The Shadow of Artificial Intelligence on the Judge’s Bench

By now, the bar has learned the hard way that careless use of large‑language models can spawn phantom citations and plausible‑sounding but nonexistent quotations—a theme I have examined with particular attention to civil liability and professional ethics for lawyers in my Altalex article.[1] The novelty of July 2025 is that a similar failure surfaced on the bench. In re CorMedix Inc. Securities Litigation No. 2:21-cv-14020 (D.N.J.) has become a test case for how resilient federal chambers are when drafting practices meet tools that accelerate the work while straining verification.

Here is, in fact, a brief account. On June 30, 2025, the District of New Jersey—per Judge Julien Xavier Neals—filed a 33‑page opinion denying defendants’ motion to dismiss, letting the securities class action move forward.[2] Three weeks later, on July 22, defense counsel submitted a detailed letter reviewing the opinion in detail and flagging more than minor glitches: precedents were cited as if they had outcomes opposite to what the courts actually held; the text attributed quotation marks to opinions that contain no such language; and two statements were presented as if drawn from the operative complaint even though the cited paragraphs do not say that.[3] The next day, July 23, the court entered a Text Order stating that the Opinion and accompanying order “were entered in error,” directed the clerk to remove them from the public docket, and indicated that a new decision would follow.[4] On August 19, 2025, a corrected Opinion issued, reaching the same bottom‑line result—denial of the motion to dismiss—while cleansing the reasoning of compromised authorities.[5]

The errors were specific and verifiable. The June 30 opinion treated Dang v. Amarin Corp. PLC as applying the core‑operations doctrine (a U.S. securities law inference that key facts about a company’s core business are so important that knowledge of them can be attributed to senior management) to defeat dismissal; in reality, Dang granted dismissal and rejected core operations. It invoked In re Intelligroup Securities Litigation as if Sarbanes‑Oxley (SOX) certifications (management statements required by U.S. law that attest to the accuracy of financial reporting and the effectiveness of internal controls) could, on their own, sustain scienter when executives knew of control failures; Intelligroup dismissed and expressly refused to infer scienter from SOX certifications. It cited a Stichting Pensioenfonds Metaal en Techniek v. Verizon decision—misplacing it in the Southern District of New York—for the idea that internal emails and memos supported scienter; the pertinent case in the District of New Jersey granted dismissal and does not discuss such internal correspondence at all. Layered onto those inversions were quotation marks that trace back to nowhere in the opinions cited, including lines imputed to City of Warwick Ret. Sys. v. Catalent, Inc. and Roofer’s Pension Fund v. Papa, and two sentences presented as if they appeared in the Third Amended Complaint when the pinpoint paragraphs do not contain them.[3]

At this point, the underlying systemic problem becomes evident. The episode is not a stray typo or an arguable gloss; it reflects a breakdown in the verification chain that ordinarily vets the doctrine and precedents underpinning judicial analysis. In modern chambers, drafting is distributed across judges, law clerks, and sometimes temporary assistants; automation tools may enter the drafting pipeline for search or summarization. When cross‑checking fails, error is not only likely—it is repeatable. In CorMedix, the danger of a “ghost precedent” was no abstraction: before the Text Order removed the opinion, litigants in a parallel securities case cited the withdrawn decision as supplemental authority, then withdrew that notice once the defects were spotlighted.[6] This illustrates how, in common law systems, such errors can become pervasive and contaminate other proceedings if not corrected promptly.

Commentary naturally asked whether generative AI was in the drafting loop. The docket is silent. But contemporaneous reporting, citing a person familiar with chambers’ process, stated that research produced with an AI tool made its way into a draft that was inadvertently docketed before review—contrary to chamber policy.[7] It is fair to note that the error pattern is consistent with unvetted AI assistance; it is not sound to convert consistency into proof. For lawyers and judges alike, that distinction matters: responsibility and remedies turn on facts established, not on plausible inferences.

Institutionally, the signal cuts both ways. There is reassurance in the court’s rapid self‑correction: within twenty‑four hours of the defense letter, a Text Order pulled the tainted opinion, and a new one followed with the reasoning cleaned up.[4][5] Yet the speed with which the error propagated—appearing as supplemental authority in another matter—shows that current guardrails are thin when the technological accelerator is engaged without methodological seatbelts.[6]

The practical lessons are straightforward. First, process transparency: chambers should study and experiment with protocols that make the drafting chain auditable, distinguishing what is written or pasted from what is merely suggested by external tools, and requiring human double‑checks for any passage that contains quotations, citations, or synthesized precedent. Second, AI governance: where use is permitted, courts should maintain an authorized‑tools whitelist, disable modes prone to fabricated references, and preserve logs for audit; where use is prohibited, the ban should be paired with real spot checks. Third, professional culture: not everything that accelerates is good law. Faster drafting is no virtue if the price is institutional credibility.

At the level of legal ethics, In re CorMedix shifts focus from sanctioning misuse at the bar to shoring up safeguards also on the bench. The point is no longer only the diligence of the advocate who files a brief; it is the robustness of the workshop where reasons are made. Precisely because the August 19 opinion denied dismissal on scrubbed reasoning, the message is not “the result changes,” but “the method matters”: even a correct outcome sits on weak foundations if it rests on apocryphal authorities.[5]

Notes

[1] Marco Perilli, Avvocato & AI: Profili deontologici e responsabilità civile (Aug. 6, 2025), Altalex, https://www.altalex.com/documents/2025/08/06/avvocato-ai-profili-deontologici-responsabilita-civile.

[2] Opinion, In re CorMedix Inc. Sec. Litig., No. 2:21‑cv‑14020 (D.N.J. June 30, 2025), ECF No. 114, available at Justia, https://law.justia.com/cases/federal/district-courts/new-jersey/njdce/2%3A2021cv14020/477937/114/.

[3] Letter from Andrew S. Lichtman, Willkie Farr & Gallagher LLP, to Hon. Julien X. Neals, U.S.D.J., In re CorMedix Inc. Sec. Litig., No. 2:21‑cv‑14020 (D.N.J. July 22, 2025), ECF No. 123, https://storage.courtlistener.com/recap/gov.uscourts.njd.477937/gov.uscourts.njd.477937.123.0.pdf.

[4] Text Order, In re CorMedix Inc. Sec. Litig., No. 2:21‑cv‑14020 (D.N.J. July 23, 2025), ECF No. 126 ("That Opinion and Order were entered in error. Accordingly, the Court directs the Clerk of Court to remove the Opinion and Order at ECF Nos. 114 and 115 from the docket. A subsequent Opinion and Order will follow."), docket noted at CourtListener, https://www.courtlistener.com/docket/60081508/in-re-cormedix-inc-securities-litigation/ (entry for July 23, 2025).

[5] Opinion (corrected), In re CorMedix Inc. Sec. Litig., No. 2:21‑cv‑14020 (D.N.J. Aug. 19, 2025), ECF No. 130, available at Justia, https://law.justia.com/cases/federal/district-courts/new-jersey/njdce/2%3A2021cv14020/477937/130/.

[6] Justin Henry, Judge Scraps Opinion After Lawyer Flags Made‑Up Quotes, Bloomberg Law (July 23, 2025; updated July 24, 2025), https://news.bloomberglaw.com/business-and-practice/judge-withdraws-pharma-opinion-after-lawyer-flags-made-up-quotes (reporting Outlook Therapeutics citation and subsequent withdrawal).

[7] Mike Scarcella, Two U.S. Judges Withdraw Rulings After Attorneys Question Accuracy, Reuters (July 29, 2025, 6:00 PM UTC), https://www.reuters.com/legal/government/two-us-judges-withdraw-rulings-after-attorneys-question-accuracy-2025-07-29/ (reporting that AI‑generated research inadvertently made its way into a draft ruling before review).