Third Circuit: Unverified AI-Generated Citations Breach Competence Duties and Warrant Attorney Discipline

I. Introduction

Stephen McCarthy v. U.S. Drug Enforcement Administration (3d Cir. Mar. 27, 2026) is a precedential attorney-discipline decision arising out of merits briefing filed in a petition for review of a DEA order. Although the case caption lists Stephen McCarthy, P.A. as petitioner and the DEA as respondent, the opinion’s operative focus is the conduct of petitioner’s counsel, Daniel A. Pallen, Esq. (“Attorney”), a member of the Third Circuit and Pennsylvania bars.

The central issue was not whether AI tools may be used in legal practice, but whether an attorney may submit to a federal appellate court legal authorities that were supplied by a non-attorney and apparently generated by AI—without verifying their existence or accuracy—then downplay the errors after the opposing party identifies them, and only investigate after the court orders production of the cited materials.

II. Summary of the Opinion

The Third Circuit held that Attorney violated Pennsylvania Rule of Professional Conduct 1.1 (competence) and, through Third Circuit Disciplinary Rule 2.1(d), became subject to federal appellate discipline. The Court declined to find a violation of Pa. R.P.C. 3.3(a)(1) (candor to the tribunal) because the rule’s text requires “knowing” falsity and, in the majority’s view, the record left uncertainty about whether Attorney “knew” the statements were false (given that he had not read the cited authorities).

As discipline, the Court imposed a reprimand under Circuit Disciplinary Rule 4.1, and directed that notice be provided to other courts and the National Disciplinary Data Bank pursuant to Circuit Disciplinary Rule 12. The Court emphasized that, going forward, similar conduct may result in harsher sanctions (including monetary sanctions), particularly because this decision establishes the Third Circuit’s first precedential guidance in this area.

Judge Roth concurred in part and dissented in part, agreeing sanctions were warranted but arguing (i) Pa. R.P.C. 3.3(a)(1) was violated (at least twice), and (ii) sanctions should have been more severe and should not be mitigated by the lack of prior Third Circuit AI-specific precedent.

III. Analysis

A. Precedents Cited

1. Pennsylvania disciplinary precedent on Pa. R.P.C. 3.3 (candor)

  • Off. of Disciplinary Couns. v. Wrona, 908 A.2d 1281 (Pa. 2006)
    The majority cited Wrona for the proposition that Pennsylvania law “suggests” reckless conduct may violate Pa. R.P.C. 3.3(a), but ultimately declined to apply a reckless standard because Pa. R.P.C. 3.3(a)(1) is textually framed in terms of “knowingly” making false statements. This citation functions as a cautionary near-miss: it signals that Pennsylvania disciplinary law may support a broader culpability standard, but the Third Circuit majority refused to expand Pa. R.P.C. 3.3 in the absence of a clearer Pennsylvania Supreme Court directive.
  • Off. of Disciplinary Couns. v. Price, 732 A.2d 599 (Pa. 1999)
    The majority relied on Price primarily to contextualize Pa. R.P.C. 3.3’s explanatory comments, reading Comment 3 as focused on statements of fact and Comment 4 as concerned with statements of law. The dissent, by contrast, invoked Price for a more demanding “reasonably diligent inquiry” concept embedded in the comments, arguing that legal argument presented as counsel’s analysis cannot be advanced without such an inquiry.

2. The emerging national body of “AI-citation” competence cases (Model Rule 1.1 analogs)

To support its finding that failure to verify citations implicates competence and thoroughness, the Court cited a line of decisions (mostly 2025) applying substantially identical competence rules (often ABA Model Rule 1.1 or state analogs):

  • In re Martin, 670 B.R. 636 (Bankr. N.D. Ill. 2025): cited for the basic ethical obligation “to review whatever cases they cite.”
  • Benjamin v. Costco Wholesale Corp., 779 F. Supp. 3d 341 (E.D.N.Y. 2025): cited as an example of a competence violation where counsel did not review AI-generated research.
  • Dehghani v. Castro, 782 F. Supp. 3d 1051 (D.N.M. 2025): cited for the proposition that failure to verify generative AI outputs can implicate competence.
  • Puerto Rico Soccer League NFP, Corp. v. Federacion Puertorriquena de Futbol, No. CV 23-1203 (RAM), 2025 WL 1080732 (D.P.R. Apr. 10, 2025): cited for sanctionable inaccuracies and non-existent citations.
  • Robert Thackston v. Daniel P. Driscoll, Secretary of the Army, No. SA-24-CV-00276-FB-ESC, 2025 WL 2715267 (W.D. Tex. Aug. 28, 2025): cited similarly.
  • Hall v. Acad. Charter Sch., No. 2:24-CV-08630-JMW, 2025 WL 2256653 (E.D.N.Y. Aug. 7, 2025): cited similarly.
  • Davis v. Marion Cnty. Superior Ct. Juv. Det. Ctr., No. 1:24-CV-01918-JRS-MJD, 2025 WL 2502308 (S.D. Ind. Sept. 2, 2025): cited similarly.

These authorities did not control the Third Circuit’s interpretation of Pa. R.P.C. 1.1, but they supplied persuasive, cross-jurisdictional confirmation of the same core point: generative AI does not alter the attorney’s non-delegable duty to verify the authorities and propositions presented to a tribunal.

3. Earlier “fake case” milestones and the normalization of the problem

  • Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023): cited as an early, high-profile instance where AI-generated fake cases were described as “unprecedented,” a characterization the Third Circuit notes is no longer accurate.
  • Johnson v. Dunn, --- F. Supp. 3d ---, No. 2:21-cv-1701, 2025 WL 2086116 (N.D. Ala. July 23, 2025): cited for compiling cases involving fabricated authorities, underscoring that the phenomenon is widespread.

4. AI guidance for the judiciary

  • Jud. Conf. Comm. on Codes of Conduct, Guidance for Ethical Use of AI in the Judiciary (Feb. 2, 2026): cited by the Court to emphasize “Accountability, accuracy, and the need for verification” as a professional baseline that extends across legal system actors.

5. Authorities highlighted in the dissent on “no warning needed”

  • Park v. Kim, 91 F.4th 610 (2d Cir. 2024): cited for rejecting the argument that AI-specific rules are required to inform lawyers of their duty to ensure accuracy.
  • Grant v. City of Long Beach, 96 F.4th 1255 (9th Cir. 2004): cited (as written) as another appellate admonition related to hallucinated cases.
  • Bevins v. Colgate-Palmolive Co., No. 25-cv-576, 2025 WL 1085695 (E.D. Pa. Apr. 10, 2025) and Bunce v. Visual Tech. Innovations, Inc., No. 23-cv-1740, 2025 WL 662398 (E.D. Pa. Feb. 27, 2025): cited to show district courts within the Third Circuit have already sanctioned failures to verify client-provided work.
  • Ader v. Ader, 87 Misc.3d 1213(A), 2025 WL 2831332 (N.Y. Sup. Ct. Oct. 1, 2025) and Smith v. Farwell, No. 2282CV01197, 2024 WL 4002576 (Mass. Dist. Ct. Feb. 15, 2024): cited for state-court sanctions arising from unverified generative-AI outputs.

B. Legal Reasoning

1. The core holding: competence (Pa. R.P.C. 1.1) is violated by failing to verify citations

The Court anchored discipline in Pa. R.P.C. 1.1’s requirements of “thoroughness and preparation reasonably necessary for the representation.” Three features of Attorney’s conduct were decisive:

  1. Non-delegable verification duty: Attorney “modified” and “simply incorporated” summaries supplied by a non-attorney without checking the existence or accuracy of the cited DEA adjudications, then signed the brief as an officer of the court.
  2. Escalation after notice: After the Government “catalogued” the failures in its response brief, Attorney still did not check the citations and filed a reply brief characterizing the errors as “immaterial misstatements” and “minor discrepancies.”
  3. Delayed remediation: Even after Attorney “conclusively deduced” AI was involved, he did nothing until ordered by the Court to produce copies and explain verification steps.

This sequence allowed the Court to distinguish ordinary, correctable errors from a sustained lack of diligence over months. The opinion’s practical rule is clear: using AI (or a non-attorney intermediary) does not reduce the lawyer’s duty to verify; it heightens the need for checking.

2. Why the majority declined to find a Pa. R.P.C. 3.3(a)(1) violation

The majority treated Pa. R.P.C. 3.3(a)(1) as textually requiring “knowing” falsity, and it viewed Attorney’s principal failure as not reading the sources rather than knowingly lying about them. The majority acknowledged that Pennsylvania authority (citing Off. of Disciplinary Couns. v. Wrona) may support a reckless standard, but declined to adopt that approach. It also found Attorney’s “good faith” assertion immaterial to the merits dispute, further narrowing the path to a 3.3 finding.

3. The dissent’s alternative: Pa. R.P.C. 3.3’s comments impose a “reasonably diligent inquiry” obligation for legal argument

Judge Roth framed the misconduct as including false legal arguments offered without any “reasonably diligent inquiry,” relying on Pa. R.P.C. 3.3 cmt. 3 and Off. of Disciplinary Couns. v. Price. On that view, counsel’s legal analysis “purporting to be” counsel’s own work cannot be advanced unless counsel believes it to be true after diligent inquiry—an inquiry Attorney admittedly did not conduct. The dissent also emphasized post-notice conduct (after the Government’s response) as independently egregious, supporting stronger sanctions.

4. Choice of sanction: reprimand, with explicit forward-looking escalation

Under Circuit Disciplinary Rule 4.1, the panel could impose any sanction short of suspension/disbarment. It selected a reprimand based on: (i) Attorney’s post–show-cause contrition and claimed corrective measures, (ii) the Court’s view that it had not previously spoken precedentially on AI guardrails, and (iii) limited notice in the show-cause process regarding Pa. R.P.C. 1.1 specifically. Critically, the Court signaled that this mitigation is unlikely to apply going forward and that monetary sanctions would have been likely absent these first-impression considerations.

C. Impact

  • Third Circuit baseline rule on AI and citation verification: This opinion effectively establishes that, in this Circuit, an attorney’s duty of competence includes verifying the existence and accuracy of authorities before filing, and promptly correcting known inaccuracies— regardless of whether AI, clients, or non-attorney staff generated the initial content.
  • Sanctions trajectory: The Court’s explicit statement that monetary sanctions were likely but for “first opportunity” mitigation functions as a warning shot: future panels may treat similar conduct as warranting harsher discipline under Circuit Disciplinary Rules 4.1–4.2.
  • Framework for litigant behavior after notice: The decision underscores that post-notice inaction (failing to investigate after the Government flags errors) can aggravate the misconduct and influence sanctions.
  • Institutional credibility and judicial resource protection: By emphasizing wasted court and government resources and harm to counsel’s credibility as an officer of the court, the opinion positions verification failures as threats to adjudicative integrity, not merely private-lawyer errors.
  • Potential future convergence with candor doctrine: The split between majority and dissent on Pa. R.P.C. 3.3(a)(1) suggests future cases may litigate whether “reasonably diligent inquiry” concepts in the comments can supply the culpability requirement for AI-driven legal argument, even where the attorney claims not to “know” of falsity.

IV. Complex Concepts Simplified

“Hallucinated” citation
A source that does not exist but is presented as real—commonly produced by generative AI tools that output plausible-looking case names, docket numbers, or quotations without grounding them in actual records.
Duty of competence (Pa. R.P.C. 1.1)
The obligation to deliver legal work with adequate knowledge, skill, thoroughness, and preparation. Here, “thoroughness” concretely meant verifying that cited adjudications existed and supported the propositions for which they were offered.
Duty of candor (Pa. R.P.C. 3.3)
The obligation not to mislead a tribunal with false statements of fact or law, and to correct certain false statements previously made. The dispute in this opinion is about the mental state required (“knowingly” vs. potentially reckless or “no diligent inquiry” standards for legal argument).
“Knowing” vs. “reckless”
“Knowing” typically implies actual awareness. “Reckless” commonly means consciously disregarding a substantial risk. The majority declined to extend Pa. R.P.C. 3.3 to reckless conduct on this record, while acknowledging Pennsylvania cases hint in that direction.
Circuit Disciplinary Rule 2.1(d)
A Third Circuit rule allowing discipline when an attorney’s conduct violates the professional conduct rules of a state bar to which the attorney is subject (here, Pennsylvania).
Reprimand
A formal sanction short of suspension/disbarment. Here, it also triggered reporting obligations to other courts and the National Disciplinary Data Bank.

V. Conclusion

McCarthy v. U.S. Drug Enforcement Administration establishes a clear Third Circuit disciplinary precedent: an attorney who submits AI-generated or non-attorney-supplied legal authorities without verification—and who fails to investigate and remediate after errors are identified—violates the duty of competent representation under Pa. R.P.C. 1.1 and may be disciplined under Circuit Disciplinary Rule 2.1(d). Although the panel imposed only a reprimand, it expressly indicated that future cases are likely to draw more severe sanctions now that the Third Circuit has spoken.