I remember, distinctly, hitting refresh over and over on Westlaw and Lexis on the night of Thursday, June 22, 2023. I was on a vacation with my wife and some friends at the beach and would not get off my laptop. Not that I am particularly well-disposed to the beach, or to a noisy beach house, but at this moment, I couldn’t wait to see how a new era of legal information history would be born.

Earlier that day, Judge Castel of the Southern District of New York had issued his opinion and order on sanctions in the now infamous case of Mata v. Avianca, 678 F. Supp. 3d 443. The case was the first high-profile instance of a lawyer using hallucinated legal authority in a brief, and plenty has been written about it, its name evoked like the bogeyman–by myself even–when discussing some of the risks of using GenAI in legal research. I pulled the filing from Bloomberg Law’s docket service within an hour of so of its publication, and I noticed something interesting. Judge Castel had chosen to discuss the hallucinated cases by name, and to attach, as an appendix to his opinion, a copy of one of the hallucinated cases cited by counsel.

The spurious case was titled “Varghese v. China Southern Airlines Co.” and allegedly originated in 2018 from the Court of Appeals for the Eleventh Circuit. (There is a case captioned Varghese v. China Shenghuo Pharmaceutical Holdings, Inc. from the S.D.N.Y. in 2008, bearing no relation to the fabricated case.) The hallucinated “Varghese,” in the judge’s PDF opinion, was watermarked “DO NOT CITE OR QUOTE AS LEGAL AUTHORITY,” and yet, in providing this cautionary example, Judge Castel created something of a problem for the legal publishers that make American case law available. He’d done the thing he warned against: he’d cited a fake case.

The next morning, I fired off some emails to trusted contacts at the legal information databases to follow up on what had happened. After all, this was a matter of first impression, and based on what little I had to that point been able to learn about the ways in which legal research companies add case law to their databases, I had a suspicion that things would be interesting.

In fact, both of the major legal research platforms, Westlaw and Lexis, initially pulled in the full text of the fake Varghese case, adding hyperlinks to the citations within, which led to other fake cases. This was not great! I am sure that my reaching out had little to do with it, but within a few days, the language of the fake case had been removed from both databases, disclaimers were added, and the fake case itself was transformed into an image inserted into the opinion, under glass, as it were, able to be examined but not touched. This was much better.

It has been nearly three years since that less than pleasant time at the beach, and by now, legal information providers have a process for handling hallucinated authorities that are cited in court opinions. They’d have to, as pro se litigants, lawyers, and at least one judge, continue to cite fake cases, real cases with fabricated quotations, or misrepresentations of holdings or propositions of law resulting from irresponsible use of GenAI. A taxonomy of hallucination types continues to evade us, but courts tend to recognize the plainer “fake case” or “fake quote from a real case” and the trickier “inaccuracies,” including misrepresentations or misstatements of the law from real cases.

This is not, by the way, a blog post about how lawyers can avoid citing hallucinated authorities (automated citation checkers can help, as can actually locating and reading the case). Nor is it about a lawyer’s professional responsibility for not citing hallucinated authorities (check out, for example, Judge Robert T. Numbers II’s recent amended order in Fivehouse v. U.S. Dep’t of Def., No. 2:25-CV-00041-M (E.D.N.C. May 7, 2026), soon to be published in the Federal Supplement). It also is not about the challenges that court staff face in checking for and sorting out hallucinations (it’s burdensome, no doubt, or the Florida Supreme Court would not have acted, in late May, to amend the state’s court rules to require that filings verify “the legal authorities identified exist and are accurately cited,” No. SC2026-0673, soon to be published in the Southern Reporter). And finally, it isn’t about how legal research systems employing retrieval augmented generation, or RAG, don’t hallucinate (they do, at rates of 17%-33%, according to a 2025 article, Varun Magesh et al., Hallucination Free? Assessing the Reliability of Leading AI Legal Research Tools, 22 J. Empirical Legal Stud. 216, see, e.g., United States v. Farris, 171 F.4th 920 (6th Cir. 2026)).

Rather, what I mean to talk about is the method the databases have chosen to represent hallucinated authority, and give some sense of the scope of the problem. Scholar Damien Charlotin maintains a database of AI Hallucination Cases. According to the data collected, as of the start of June 2026, there have been 1,058 cases in the U.S. reporting on either a litigant or attorney’s improper use of GenAI in a court filing. Interestingly, the split is about 60% pro se litigant to 40% attorney, so please do not think that the problem is confined just to non-legally-trained individuals. Charlotin’s data points to approximately 720 federal cases discussing the use of hallucinated authority and about 340 state cases.

So, what about the legal research platforms? Both employ clear warnings before any case that cites to hallucinated authority. Usually these cases are disciplinary in nature, though the response of courts has really been quite varied, with the issue garnering more or less attention based on the conduct of whoever was responsible and how widespread the hallucinations were. On Westlaw, within a day or two of such a case being added to the database, you will see the following notice near the top of the page:

Editor’s Note: This document contains discussion of unverified citations, likely generated by AI tools used by a party or counsel. These citations are not attributable to the Court. The unverified citations have been preserved as part of the official record, but links are unavailable.

This language changed some time in the past month or so, and the only reason I figured it out is because the WestClip alert I had set up to track these cases stopped working around then.

On Lexis, the approach to notifying users about the discussion of, and citation to, hallucinated cases has changed even more recently (indeed, as a draft of this post was being prepared for publication). As of this writing, an evolving strategy is being rolled out across the platform. After using warning language that had been in place since the time in late June 2023 when I was emailing with the Lexis representative, the company is now taking a slightly different path. Soon, users will see the following in a box at the top of the page:

Notice: This decision contains references to unverifiable citations that may be invalid. They are relevant to the decision and have not been edited or altered.

The notice will be preceded by a pink triangle Shepard’s signal icon with a red exclamation mark in the center. As Lexis is implementing this new warning, the text of the warning is not yet indexed and full-text searchable, but I have been assured it will be soon. A plan for more additions to the warning is also in the works: users will be able to click to advance to the first suspected hallucinated case, with an explanation of what kind of hallucination it is (described by a Lexis representative as “a fictional citation, a fictional case name associated with a real but unrelated citation, a fictional quote attributed to a real case name and citation, or a fictional holding attributed to a real case name and citation.”)

To summarize then, both of the major legal research platforms are taking slightly different approaches: if a court discusses or cites fake, misrepresented, or mischaracterized authority, a warning will be issued, and while the citation will remain in the opinion, it will not be hyperlinked to anything on the platform. Both databases have struggled on this last point, occasionally hyperlinking from a fake case to the real case with the same citation, but Lexis’ new move is intended to prevent that error from happening, essentially “quarantining” cases suspected of being hallucinations from being linked. Surely Westlaw will soon follow suit in a similar way.

So how many cases are captured by each database as having discussed or cited to what’s suspected to be a GenAI hallucination? As of June 5, 2026, Lexis has 889 cases, and counting, with the older, still searchable, disclaimer. Of those, 673 are federal and 216 are state or territorial. Westlaw has 797 cases, and counting, with their disclaimer. Of those, 645 are federal and 152 are state or territorial.

Westlaw has even added a new entry to its more-than-century-old Topic and Key Number System: 46H Attorneys and Legal Services k789 Use of artificial intelligence (46Hk789 if you are hip), with 54 headnotes already so classified.  

I’ve been spending a little time looking at these cases to try to get a better understanding of when lawyers might be more apt to misuse GenAI, and in what kinds of circumstances. As usual when it comes to attorney misconduct, the story is often sad, filled with bad decisions compounding, with a few cases of unremorseful willfulness sprinkled in. I’ve read family law cases with AI hallucinations in them, defamation cases, and commercial disputes. And my two big impressions are that the problem we’re seeing is widespread and does not discriminate, and, that, somewhat chillingly, we are only seeing the instances when someone got caught, and their getting caught was reported by the court.

I don’t know what to do with this information. To me, it does not reflect any kind of unknown failure in either gen AI systems or in attorney misconduct–we know the risks, and these are them. As librarians, we need to continue teaching concepts like AI literacy and professional responsibility in our research courses, and we need to continue keeping up to date with changes in technology and its application in our field.

For one small mercy, and a sign of success, we can return to the point we started. Mata v. Avianca, the “patient zero” hallucination case, has been cited by other courts 289 times according to Shepard’s on Lexis and 265 times according to KeyCite on Westlaw. The fake case “Varghese v. China Southern Airlines Co.,” cited by Judge Castel way back in June 2023, has only been cited by one other court, in a case disciplining an attorney and describing the danger of hallucinated cases, Fletcher v. Experian Information Solutions, Inc., 168 F.4th 231, 233 (5th Cir. 2026). And no, it is not hyperlinked on either Lexis or Westlaw.