Lawyers continue to file hallucinated cases, which is, frankly, kind of an accomplishment at this point. Legal tools exist that guard against such hallucinations and even more tools exist to double check to make sure no hallucinations slipped in there. And then, of course, there’s being an actual lawyer and proofreading your work before filing it.
Even though the blame rightly falls on the attorneys failing to check their cites, there’s still room for other actors to help save these lawyers from themselves.
During last week’s Legaltech Week roundtable, a February blog post by law librarian Shay Elbaum was brought to my attention that deserves some more attention. The post discusses a point raised in Flycatcher Corp. v. Affable Avenue. It’s an otherwise bog standard AI hallucination sanction case, except the lawyer involved made the argument that when a generative AI tool invents a citation in a proprietary format, it makes it difficult to verify it without access to the proprietary source.
In other words, when the AI spits out “2022 WL 4637582,” or “2024 LEXIS 2847243,” there are lawyers out there who can’t easily check that. As Elbaum notes, “not only is it difficult for you to read the case, but it’s also difficult for you to even be sure it’s a real case.” The attorney in this case, as it happens, did have access to Westlaw, so this argument did not go over well. But it’s nonetheless an interesting wrinkle.
Lawyers shouldn’t include cases they can’t verify, but consider that we now have judges sanctioning lawyers for simply not catching the other side’s mistakes. A cite in an opposing brief that the lawyer can’t verify leaves those who did nothing wrong vulnerable just because they aren’t subscribed to the big databases.
On a quick review of the latest fabricated citations in Damien Charlotin’s database, most of the recent filings with fabricated citations included at least one fake citation with the format “2022 WL 4637582,” supposedly pointing to an unpublished case available on Westlaw. (I’ll call these “WL citations” in the rest of this post.) I didn’t find any hallucinated citations with the equivalent Lexis format for unpublished cases, but my review was far from thorough and I assume they exist.
This isn’t a knock on Westlaw or Lexis, but their proprietary formats make them more vulnerable to ending up in embarrassing hallucination stories. When a brief from opposing counsel cites a phantom F.3d case, anyone with a free database can catch it in 30 seconds. When the same brief cites a phantom WL case, the lawyer on the other side — the solo, the legal aid attorney, the self-represented litigant who couldn’t swing a Thomson Reuters subscription if their docket depended on it — has no move.
Presumably both major legal publishers would say, “they could buy a subscription,” so let’s agree to roll our eyes and then move on to some entirely obvious solutions that have nonetheless not been implemented:
If Westlaw doesn’t want the WL citations to become red flags for potential hallucinations, it needs to provide a method for everyone to verify these citations. Lexis, too; I’m just picking on Westlaw more because their database-specific format is hallucinated more frequently. And courts can act, too, without even needing to directly regulate AI use. Federal Rule of Appellate Procedure 32.1(b) already requires attorneys citing unpublished cases to provide copies of the cases with their filings. Adding a similar requirement to the Federal Rules of Civil Procedure and their state equivalents, or to local rules and standing orders, would even the playing field and reduce hallucinations in filings.
I’d go further than just a checker and say they should allow anyone to read a raw unpublished decision if the user enters the precise citation they’re looking up. It doesn’t give away the data moat to let people see the case based on entering a precise citation. The real value of these databases is in researching concepts, not calling up specific documents. Strip the result of key cites if necessary… but just give them access to the opinion.
There may have been a day where Thomson Reuters and LexisNexis made money off gatekeeping unpublished opinions, but that’s not the cash cow in a world where they’re selling state-of-the-art AI research tools. It’s not worth becoming the common link in a bunch of hallucination cases just to jealously guard some Eastern District of Missouri ruling on a random motion in limine. A free public lookup would fix that and what the publishers lose is basically a rounding error.
As for rules requiring lawyers to attach copies of unpublished opinions, there are a few districts doing that, but it should be universal. The elegance is that a rule forcing lawyers to attach copies of the case would force them to confirm the case is real in the first place. Maybe if everyone did that with all their cases, we wouldn’t be in this mess. And, maybe more to the point, the rule must be universally enforced. Because Elbaum notes that the Middle District of Pennsylvania has such a rule now and:
I took a peek at recent filings in the Middle District of Pennsylvania and found plenty of citations to unpublished opinions, but no attached copies.
Swing and a miss. But that’s why we need a belt and suspenders approach here.
For two years we’ve treated hallucinations as the personal failings the lawyers. Because they are. But one of these days, those mistakes are going to work their way into the system in the form of rubberstamped rulings that get repeated as precedent. So there’s an obligation for all of us to save these folks from themselves before they make everything worse for everybody.
Joe Patrice is a senior editor at Above the Law and co-host of Thinking Like A Lawyer. Feel free to email any tips, questions, or comments. Follow him on Twitter or Bluesky if you’re interested in law, politics, and a healthy dose of college sports news.
The post Those ‘WL’ Citations Are An Open Invitation To AI Hallucinations appeared first on Above the Law.
Lawyers continue to file hallucinated cases, which is, frankly, kind of an accomplishment at this point. Legal tools exist that guard against such hallucinations and even more tools exist to double check to make sure no hallucinations slipped in there. And then, of course, there’s being an actual lawyer and proofreading your work before filing it.
Even though the blame rightly falls on the attorneys failing to check their cites, there’s still room for other actors to help save these lawyers from themselves.
During last week’s Legaltech Week roundtable, a February blog post by law librarian Shay Elbaum was brought to my attention that deserves some more attention. The post discusses a point raised in Flycatcher Corp. v. Affable Avenue. It’s an otherwise bog standard AI hallucination sanction case, except the lawyer involved made the argument that when a generative AI tool invents a citation in a proprietary format, it makes it difficult to verify it without access to the proprietary source.
In other words, when the AI spits out “2022 WL 4637582,” or “2024 LEXIS 2847243,” there are lawyers out there who can’t easily check that. As Elbaum notes, “not only is it difficult for you to read the case, but it’s also difficult for you to even be sure it’s a real case.” The attorney in this case, as it happens, did have access to Westlaw, so this argument did not go over well. But it’s nonetheless an interesting wrinkle.
Lawyers shouldn’t include cases they can’t verify, but consider that we now have judges sanctioning lawyers for simply not catching the other side’s mistakes. A cite in an opposing brief that the lawyer can’t verify leaves those who did nothing wrong vulnerable just because they aren’t subscribed to the big databases.
On a quick review of the latest fabricated citations in Damien Charlotin’s database, most of the recent filings with fabricated citations included at least one fake citation with the format “2022 WL 4637582,” supposedly pointing to an unpublished case available on Westlaw. (I’ll call these “WL citations” in the rest of this post.) I didn’t find any hallucinated citations with the equivalent Lexis format for unpublished cases, but my review was far from thorough and I assume they exist.
This isn’t a knock on Westlaw or Lexis, but their proprietary formats make them more vulnerable to ending up in embarrassing hallucination stories. When a brief from opposing counsel cites a phantom F.3d case, anyone with a free database can catch it in 30 seconds. When the same brief cites a phantom WL case, the lawyer on the other side — the solo, the legal aid attorney, the self-represented litigant who couldn’t swing a Thomson Reuters subscription if their docket depended on it — has no move.
Presumably both major legal publishers would say, “they could buy a subscription,” so let’s agree to roll our eyes and then move on to some entirely obvious solutions that have nonetheless not been implemented:
If Westlaw doesn’t want the WL citations to become red flags for potential hallucinations, it needs to provide a method for everyone to verify these citations. Lexis, too; I’m just picking on Westlaw more because their database-specific format is hallucinated more frequently. And courts can act, too, without even needing to directly regulate AI use. Federal Rule of Appellate Procedure 32.1(b) already requires attorneys citing unpublished cases to provide copies of the cases with their filings. Adding a similar requirement to the Federal Rules of Civil Procedure and their state equivalents, or to local rules and standing orders, would even the playing field and reduce hallucinations in filings.
I’d go further than just a checker and say they should allow anyone to read a raw unpublished decision if the user enters the precise citation they’re looking up. It doesn’t give away the data moat to let people see the case based on entering a precise citation. The real value of these databases is in researching concepts, not calling up specific documents. Strip the result of key cites if necessary… but just give them access to the opinion.
There may have been a day where Thomson Reuters and LexisNexis made money off gatekeeping unpublished opinions, but that’s not the cash cow in a world where they’re selling state-of-the-art AI research tools. It’s not worth becoming the common link in a bunch of hallucination cases just to jealously guard some Eastern District of Missouri ruling on a random motion in limine. A free public lookup would fix that and what the publishers lose is basically a rounding error.
As for rules requiring lawyers to attach copies of unpublished opinions, there are a few districts doing that, but it should be universal. The elegance is that a rule forcing lawyers to attach copies of the case would force them to confirm the case is real in the first place. Maybe if everyone did that with all their cases, we wouldn’t be in this mess. And, maybe more to the point, the rule must be universally enforced. Because Elbaum notes that the Middle District of Pennsylvania has such a rule now and:
I took a peek at recent filings in the Middle District of Pennsylvania and found plenty of citations to unpublished opinions, but no attached copies.
Swing and a miss. But that’s why we need a belt and suspenders approach here.
For two years we’ve treated hallucinations as the personal failings the lawyers. Because they are. But one of these days, those mistakes are going to work their way into the system in the form of rubberstamped rulings that get repeated as precedent. So there’s an obligation for all of us to save these folks from themselves before they make everything worse for everybody.
Joe Patrice is a senior editor at Above the Law and co-host of Thinking Like A Lawyer. Feel free to email any tips, questions, or comments. Follow him on Twitter or Bluesky if you’re interested in law, politics, and a healthy dose of college sports news.
