The Volokh Conspiracy
Mostly law professors | Sometimes contrarian | Often libertarian | Always independent
Reminder: AI Might Have Hallucinated Quotations Even if You've Confirmed the Citations
"Lead counsel reviewed the cited cases in LexisNexis and confirmed that they were actual decisions relevant to the legal issues in this matter. However, the quotations included in the initial draft of the Memorandum were not independently verified against the official opinions word-for-word."
From a Response to an Order to Show Cause in Pennantia, LLC v. Rose Cay Maritime, LLC (S.D.N.Y.), filed by lawyers from a 10-lawyer firm:
[I.] Introduction and Acknowledgment of Responsibility
Undersigned counsel acknowledge the Court's concerns and accept full responsibility for the errors identified in Defendant's opposition Memorandum of Law (ECF No. 32)("Memorandum"). Specifically, the Memorandum included quotations that had not been confirmed against the cited judicial opinions prior to filing. Counsel recognize that this was a serious lapse in professional obligations under Park v. Kim, 91 F.4th 610 (2d Cir. 2024), deeply regret the error, and extend their sincere apologies to the Court, the parties, and opposing counsel.
[II.] Explanation of Circumstances
Lead counsel reviewed the cited cases in LexisNexis and confirmed that they were actual decisions relevant to the legal issues in this matter. However, the quotations included in the initial draft of the Memorandum were not independently verified against the official opinions word-for-word.
The error occurred during preparation of the Memorandum under a compressed briefing schedule in response to Plaintiff's application for a temporary restraining order and preliminary injunction. On July 28, 2025, undersigned lead counsel used a generative artificial intelligence program to generate an initial outline addressing, among other issues, whether interest may be included in the value of a maritime lien. The draft outline contained citations to authorities, … together with purported quotations attributed to those cases. Counsel confirmed through Lexis Nexis that the cases were valid and relevant to maritime lien law, but failed to verify the accuracy of the quoted language against the official opinions.
Undersigned lead counsel departed for a previously scheduled vacation on August 2. Between July 28 and August 4, the Memorandum underwent more than eleven rounds of revisions; however, during that process the accuracy of the quotations initially generated by the AI program was never verified. Importantly, the Memorandum also cited other authorities … that supported the proposition at issue. The inclusion of the fictitious quotations was inadvertent and occurred without any purpose or intent to mislead the Court or opposing counsel.
Undersigned counsel emphasize that there was never any intent to mislead the Court or opposing counsel. The AI program was used solely as a time-saving aid in the initial drafting process. The subsequent failure to confirm the quotations was inadvertent, yet it constituted a breach of professional responsibility. Undersigned counsel fully recognize that the duty to verify authority rests squarely with the attorney in all circumstances.
[III.] Corrective Measures Implemented
Since the Court's August 26, 2025 Order, Lennon Murphy & Phillips LLC has taken immediate and concrete steps to ensure that this type of error does not recur:
[1.] Conducted a manual review of all citations and quotations contained in Dove Cay's submissions before this Honorable Court that resulted in the redline of Dove Cay's Memorandum of Law (ECF No. 57-1).
[2.] All citations and quotations in any filing must now be verified directly against official sources (Lexis, Westlaw, or court websites). Attorneys must certify this verification before submission.
[3.] The firm will implement training for all attorneys regarding the proper and ethical use of generative artificial intelligence in the practice of law, with particular emphasis on Rule 11 obligations.
[4.] AI-generated text may only be used as an internal drafting aid, and only after complete, manual verification of all authorities, quotations, or other statements by the responsible attorney.
These safeguards are designed to eliminate the risk of inaccurate citations in future submissions.
The filing goes on to argue that sanctions shouldn't be imposed, partly because "the cases cited here … are genuine, published decisions, relevant to maritime lien jurisprudence," and "[t]he error was confined to two fictitious quotations generated by an artificial intelligence tool and mistakenly attributed to genuine cases during an expedited initial drafting process." And the court (Judge Sidney Stein) agreed, in a Sept. 26 order:
In light of counsel's acknowledgment of the fact that the quotations at issue were not independently verified against the actual opinions and acceptance of responsibility and the remedial measures subsequently put in place to avoid a repetition of the errors, the Court will take no further action on this matter at this time.
I think the filing made as good a case as possible for not imposing sanctions, in part because of what struck me as its candid, regretful, and professional tone.
Editor's Note: We invite comments and request that they be civil and on-topic. We do not moderate or assume any responsibility for comments, which are owned by the readers who post them. Comments do not represent the views of Reason.com or Reason Foundation. We reserve the right to delete any comment for any reason at any time. Comments may only be edited within 5 minutes of posting. Report abuses.
Please to post comments
The attorneys could do a CLE training on “how to apologize to the Court”.
Nicely done.
Agreed. And that would be a very useful CLE session.
Although any such CLE should note that the "how" is less important than the fact of the act. Attorneys who double down or make excuses are billions of times more likely to be sanctioned than attorneys who just forthrightly admit fault and apologize. To the court and opposing counsel. No magic words required. Just express guilt and remorse.
I did not expect a 21st century headline to cover AI's learning to be lazy.
AI. Is there anything it cannot mess up?
It's starting to sound like the 21st century update of Frankenstein's monster.
I guess I don't get out enough. I continue to be floored that apparently a good portion of the legal community was not already doing this as a matter of course before LLMs came along.
AI is only highlighting misconduct that has been a long-standing problem. Far too many lawyers (and judges) pretend to present controlling or persuasive legal authority when, in fact, they are misrepresenting the law. Falsehoods by one lawyer, law firm or judge are repeated by others. Even SCOTUS justices do it.
Consider, for example, how SCOTUS justices lied to us in Trump v. Anderson (regarding the power of states to control state elections). That decision was so dependent on lies that it's not surprising no justice wanted to claim responsibility for it (it was issued anonymously (per curiam)). Trump v. Anderson was founded on at least two lies, and SCOTUS justices pretended to "prove" both lies by quoting a prior decision, U. S. Term Limits, Inc. v. Thornton, 514 U.S. 779 (1995).
One lie was the obvious fairytale-type falsehood that all "federal officers" '‘owe their existence and functions to the united voice of the whole, not of a portion, of the people.' ” That's obviously false in multiple respects. In no election do "the whole people" speak with a "united voice."
No president since George Washington can even potentially plausibly claim to have been elected by "the united voice of the whole" of "the people." As every election for national office highlights, the multitude of voters clearly do not speak with one voice. Many don't speak (vote) at all. Moreover, we clearly have no national election.
Every election that culminates in or contributes to a federal employee being elected is a state election. As states like Texas, Missouri, North Carolina, Utah and California have highlighted recently, state legislatures make state laws, e.g., governing who can vote and where they can vote (including redistricting/gerrymandering) to determine how the president is chosen. State legislators are actively working to use partisan gerrymandering to rig not only the upcoming congressional elections, but also the next presidential election. The president clearly has not been (and SCOTUS justices clearly did not believe the president has been or would be) chosen by "the united voice of the whole" of "the people."
As Article II established clearly, the president must be chosen by the votes of only state electors. (That was the point of Trump's Jan. 6 effort to "stop the steal" (by stopping the count of electors' votes) and the point of Trump's efforts to have false slates of electors votes counted for states like Arizona and Michigan). State electors must be chosen according to state law and their votes are expected (or required) to reflect (according to state law) the votes of state voters who often have been corralled into districts by gerrymandering state legislators.
Article II of our Constitution proves the truth of Madison's statement in The Federalist No. 45 that "Without the intervention of the State legislatures, the President of the United States cannot be elected at all. [State legislatures] must in all cases have a great share in [the president's] appointment, and will, perhaps, in most cases, of themselves determine it."
Another obvious falsehood by SCOTUS justices in Trump v. Anderson was that "powers over [the] election" of the president "must be specifically 'delegated to, rather than reserved by, the States.' ” That contention was clearly contrary to the plain text of our Constitution. Our original Constitution was amended very promptly to identify and distinguish "delegated" powers from "reserved" powers in the Tenth Amendment. It's not even plausible that any SCOTUS justice, much less all SCOTUS justices, failed to grasp the significance of the plain text of the single sentence of the Tenth Amendment.
The Tenth Amendment summarized our Constitution in a single sentence: "by the Constitution" We the People "delegated to the United States" limited "powers;" We "prohibited by it [our Constitution] to the States" certain powers; We "reserved to the States" certain powers; and We "reserved" to "the people" the remaining powers.
Stated another way, the Tenth Amendment emphasized that all powers relevant here were "reserved to the States" except to the extent that our Constitution "delegated" limited powers "to the United States" government or "prohibited" certain powers "to the States."
My point is that obvious misrepresentations of controlling legal authority are not new and they are not limited to AI. AI mimics the lies and other falsehoods about the law by lawyers and judges.
Correction: California's redistricting was approved by voters, not by the mere state legislature.
I've had AI hallucinate quotes from docs I uploaded directly into the chat.