The Volokh Conspiracy
Mostly law professors | Sometimes contrarian | Often libertarian | Always independent
Apparent AI Hallucinations in Briefing From Both Parties
That's a new one on me, from a UK trademark appeal; the plaintiff, who was self-represented, admitted to using generative AI, and the defendant's lawyer was strongly suspected of so doing:
8. At the start of the hearing, I asked Dr Soufian if he had drafted the documents and he said he had drafted it with the assistance of Chat GPT. I pointed out the numerous errors in the citations and problems with the skeleton and he politely apologised and did so unreservedly. Before moving on, it is worth noting that most of the skeleton produced by Chat GPT was made up of arguments purportedly relating to the evidence in the case. However, the factual issues highlighted were largely not relevant to the issues before me and the proposed arguments were not very helpful. In other words, even aside from the fabricated citations, the output of Chat GPT was in fact unhelpful to him.
9. In the case of Mr Caddy, who is a trade mark attorney, his skeleton argument dated 6 June 2025 included three cases which existed and were correctly cited. But it was unclear to me the cases cited stand for the propositions claimed by Mr Caddy.
10. During the Hearing, I asked Mr Caddy to identify the part of the judgments which supported the propositions made. He said, "I cannot actually remember that now, to be honest with you". I gave him time to read the judgments so as to find the relevant paragraphs. He could not do so. He then said he got the references from a "previous edition of Kerly's Law on Trade Marks". I could not find any support for the propositions (or anything similar) in Kerly during the hearing. He then said he could not remember where he got them from, saying maybe it was Wadlow [on the Law of Passing Off] but said he went on to say that did not make much sense. After the hearing, I checked Wadlow and could not find anything matching the propositions.
11. A few hours after the hearing, Mr Caddy sent an email to the Secretariat, he said in the email that he had not been expecting to "make out my own side's case more so than had been done in the skeleton". He then went on to try and show support for the propositions in the three cases cited in the skeleton. In my view, nothing in the email improved Mr Caddy's position and the quotation above clearly makes it worse….
30. I cannot say how Mr Caddy came up with the propositions of law he put forward for the three cases. I emphasise once more that a fabricated citation is not just an entirely made up case, but also includes citing a case for a proposition where there is absolutely no basis in the case for that proposition to be made. As I have already mentioned, even after his clarification I struggle to see how the cases he cites support his propositions more than in the most general and abstract sense.
31. Three things about Mr Caddy's conduct concern me greatly. First, he did not know when asked in the hearing where he obtained the propositions of law he included in a skeleton argument, even though the document was dated less than a week before the hearing. Secondly, he appears to think that he is not expected to be ready to expand on points made in his own skeleton argument at the hearing. Thirdly, he appears to think it is acceptable to use out-of-date textbooks. It is necessary for all lawyers to have access to relevant and up-to-date textbooks and relevant case reports whether online or printed and whether within their own firm or using one of the law libraries available to them, which for those who are London-based (like Mr Caddy) include the Intellectual Property Reading Room at the British Library.
32. In the end, I have decided I will not refer Mr Caddy to IPREG on this occasion. I accept that it is possible that he found these references and simply read them (misunderstood them) in a way I cannot follow or understand. However, even if this is the case he needs to seriously reflect on how he conducts his practice in the future, including how he undertakes legal research, how he drafts skeleton arguments, and he must ensure that when he appears before a tribunal he is properly prepared to do so. Advocates should always be prepared to explain to a court or tribunal what they have included in skeleton or other written arguments.
Check out Damien Charlotin's AI Hallucination Cases database, which is now at 210 cases (more than half from the U.S.), as well as Peter Henderson's AI Law Tracker, now at 223 cases.
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
I've been running my briefs through ChatGPT for the last couple of weeks and it wants to do this to me all of the time. Even if I tell it to only give me real cases it gives me fake cases.
My practice is rather focused so when it cites the fake cases it gives rather outcome determinative statements that I immediately recognize is not a real case because it would make the issue I am arguing a slam dunk and I certainly would have heard of that case before.
It is almost like the AI "wants" to help so much that it just reaches out and makes up something so you can win.
For example, I was trying to argue that fornication had long been against my state's public policy and it came back with a fake case, State v. Sigley from 1897 with a pinpoint cite stating that the fornication law had been upheld against a constitutional attack. I knew immediately that was wrong because in 1897 nobody would have made such an argument---such an argument wouldn't have been thought of until at least the 1960s.
I asked it if that was a real case, and it immediately admitted that it wasn't. I wish the model could train itself not to do that.
But until then attorneys need to understand the limitations and even if the case seems plausible, just LOOK IT UP yourself and verify what the AI is telling you.
Could running draft legal briefs through ChatGPT or any AI tool compromise attorney client privilege?
I have not looked into it closely, but I'd think the Work Product privilege would be more applicable here anyway.
How would that make the disclosure of confidential material (and I'm considering drafts as confidential) to an AI tool any more advisable? ChatGPT cannot seriously be deemed a consulting agent of the attorney.
I don't see how it would. Anything in a legal brief is for consumption by the court and the other party. If it is something the client wants to keep secret, it's not in a brief.
Drafts? You would produce drafts without any claim to privilege if requested in discovery?
Riva, where does wvattorney13 reference draft briefs? The comment, "Anything in a legal brief is for consumption by the court and the other party" obviously refers to submissions to a court.
The initial comment related to material that had not been yet been submitted. Some may call that a draft. My question clearly related to drafts. I thought that was clear. Apparently not.
No. I don't see how running it through AI would be any different than running it through a Word spellchecker.
I'm just failing to see how using a computer tool would defeat a privilege or cause something not subject to discovery to now be subject to discovery.
Because putting confidential material through a third-party AI tool potentially compromises your confidential information both by the AI potentially retaining data and by uploading the information to servers that might not be secure. This is most definitely NOT the same thing as checking the document on whatever word processing programming you may be using on your private computer. What enforceable guarantees of confidentiality do you get from an AI tool? My guess is none. You'd be better off trusting your bartender.
This is correct. I've had to draft policies instructing people not to put sensitive info into chat gpt.
This might be an unexplored aspect to so many lawyers using the tool to write briefs, in addition to the legal inaccuracies.
The lawsuit is a sworn statement. Intentional misleading of the court should be punished like perjury.
That being said, all fictitious lawyer statements should be severely punished. Anywhere the word, constructive, is used, requires punishment. Whenever the plain language of the constitution is denied, the charge should be insurrection against the constitution. That is punishable by 10 years in fed stir. Arrest the Supreme Court Justices and put them in gen pop. That includes their allowing any executive regulation to be in effect without formal Congressional approval by vote. Include lying by omission, as when a prosecutor withholds exculpatory evidence. Prison term. We are sick of this lying garbage profession destroying our nation for their own enrichment.
"The lawsuit is a sworn statement. Intentional misleading of the court should be punished like perjury."
A verified complaint is a sworn statement. Unsworn statements in pleadings are not perjury.
It is unethical for a lawyer to intentionally mislead a court, not to mention ineffective advocacy. But absent an oath or unsworn declaration under penalty of perjury (see, 28 U.S.C. § 1746), it is not criminal.
It's not criminal, but it's a Rule 11 violation. And hence sanctionable.
In addition, being sloppy and failing to check your cites is not "intentionally" misleading the court. It is negligent or possibly reckless but not intentional.
I am wondering how an attorney can specialize in the public policy on fornication. Are there a lot of such cases? Do people pay you to write briefs on fornication?
I have written briefs challenging the constitutionality of laws prohibiting prostitution, based on the rights of privacy and substantive due process. There is a wide body of case law regarding fornication.
Honestly, there's enough terrible lawyering out there that I would buy that they would cite a case for a proposition that really isn't on point without the help of AI.
I think there's a few extra questions to make this determination:
Was there a pincite? If so, that suggests a lawyer who is expected to be diligent (a pincite to something completely irrelevant would suggest AI hallucination).
How far off are we talking about? Is it at least the same subject matter?
Ultimately, I don't think it makes a difference whether the lawyer is bad all on their own or with the help of AI, though.
I've definitely seen that kind of citation long before AI came into use. Your last sentence is the key -- the attorney is responsible for the cases he cites, however he got hold of them!
Back in the day, my favorite professor in law school (the reason I went all "evil law") said the following to the class-
If you want to pass this class, you have to learn the black-letter law. But that's the basic level of competence. Places like Cravath aren't going to pay you to regurgitate the black-letter law. They pay you to win even if the black-letter law is against you. So if you want to excel in this class, you need to understand the black-letter law, and then be able to argue why it doesn't apply.
I've seen attorneys using AI to review their briefs or "help" with their arguments, and I'm reminded of what that professor told me. See, the real problem isn't just the hallucinations. If you've advanced past the level of basic mouth-breathing, you can check the citations. No, the real problem is that you will consign yourself to a level of perfectly cromulent reasoning.
If you're a bad attorney (and you correct hallucinations), then AI can help you. But that's it- it raises the floor for you, but you're always going to have a very low ceiling.
From what I've seen so far. Who knows? Maybe in two years, or ten years, it will be different. But not today.
So if you want to excel in this class, you need to understand the black-letter law, and then be able to argue why it doesn't apply."
You need to do that credibly. I've seen lawyers make silly arguments in this vein. "That case is distinguishable because the party wore a blue shirt when he signed the contract, but mine was wearing a yellow shirt."
The lash should be applied to that scum bag Cravath lawyer and to that professor that mentored him.
That sounds like a pseudo-inspiring thing for a law professor to say but how far do you typically get when arguing against black letter law?
Sure, the judge and your colleagues might think you are a really clever lawyer, but when the client loses, because black letter law is against you, what does that really mean?
If you were saying that the mark of a good lawyer is to be able to distinguish seemingly adverse cases and argue why they don't apply to your factual situation, then we might be on the same page. But AI can help with that. It's just a tool not the master.
It wasn't a question of inspiring. It was the truth. Obviously, it helped that it was my BizOrgs and Corporations professor, and he had actually practiced for some time at Cravath (in addition to federal clerkships, etc.).
The point isn't some spurious song-and-dance "Red Car / Green Car" issue. It was about truly understanding the law- understanding how it all works together, understanding the differences between legal and equitable principles and how they interact, understanding when procedural arguments are effective, understanding when you can appeal to L&E reasoning (state incorporation of Delaware law, for example) and when it isn't (due process). About how to use counter-intuitive doctrines that your state might not have recognized yet that would apply. Or even something as simple as the correct argument as to why the trial court should apply (sub silento) the majority rule for a particular issue due to the factual circumstances instead of the minority rule currently used, but unexamined for fifty years.
Or, for that matter, the proper basis for an extension of modification of the law.
And I am perfectly comfortable with my assessment of AI technology (general AI) as it stands. I think that specialized AI tools have been useful for certain research tasks, but using AI makes bad attorneys, at best, perfectly cromulent. Good attorneys that use them lower their ceilings.
Good luck with what you are doing.
Places like Cravath aren't going to pay you to regurgitate the black-letter law. They pay you to win even if the black-letter law is against you."
That's true, but we are talking about a rarefied stratum of lawyers and cases there. Also, the smarter and better you are as a lawyer, especially as you gain experience in a field, the "blacker" the law gets (and the genuinely gray areas become smaller). Just look at Prof Kerr as a case in point.
And even then, lots of times it is not about "changing the law", but rather marshalling the facts in such a way as to make them appear to fit into the applicable law.
The corollary is the case (and there are a lot of these) where "good" lawyering is counterproductive to your client. One of my former employers used to have a lot of low dollar (sub $10k) claims against them for workers comp, wrongful term, discrimination etc. Given the exposure, it made no sense to do a thorough investigation followed by an in-depth and sophisticated legal analysis. Cromulence was the name of the game there.
"...perfectly cromulent reasoning..."
Somebody went to college.
Seems like Mr Caddy doesn't understand that the coverup is worse than the crime. I wonder how long he's been a lawyer.
Exhibit 2053 that AI is garbage. (For those old enough to remember, GIGO.)
AI is garbage. But it's rich garbage. By its mechanisms and some luck, it can aid in the development of intelligent ideas even if it's not capable (strictly speaking) of producing intelligent ideas.
I'm a curmudgeon to the subject. I am for the most part retired and not concerned about my productivity. But if I were concerned about my productivity, I'd be looking at how AI might be able to help me to up my game.
I'd be looking to prove you wrong, even though almost every emotional and logical fiber in my being says GIGO, GIGO, GIGO.
We need to completely shutdown the practice of law until we figure out what is going wrong.
The lawyer profession is not putting out a defective product. It is putting out a catastrophically dangerous product. Recall it until fixed. Every self-stated goal of the 200 subjects of the law is in utter failure. It is totally toxic garbage. The direct damage done by the lawyer profession is double that of organized crime. The indirect damage is 10 times more damaging.
[misplaced]
SPOILER ALERT
One of my guiding axioms is "don't make the perfect the enemy of the good".
I am convinced AI is here to stay, improving at a frightening rate, and truth be told is one of the better areas to invest in. In fact here is what the Gemini AI engine has to say about how Westlaw is using AI. Something that produces better results than cheaper alternatives
Westlaw, as a leading legal research platform, has been integrating AI for years, but the rise of generative AI has accelerated its efforts. Thomson Reuters, the company behind Westlaw, has made significant investments to embed AI capabilities across its products, moving beyond simple keyword searches to more sophisticated, conversational, and predictive tools.
Here's a breakdown of how Westlaw integrates AI:
1. Generative AI Tools
This is the most recent and transformative type of AI integration, often using large language models (LLMs) to create new content based on a user's prompt.
AI-Assisted Research: This feature, available in Westlaw Precision, allows users to ask complex legal questions in natural, conversational language. The AI provides a narrative answer with direct citations and links to the underlying legal sources, such as cases and statutes, ensuring the response is "grounded" in reliable data.
CoCounsel: Thomson Reuters acquired CoCounsel, a legal AI assistant, and is integrating its capabilities into the Westlaw platform. This tool can perform tasks like:
Summarizing large documents.
Drafting legal documents (e.g., deposition questions, client emails).
Creating timelines of events from multiple documents.
AI Jurisdictional Surveys: This tool uses AI to quickly analyze and compare statutes and regulations on a specific topic across multiple jurisdictions, which is a traditionally time-consuming task.
Claims Explorer: A generative AI tool that helps identify potential legal claims based on a user's description of a fact pattern. It can suggest relevant claims and provide links to supporting legal authority.
2. Extractive and Predictive AI
Before the age of generative AI, Westlaw used other forms of AI and machine learning to enhance its search and analysis tools. These are still critical to the platform's functionality.
WestSearch Plus: This AI-powered search engine uses natural language processing to understand a user's query and provide more relevant results. It goes beyond simple keyword matching to identify the core legal concepts in the user's request.
Quick Check: This document analysis tool uses AI to review a legal brief or memorandum. It can identify relevant cases the author may have missed and flag any cited authority that may be "bad law" (e.g., overruled or negatively treated). This helps lawyers ensure their arguments are well-supported.
Litigation Analytics: This feature uses AI to analyze massive amounts of litigation data. It provides data-driven insights into judges, courts, attorneys, and law firms, allowing lawyers to make more strategic decisions about their cases. For example, it can help predict the likely outcomes of a motion or the typical length of a case before a specific judge.
The Foundation of Westlaw's AI
A key advantage for Westlaw's AI is its extensive and meticulously curated legal database. Unlike general-purpose LLMs, Westlaw's AI tools are trained on and access a "closed universe" of high-quality legal content. This includes over a century of human editorial work, such as the famous West Key Number System, which helps ensure the AI's outputs are accurate and verifiable. Westlaw uses a technology called Retrieval Augmented Generation (RAG), which combines the power of an LLM with its own proprietary database to provide accurate, up-to-date, and well-cited answers. This process is designed to reduce the risk of "hallucinations" (producing fake or incorrect information) that are a common concern with public AI models.
ChatGPT writes credible briefs and lawsuits with real cases for $20 a month. Give it the talking points. It does so in 30 seconds. It will then provide the opposing arguments. Then rebut these arguments ahead of time in a revised brief or lawsuit. Bye, lawyer wankers. It cites real Rules of Procedure, and verifiable, real cases.
I retired back when but was always happy with what Westlaw produced pre-AI. Given the price I no longer have access but am wondering if anyone (including EV) has ever used the current version and if so what impression did it leave. Asking for a friend.