether harsh sanctions are justified given the typically minimal substantive harm in such cases. It argues that sanctions for such errors must be principled, proportionate, and consistent with existing jurisprudence and proposes a four-pillar framework for judges to evaluate AI-related citation misconduct, balancing the need to deter careless practice with the imperative to preserve fairness and encourage responsible use of valuable new technology.

In Johnson v. Dunn,¹ a Northern District of Alabama federal judge sanctioned three experienced attorneys for submitting two motions containing five erroneous citations generated by ChatGPT. The case, analyzed in a previous article by one of the authors,² is an escalation in the judiciary’s handling of errors created by generative AI moving beyond fines and reprimands to applying sanctions causing reputational damage with potential career-altering consequences.

AI hallucination incidents are increasing,³ and they require serious attention.⁴ However, whether severe, public sanctions are the appropriate solution deserves careful consideration, especially when case-specific harm is minimal. What distinguishes citation hallucinations from other legal errors? And why did their appearance in Johnson v. Dunn lead to such broad penalties?

We argue that future judicial responses to AI-related citation errors should be guided by measured professional judgment. In exploring this question, we address:

• Whether judicial reaction to hallucinations is outpacing actual harm;
• The role opposing counsel should play in raising or resolving citation errors; and
• How a principled, repeatable framework can guide sanction decisions without unnecessary overreaction.

Sanctions should remedy wrongs, deter future misconduct, and protect the integrity of the justice system while remaining fair to all parties and participants. A calibrated approach protects litigants, conserves judicial resources, and achieves true deterrence without overreach.⁵

Context and Controversy: What Happened in Johnson v. Dunn

In Johnson v. Dunn, three experienced litigators⁶ from a prominent national law firm were sanctioned for submitting two motions that contained five fabricated case citations generated by ChatGPT. Opposing counsel flagged the citation anomalies in their responses, prompting the court to independently verify the errors. Following a show cause order and a hearing with all attorneys involved, the U.S. District Judge imposed sanctions on the three lawyers connected to the filings.

Citing its inherent authority,⁷ the court found that each attorney’s conduct was “tantamount to bad faith.” The resulting sanctions were severe: public reprimands, disqualification from the case, referrals to the Alabama State Bar, and an unusual order requiring each sanctioned attorney to provide the sanctions order to every client, colleague, opposing counsel, and presiding judge in active matters.

In her order, Judge Manasco acknowledged that these sanctions exceeded those imposed in earlier AI hallucination cases.⁸ Nonetheless, she emphasized that the court’s inherent authority supports sanctions not only to rectify misconduct, but also to deter future violations⁹ and preserve judicial integrity. As she put it:

The court further finds that no lesser sanction will serve the necessary deterrent purpose, otherwise rectify this misconduct, or vindicate judicial authority.¹⁰

By presenting these rationales in the alternative, the court left unclear which factor primarily justified the stiffer penalties. This ambiguity matters. As we’ll discuss below, only one of the three sanctioned attorneys knowingly used ChatGPT. The other two were found to have failed in their supervisory duties but were unaware that AI was used to generate content in the motions. The motions did not misstate the underlying law; the only errors were the citations themselves. The opposing party suffered no substantive prejudice. Had the court limited its response to awarding costs or fees for the extra work caused by the errors, it would have aligned with existing precedent in similar cases.

That leaves deterrence and “vindicating judicial authority,” two expansive rationales, as the principal justifications. The role each attorney played and how these fabricated citations made their way into the final filings informs the analysis of whether these justifications warrant such sweeping sanctions.

Dropping the Ball Despite Safeguards

How could these erroneous submissions happen in a nationally prominent firm with written AI policies, an Artificial Intelligence Committee, and seasoned litigators? The underlying litigation was high-stakes prison constitutional violation litigation for the State of Alabama. Yet the two motions at issue, a motion to compel and a motion for leave to depose a party, were routine discovery filings.

The two motions containing hallucinated citations generated by ChatGPT were prepared by two of the firm’s lawyers and reviewed by a third. The three lawyers had experience working together. One was of counsel to the firm, the second was a partner and assistant practice group leader in the firm’s constitutional and civil rights litigation group, and the third was a partner and practice group leader of the constitutional and civil rights litigation group.

The drafting process was routine and familiar. One attorney (the of counsel drafting lawyer) prepared the initial drafts and forwarded them for review. A second attorney (the reviewing lawyer, a firm partner and an assistant practice group leader) revised the drafts and added content. Critically, that new content included citations generated by ChatGPT, which he did not verify through Westlaw, PACER, or any other authoritative source despite firm policy prohibiting such unsupervised use. He later admitted this was the first and only time he had used ChatGPT for legal drafting and acknowledged it was a serious error in judgment.

A third attorney (the responsible lawyer), a firm partner and the practice group leader, reviewed only the motion to compel, giving it a brief scan focused on factual background, headings, and substantive argument. He did not check the citations or know that ChatGPT had been used.

When the reviewing lawyer returned his edits, the drafting lawyer checked them for factual accuracy, grammar, and style, but not for the accuracy of the citations. He signed and filed the motions, later acknowledging that by doing so he accepted ultimate responsibility for their content, even though he had no reason to suspect that a partner’s edits were noncompliant with firm standards.

In response to the court’s show cause order, the firm and all three attorneys admitted the ChatGPT citations were wrong and sanctionable. The reviewing lawyer even urged that he should bear sole responsibility, given that he had injected the faulty citations into the motions.

How Could this Mistake Happen?

The reasons for submitting hallucinated citations are varied. A lawyer may not appreciate that an LLM is merely a prediction machine capable of making confident but false connections. Most general-purpose LLMs excel at “chat language”—polished, convincing, and delivered with unwarranted certainty. Lawyers may plan to verify citations later, only to have that step lost in the pressures and urgency of litigation. In team settings, miscommunication can occur when each lawyer assumes another is checking the citations. None of this is excusable, but it falls short of malevolence.

The solution is deceptively simple: check the citations. Commercial legal platforms such as Lexis and Westlaw include robust citation verification tools,¹¹ and every lawyer has access to manual verification methods. But the simplicity of a solution does not convert a missed step into an intentional or corrupt practice. Just because a solution is simple does not mean that failing to apply it is an unequivocal sign of gross dereliction. Straightforward safeguards may be missed without malice, and such lapses should not automatically trigger the most severe sanctions. And for the lawyers who were in a reviewing or supervising role, their conduct must be assessed understanding that they had no reason to expect AI was used in revisions by one of the team.¹²

Was the Behavior of All Three Recklessness Tantamount to Bad Faith?

The record shows no intent to deceive or to gain an unfair advantage, only a breakdown in communication and verification. The lawyers’ roles and knowledge were not interchangeable. Two had no idea ChatGPT was used, and each relied, in different ways, on the assumption that experienced colleagues were meeting the same professional and firm standards they themselves followed.

Certainly, a lawyer who signs a filing is responsible for its contents. But is it recklessness tantamount to bad faith to rely on an experienced partner’s edits without rechecking every citation? Similarly, should a practice group leader be sanctioned at that level for failing to spot-check citations in a routine motion prepared by trusted subordinates?

The reviewing lawyer’s conduct of using ChatGPT in defiance of firm policy and failing to verify the output was unquestionably wrong. Yet, does that error, in the context of immediate admissions, an absence of prejudice to the opposing party, and the availability of narrower remedies, truly rise to the “bad faith” threshold? Even where sanctions are merited, a court should impose the least severe sanctions necessary to achieve the goal where the purpose of imposing sanctions is deterrence.¹³

Of course, if an avalanche of hallucinations were causing a breakdown and dysfunction of our judicial system, the problem should be sternly addressed. However, are the numbers of reported hallucination cases¹⁴ that shocking given the literally millions of annual filings in federal and state courts and the increasingly widespread use of generative AI? The courts must respond, but proportionality matters.

Judges, like lawyers, sign off on others’ work and are accountable for errors — yet their own citation mistakes are rarely treated as bad faith. Sanctions should fit the harm and culpability, not reflexively reach for the harshest penalty.

Hon. Ralph Artigliere (ret.) and Professor William F. Hamilton.

The court’s concern over hallucinated or erroneous citations is legitimate, but proportionality matters. Judges themselves have, on occasion, withdrawn orders after discovering factual or citation errors — sometimes tied to staff work or unauthorized AI use.¹⁵ Those mistakes, while damaging to public confidence, are generally not treated as recklessness tantamount to bad faith. Judges remain fully responsible for what they sign, regardless of whether the inaccuracy came from counsel, staff, or technology.¹⁶ This comparison is not to excuse lawyer misconduct, but to show that such errors, wherever they originate, call for sanctions tailored to culpability, harm, and deterrence — not the harshest penalties available.

From Case Study to Guiding Principles

The Johnson v. Dunn decision offers a vivid, and controversial, example of a court responding to AI-related citation misconduct. It also reveals a risk. When judicial concern about emerging technology is not grounded in careful precedent, fears can be exaggerated, precedents can be discounted, and sanctions can shift toward serving a broader indiscriminate and likely ineffective “messaging” purpose rather than fitting the specific misconduct. The result can be confusion among practicing attorneys and a chilling effect on the use of valuable tools and technology by firms and their lawyers. It may also trigger a trend of harsher sanctions in subsequent cases.¹⁷

Consider the more measured response in In re Martin,¹⁸ decided just two weeks earlier on July 18, 2025. There, a bankruptcy attorney filed a brief containing fabricated quotations and non-existent authorities generated by artificial intelligence. The lawyer accepted responsibility, expressed remorse, and withdrew the brief. After reviewing multiple hallucination cases,¹⁹ including some also discussed in Johnson,²⁰ the court ordered the attorney and a senior partner to attend a specialized national AI training program for bankruptcy practitioners and imposed a $5,500 fine.

Likewise, in Lacey v. State Farm,²¹ decided on May 5, 2025, U.S. Magistrate Judge (ret.) Michael Wilner, sitting as a special master, faced a situation factually similar to Johnson. The plaintiff, Mrs. Lacey, was represented by a large team of attorneys from two nationally prominent firms. In preparing a submission to Judge Wilner, one lawyer used content from generative AI that included numerous erroneous citations and legal authorities. Even after Judge Wilner specifically asked for clarification on two of the citations, the resubmitted document corrected only those two, leaving other erroneous citations in place.²²

Judge Wilner applied Rule 11 and inherent authority sanctions but limited them to denying the relief the plaintiff requested²³ and awarding $31,000 in attorney’s fees and costs to be paid by the law firms. He explained his reasoning for not sanctioning the individual attorneys, including the reviewing lawyers and the attorney who used AI:

In a further exercise of discretion, I decline to order any sanction or penalty against any of the individual lawyers involved here. In their declarations and during our recent hearing, their admissions of responsibility have been full, fair, and sincere. I also accept their real and profuse apologies. Justice would not be served by piling on them for their mistakes.²⁴

As in Johnson, the declarations and responses in Lacey included genuine apologies and honest admissions of fault. Yet Judge Wilner chose not to even name the individual lawyers in the order and exercised restraint in considering the professional impact of sanctions.

The conduct in Martin and Lacey was strikingly similar to Johnson, but the sanctions were far less severe. Differences in courts, case types, and procedural contexts can justify some disparity. However, as in Lacey, two of the lawyers in Johnson had no knowledge that AI had been used, and the conduct of the third attorney did not materially differ from that of the attorney in Martin. In fact, the lapses in supervision and citation checking in Lacey could be viewed as more serious than the attorneys in Johnson.

Not all judges see AI-related citation errors as warranting the most severe sanctions. Hon. Xavier Rodriguez of the U.S. District Court for the Western District of Texas, quoted recently in MIT Technology Review, offered this perspective:

I think there’s been an overreaction by a lot of judges on these sanctions. The running joke I tell when I’m on the speaking circuit is that lawyers have been hallucinating well before AI. Missing a mistake from an AI model is not wholly different from failing to catch the error of a first-year lawyer. I’m not as deeply offended as everybody else.²⁵

The disparity in sanctions between Johnson and other similar cases underscores the need for a principled, repeatable framework for evaluating AI-related citation errors. Such a framework should preserve judicial integrity while calibrating sanctions to the actual harm, degree of culpability, and remedial actions taken. A “four-pillar” framework can help strike that balance.

A Framework for Hallucination Sanctions: A practical four-pillar test judges can apply

The following framework (or scorecard) proposed by the authors is a sanctions proportionality matrix. When applying the matrix score, each pillar should receive a High, Medium, or Low score that is explained by the court. Sanctions escalate when scores across multiple pillars are High.

1) Harm to the case and system

• Case prejudice (delay, expense, lost merits opportunity)
• Systemic harm (required excessive court time, damaged public confidence, required record correction)
• Contamination (did the error propagate into other filings or rulings?)

2) Culpability of each lawyer

• Knowledge/intent (submitter knew or should have known; ignored firm policies or court guidelines)
• Role-specific duty (signer’s non-delegable Rule-like duty; supervisor’s oversight; drafter’s verification)
• Prior notice/training/policies (clear warnings; explicit bans; prior incidents)

3) Remediation and candor

• Speed and completeness of correction (self-report vs. opponent-revealed)
• Scope of remedial audit (single document  vs. firm-wide; the utilization of outside validation audit)
• Cost shifting (payment of incurred fees and costs of the non-offending party)

4) Deterrence context

• Recency and frequency (pattern vs. one-off)
• Alternative, narrower remedies available (targeted education, fee award, filing restrictions)
• Collateral effects (impact on other clients uninvolved in the case; reputational impact beyond impacted cases)

The Sanctions Ladder that Maps to the Four Pillar Test

• Low/Medium overall: corrective filing; fee shifting; written admonition; mandatory CLE.
• Medium/High: public reprimand; targeted publication to parties in the case; limited filing restrictions in the case.
• High: removal from the case; court-specific practice suspension; bar referral.
• Extreme (multiple Highs with aggravators): broader suspension or referral with interim restrictions.

A Four-Pillar Framework for Sanctioning AI-Related Citation Misconduct Applied to Johnson v. Dunn

Drawing from established sanctions jurisprudence under a court’s inherent authority, and professional responsibility standards, let’s apply the four-pillar test in Dunn v. Johnson.

1. Harm to the Proceeding or the Parties

• Johnson v. Dunn involved fabricated citations, but the underlying law was not disputed, the motions were routine discovery requests, and opposing counsel was not materially disadvantaged beyond the work of identifying and pointing out the errors.
• Rationale: Courts should distinguish between AI-related errors that actually mislead the court or prejudice a party, and those that are promptly corrected with minimal substantive impact. The guiding question should be: Did the AI-generated error meaningfully alter the court’s decision-making or burden the opposing party beyond correcting the error?

2. Culpability and State of Mind

• In Johnson v. Dunn, only one attorney knowingly used ChatGPT, and he admitted the conduct was contrary to firm policy and professional obligations. The other two relied on his edits without knowledge of AI involvement.
• Rationale: Sanctions law recognizes a spectrum from negligence to recklessness to bad faith.²⁶ Recklessness tantamount to bad faith justifies harsher measures; mere negligence typically does not.
• The guiding question should be: Was the misconduct the result of intentional disregard, reckless indifference, or an understandable (though sanctionable) lapse?

3. Remediation and Cooperation

• After discovery of the errors, the Johnson v. Dunn attorneys admitted fault, apologized, and the firm undertook extensive remedial steps, including reviewing 2,400 citations in 330 filings across 40 federal dockets, finding no other fabricated citations.
• Rationale: Sanctions analysis should account for whether the lawyer or firm took prompt, good-faith action to correct the record, cooperate with the court, and prevent recurrence. The guiding question should be: Did the response demonstrate acceptance of responsibility and genuine efforts to fix the problem?

4. Deterrence and Proportionality

Courts have a wide variety of sanctions available for the purpose of deterrence, but the court should use sanctions that are limited to what suffices to deter repetition of the conduct or comparable conduct by others similarly situated.²⁷

• The Johnson v. Dunn court justified its sanctions as necessary to deter future misconduct and “vindicate judicial authority,” even though prior cases imposed far less severe penalties for similar AI-related errors.
• Rationale: A deterrence rationale is valid, but sanctions should be calibrated so they are neither symbolic overreach nor a mere “cost of doing business.” Disbarment level penalties should be reserved for misconduct with high malevolence and actual harm. The guiding question should be: Is the sanction proportionate to the harm and culpability, and will it deter future violations without chilling legitimate advocacy or technology use?

Why This Matters

Applied to Johnson v. Dunn, our four pillar framework might have led to reduced and differentiated sanctions—still meaningful, but tailored to each lawyer’s role, knowledge, and actions. Going forward, courts can use these pillars to ensure AI-related sanctions are consistent, fair, and aimed at the right target: protecting the integrity of the justice system without reflexively imposing career-altering penalties for every hallucination incident.

All Hands on Deck: The Role of Opposing Counsel

Proportionality in sanctions begins before the court ever gets involved. In Johnson v. Dunn, opposing counsel identified the erroneous citations in the submissions. How might they have responded other than by calling out the errors in their response memorandum to the court?

The legal system works best when all participants, including counsel on both sides and the court, accurately identify the law governing the case. Opposing counsel’s first obligation is to their client, but professional courtesy and efficiency favor raising citation errors promptly and directly with the other side when circumstances allow. Often, a phone call or email can resolve the issue faster, more cheaply, and without lasting damage to reputations. In many cases, a cooperative fix preserves credibility on all sides and avoids unnecessary motion practice. There is a difference between protecting your client’s interests and turning an error into a public spectacle when it could otherwise be corrected quickly.

Of course, not all errors can or should be handled informally. Where an error is repeated, ignored,²⁸ or causes material prejudice, formal court involvement is justified. But when an isolated, admitted mistake can be remedied without impairing the client’s case, immediate escalation to sanctions invites disproportionate outcomes and undermines the profession’s cooperative norms.

Courts, too, benefit from a cooperative approach. When opposing counsel and the filing attorney resolve errors early, judicial resources are preserved for disputes that genuinely require adjudication. Sanctions then become a tool of last resort, reserved for bad faith, repeated violations, or conduct causing real harm—not an inevitable consequence of every lapse.

Closing Note for All Lawyers

AI is here to stay—but so are your professional duties. Don’t trust, verify. Whether the source is ChatGPT, a premium legal AI tool, or a colleague you’ve known for twenty years, the rule is the same: open the case, read it, confirm it, and make sure it says what you claim it says. That’s how you protect your client, your license, and the integrity of the justice system.

Practice Pointers to Prevent AI-Related Citation Errors

For Practitioners

• Treat general-purpose LLMs (e.g., ChatGPT, Claude, Copilot) as brainstorming tools only—never for citations.
• Even with legal-specific AI tools, verify every assertion: open the case, confirm the proposition, check controlling status, and review subsequent history.
• Teamwork protocols:
  • Clarify responsibilities for adding and verifying citations.
  • Log who added each citation, how and when it was checked, and by whom.
  • Before signing, review the log, sample-check cites, and require verification notes.

For Opposing Counsel

• Verify the other side’s citations when drafting responses—don’t assume accuracy.
• Raise errors promptly but professionally; a direct phone call or email may resolve the issue without court involvement.
• Document discovery of citation errors and any resolution attempts to preserve the record.
• Consider whether a motion is necessary, or whether corrective action short of judicial intervention will protect your client.

For Courts

• When hallucinations appear, focus first on the case-specific harm and the responsible party’s intent, not the broader tech trend.
• Distinguish between bad-faith misuse and negligent oversight; calibrate sanctions accordingly.
• Consider remedial orders—such as costs, education, or monitored filings—before imposing career-altering penalties.
• Apply consistent standards so similar conduct results in similar consequences, regardless of public attention or AI involvement.

For Law Firms

• Establish clear AI use policies: define approved tools, prohibited uses, and verification protocols.
• Train all attorneys and staff on AI’s risks, including hallucinations and confirmation bias.
• Require verification logs for all filings; mandate at least one independent citation check before submission.
• Hold everyone accountable—from junior associates to practice group leaders—for ensuring accuracy and compliance with firm policy.
• Audit compliance periodically to ensure policies are being followed in practice.

CONCLUSION

Johnson v. Dunn illustrates both the challenges and the stakes of AI-related misconduct in legal filings. The court’s sanctions were unquestionably severe, more severe than in prior hallucination cases such as Mata v. Avianca, In re Martin and Lacey v. State Farm, and they raise important questions about consistency, proportionality, culpability, and deterrence. While protecting the integrity of the judicial process is paramount, sanctions that overshoot the facts and intent risk undermining fairness and chilling responsible technology use.

Ultimately, sanctions for AI-related citation errors should fit the misconduct and the moment. Proportionality protects both the integrity of the judicial process and the fairness owed to practitioners navigating new technological terrain.

The four-pillar framework we propose—harm, culpability, remediation, and proportionality—offers courts a principled way to separate egregious, bad-faith misuse of AI from sanctionable but less culpable errors. Applied consistently, it can ensure that sanctions remain calibrated to the specific misconduct, that penalties deter without destroying careers unnecessarily, and that justice is served without yielding to institutional overreaction or public alarm.

The legal profession’s challenge is not to fear AI, but to master it responsibly. Judges, lawyers, and firms all have roles to play in setting clear expectations, verifying accuracy, and fostering a culture where technology is a tool, not a shortcut. Meeting that challenge can help us avoid the next Johnson v. Dunn, not by luck, but by design.

Notes

1. 2025 U.S. Dist. LEXIS 141805 (N.D. Ala. July 23, 2025). ↩︎
2. Artigliere, R., When AI Policies Fail: The AI Sanctions in Johnson v. Dunn and What They Mean for the Profession, August 1, 2025, found at https://www.jdsupra.com/legalnews/when-ai-policies-fail-the-ai-sanctions-9043268/. ↩︎
3. A website tracking AI hallucination cases by lawyers, pro se litigants, judges, experts, and others is found at https://www.damiencharlotin.com/hallucinations/. The numbers appear to rise daily and there is no doubt this is a serious concern for courts and trial lawyers on both sides of every case. Another source with even larger numbers of hallucinations is found at https://www.ailawlibrarians.com/2025/08/10/coming-soon-the-interactive-genai-legal-hallucination-tracker-sneak-peek-today/. ↩︎
4. Is there a continuing series of missteps that merit deterrent sanctions? Yes, but the number of reported incidents is still a small fraction of all court filings, and the errors are caught by the filer, the opposing party, or the court without damage to the case. There is no empirical evidence that the “hallucination problem” is overwhelming. The reported cases are relatively few given the millions of yearly filings in federal and state courts. If the opposing party is harmed or costs occur, sanctions can restore the balance. The solution is proportionate sanctions and education. A chart of hallucination court cases worldwide is found at AI Hallucination Cases Database – Damien Charlotin. ↩︎
5. This article deals with federal cases, which are reported sooner and more frequently than state court cases. Nonetheless, most state sanctioning authority is based on frameworks with the same underlying principles as federal law. The concepts and recommendations we put forth are equally applicable in state court cases involving similar issues and circumstances. ↩︎
6. Two other lawyers whose names appeared on one or both motions were not sanctioned because they had nothing to do with drafting or reviewing the submissions. The law firm was not sanctioned because the court reasoned that the firm had AI policies and practices in place that showed it acted reasonably in its efforts to prevent AI misconduct and doubled down on its precautionary and responsive measures when this order to show cause unfolded. The court saw no evidentiary basis for a finding that the firm acted in bad faith or with such recklessness that its conduct was tantamount to bad faith. ↩︎
7. The court found that it could not apply Rule 11 sanctions because these were discovery motions. 2025 U.S. Dist. LEXIS 141805 *35-6. Local Rule 83.1(f) provides that attorneys may be disciplined for acts or omissions that are inconsistent with the Alabama Rules of Professional Conduct, but Alabama Rule of Professional Conduct 3.3 provides that a lawyer shall not “knowingly” make a false statement of material fact or law to a tribunal. Therefore, the court turned to inherent authority sanctions, requiring “a finding of subjective bad faith or something tantamount to it is necessary to support a sanction issued pursuant to a court’s inherent power.” Id. at *38. ↩︎
8. Among the many cases Judge Manasco cited in the Johnson v. Dunn case were: Mata v. Avianca, Inc., 678 F. Supp. 3d 443, 448 (S.D.N.Y. 2023); see, e.g., Dehghani v. Castro, No. 2:25-cv-00052-MIS-DLM, 2025 U.S. Dist. LEXIS 90128, 2025 WL 1361765 (D.N.M. May 9, 2025); Bevins v. Colgate-Palmolive Co., No. 25-576, 2025 U.S. Dist. LEXIS 68399, 2025 WL 1085695 (E.D. Pa. Apr. 10, 2025); Ferris v. Amazon.Com Servs., LLC, No. 3:24-cv-304-MPM-JMV, 2025 U.S. Dist. LEXIS 73784, 2025 WL 1122235 (N.D. Miss. Apr. 16, 2025); United States v. Hayes, 763 F. Supp. 3d 1054 (E.D. Cal. 2025), reconsideration denied, No. 2:24-cr-0280-DJC, 2025 U.S. Dist. LEXIS 68016, 2025 WL 1067323 (E.D. Cal. Apr. 9, 2025); Sanders v. United States, 176 Fed. Cl. 163 (Fed. Cl. 2025); Wadsworth v. Walmart Inc., 348 F.R.D. 489 (D. Wyo. 2025); Gauthier v. Goodyear Tire & Rubber Co., No. 1:23-cv-281, 2024 U.S. Dist. LEXIS 214029, 2024 WL 4882651 (E.D. Tex. Nov. 25, 2024); Park v. Kim, 91 F.4th 610 (2d Cir. 2024). See 2025 U.S. Dist. LEXIS 141805 *32. Also cited were the following: Versant Funding LLC v. Teras Breakbulk Ocean Navigation Enters., LLC, No. 17-cv-81140, 2025 U.S. Dist. LEXIS 98418 , 2025 WL 1440351, at *7 (S.D. Fla. May 20, 2025) (imposing monetary sanctions ranging from $500 to $1,000); Ramirez v. Humala, No. 24-cv-424-RPK-JAM, 2025 U.S. Dist. LEXIS 91124, 2025 WL 1384161, at *3 (E.D.N.Y. May 13, 2025) (imposing a $1,000 monetary sanction); Nguyen v. Savage Enters, No. 4:24-cv-00815-BSM, 2025 U.S. Dist. LEXIS 37125, 2025 WL 679024, at *1 (E.D. Ark. Mar. 3, 2025) (imposing a $1,000 monetary sanction); Wadsworth, 348 F.R.D. at 499 (imposing monetary sanctions ranging from $1,000 to $3,000); Gauthier, 2024 U.S. Dist. LEXIS 214029, 2024 WL 4882651 at *3 (imposing a $2,000 monetary sanction); see also Lacey v. State Farm Gen. Ins. Co., No. 2:24-cv-05205-FMO-MAA, 2025 U.S. Dist. LEXIS 90370, 2025 WL 1363069, at *5 (C.D. Cal. May 5, 2025) (ordering plaintiff’s lawfirms to pay defendants $31,100 in fees and costs). 2025 U.S. Dist. LEXIS 141805 *58-59. ↩︎
9. The court stated: “… there is persuasive precedent in this Circuit for calibrating sanctions issued pursuant to a court’s inherent power to serve deterrence purposes. See Boe v. Marshall, 767 F. Supp. 3d 1226, 1295-97 (M.D. Ala. 2025) (imposing sanctions for bad-faith judge-shopping to serve deterrence purposes).” ↩︎
10. 2025 U.S. Dist. LEXIS 141805 *60-1. ↩︎
11. WestCheck (Thomson Reuters) integrates with your word processor or can be used standalone to analyze legal documents. Shepard’s BriefCheck (LexisNexis) runs through Lexis and uses Shepard’s citation service to validate legal authorities. ↩︎
12. The court expressly noted that two of the three sanctioned attorneys did not know the reviewing lawyer had used ChatGPT and had no reason to expect AI was used in the revisions. The drafting lawyer testified that he had never used a publicly accessible generative AI chatbot for legal citations, was unaware any partner had done so, and typically relied on partners’ edits without rechecking their citations. The supervising lawyer similarly stated he was unaware of any AI use by team members and did not use such tools himself. ↩︎
13. Mata v. Avianca, supra n. 8 at 465. ↩︎
14. For an idea of numbers of cases, see the two websites tracking these incidents, supra n. 3. ↩︎
15. See Scarcela, M., Two US judges withdraw rulings after attorneys question accuracy, Reuters.com (Jul. 29. 2025) found at https://www.reuters.com/legal/government/two-us-judges-withdraw-rulings-after-attorneys-question-accuracy-2025-07-29/; Schlegel, S., Judges Using AI Wrongly Is Not Just a Problem. It’s a Crisis Waiting to Happen, Jul. 24, 2025, found at Judges Using AI Wrongly Is Not Just a Problem. It’s a Crisis Waiting to Happen. ↩︎
16. Similarly, in Florida state court, judges are taught and the law provides that delegating the task of drafting sensitive, dispositive orders to counsel, and then uncritically adopting the orders nearly verbatim would undermine the appearance of justice. See, e.g., Corporate Mgmt. Advisors, Inc. v. Boghos, 756 So. 2d 246 (Fla. 5^th DCA 2000). Simply put, judges should verify the facts and caselaw they depend upon in reaching a decision regardless of the source. Done right, this is yet another safeguard against injustice from a false citation. ↩︎
17. Signaling a dangerous trend, in a case following and citing Johnson v. Dunn, a federal Magistrate Judge in Arizona issued severe sanctions against an attorney from Washington state. The case involved numerous citation errors purportedly the result of use of AI in a brief prepared by a contract attorney and reviewed, signed, and submitted by the attorney who was sanctioned under Rule 11. The sanctions were in some respects more severe than those in Johnson v. Dunn. The sanctions included: 1. The pro hac vice status of the attorney in Arizona federal court was revoked and Counsel was removed from the case; 2. Plaintiff’s Opening Brief was stricken; 3. The attorney was ordered to promptly serve a copy of the Order on Plaintiff; 4. The attorney was ordered to write a letter to the three Judges to whom she attributed fictitious cases, notifying them of her use of fake cases with their respective names attached; 5. The attorney was ordered to transmit a copy of this Order to every Judge who presides over any case in which Counsel is attorney of record; and 6. The Clerk of Court’s Office was directed to serve a copy of the Order on the Washington State Bar Association, of which Counsel is a member; and 7. If Counsel is a member of any other state’s bar, she was ordered to serve a copy of the Order on that state’s bar office. See Mavy v. Comm’r of SSA, 2025 U.S. Dist. LEXIS 157358 at *28-32 (D. Az. Aug. 14, 2025). ↩︎
18. 2025 WL 2017224 (U.S. Bankr. Ct. N. D. Ill. Jul. 18, 2025). ↩︎
19. Id. at *5. ↩︎
20. See list of cases discussed in Johnson v. Dunn, supra n. 8. ↩︎
21. Case No. CV 24-5205 FMO (MAAx) (C.D. Cal. May 6, 2025) found at Lacey v. State Farm | DocumentCloud. ↩︎
22. “During my initial review of Plaintiff’s brief, I was unable to confirm the accuracy of two of the authorities that the lawyers cited. I emailed the lawyers shortly after receiving the brief to have them address this anomaly. Later that day, K&L Gates re-submitted the brief without the two incorrect citations – but with the remaining AI-generated problems in the body of the text. An associate attorney sent me an innocuous e-mail thanking me for catching the two errors that were ‘inadvertently included” in the brief, and confirming that the citations in the Revised Brief had been ‘addressed and updated.’ I didn’t discover that Plaintiff’s lawyers used AI – and re-submitted the brief with considerably more made-up citations and quotations beyond the two initial errors – until I issued a later OSC soliciting a more detailed explanation.” Id. at ¶ 8-9. ↩︎
23. As a “deterrent” sanction, the court struck the briefs submitted by the offending party and denied the discovery relief Plaintiff sought in the proceedings leading to the bogus briefs. This aspect is notable because it penalized the innocent party as part of the deterrence rationale. The Special Master explained: “If the undisclosed use of AI and the submission of fake law causes a client to lose a motion or case, lawyers will undoubtedly be deterred from going down that pointless route.” Id. at ¶ 20. He further acknowledged: “I recognize that Mrs. Lacey is clearly not at fault for the AI debacle, but will bear this outcome as a consequence of her lawyers’ actions. She will not, however, be financially responsible for the monetary awards described in this order. Those will fall solely on the lawyers and their firms.” Id. at ¶ 24. ↩︎
24. Id. at ¶ 25. ↩︎
25. O’Donnell, J., Meet the early-adopter judges using AI, MIT Technology Review (August 11, 2025) found at https://www.technologyreview.com/2025/08/11/1121460/meet-the-early-adopter-judges-using-ai/. ↩︎
26. See, e.g., Chambers v. NASCO, Inc., 501 U.S. 32, 47, 111 S. Ct. 2123, 115 L. Ed. 2d 27 (1991); Roadway Exp., Inc. v. Piper, 447 U.S. 752, 767, 100 S. Ct. 2455, 65 L. Ed. 2d 488 (1980) (stating that inherent powers require a finding that “counsel’s conduct . . . constituted or was tantamount to bad faith”). Both of these cases were cited by the court in Johnson v. Dunn. ↩︎
27. Mata v. Avianca, supra n. 8 at 465. ↩︎
28. The lawyer in Mata v. Avianca made the mistake of not admitting and correcting the hallucination errors when they were brought to his attention by the other side. “Despite the serious nature of Avianca’s allegations, no Respondent sought to withdraw the March 1 Affirmation or provide any explanation to the Court of how it could possibly be that a case purportedly in the Federal Reporter or Federal Supplement could not be found.” Mata v. Avianca, supra n. 8, at p. 450. ↩︎

Assisted by GAI and LLM Technologies per EDRM GAI and LLM Policy.