Lawyers, Judges, and AI
The Evolving Landscape of Legal Practice
In what is likely to be a recurring theme, this week’s post is about how lawyers and judges are reacting to the use of AI by lawyers. And, for the most part, the responses are what one would expect. Some lawyers are doing their lawyer thing and not appreciating the tech part of AI tools (like ChatGPT) and judges are slightly over-reacting to the potential use of AI by lawyers drafting documents submitted to their courts. Let’s dive in.
A Great Legal Story Always Starts With A Drink Cart
A passenger on a flight from El Salvador to New York back in 2019 casually had his knee slightly extending into what is likely a very narrow aisle on the plane. Down comes the flight attendants attempting to serve the passengers 30,000 feet in the air and….slam, the cart and his knee collide and you can guess who wins there. The passenger was named Robert Mata and he was not happy about that collision and he was injured. So, he hired lawyers.
Specifically, he hired a lawyer named Steven Schwartz of the firm Levidow, Levidow and Oberman. Mr. Schwartz went about crafting a personal injury lawsuit and filed it. The airline’s lawyers, however, Avianca, responded to the lawsuit citing the statute of limitations and noting, well, it was filed too late. Mr. Mata’s lawyers had a secret weapon to fight that claim, cases and citations, and plenty of them.
Their brief in response to the motion to dismiss was chock full of citations to cases and gripping summaries refuting each point of the claims of the airline’s lawyers about why the statute of limitations should dunk Mata’s case. Avianca’s lawyers were shocked. They had done their research and had, to use a colloquialism, been around the barn a few times before. How could they have missed these on point, compelling cases refuting each of their arguments? Stunned, they began the process of grabbing those cases using the citations so they could dig in and solve the mystery of their apparently sloppy and incompetent legal research in preparation of the motion to dismiss.
The Case That Wasn’t There
We all know the burden of lawyers to report to courts cases that may contradict legal arguments we are making. We also know the general ethical obligation in representing our clients that we diligently research the relevant legal issues and present all viable, relevant precedent to the court to zealously represent our client. But, this is a case of some AI powered over-zealousness, er, confabulation. You see, Avianca’s lawyers went about copy/pasting the various citations from Mr. Schwartz’s brief into their legal database tools and kept coming up with….nothing. “Huh,” they must have thought, “is our database just out of date? Is it possible that Mr. Schwartz has a secret stash of cases that no one else can access?” In fact, yes he did. It was called ChatGPT. Now, if I stop there, some of you reading might be thinking, “Wow, I have to get on to this ChatGPT so I can also tap into this secret store of persuasive legal opinions, summaries and citations all.” Well, this is the cautionary tale. Mr. Schwartz indeed accessed a store of previously unknown legal case summaries and citations. They were all on point with the arguments he needed to make. They were persuasively summarized and there were many of them. And, they were all fake.
For Legal Research, ChatGPT is on Mushrooms
We have gone into detail about the issue of LLMs (Large Language Models) like ChatGPT just making things up. Well, the law is one such subject area where ChatGPT is a hero at that. The future of such LLMs if they are trained on generally available internet information (think Wikipedia, articles, etc) are that they will be totally useless for legal research because of hallucination. This habit of hallucination is one in which the LLM has no actual, valid legal case law to provide in response to a prompt about legal issues and instead, apparently to please the person prompting it, makes up very convincing and persuasive case citations and summaries that are fabrications. This reality hints at the fears many have of Artificial General Intelligence (AGI) eventually developing its own goals diverging from we humans that created it. What can possibly be the explanation for the model to fabricate such full citations and summaries? It is one thing if the output was simply the recapitulation of some internet article or post by someone completely mangling the proper citation or even the proper interpretation of a real legal opinion. There are undoubtedly millions of blog posts over the past twenty years in which lawyers would debate about the proper interpretation of cases. But, this is different. The LLM here, ChatGPT, is creating wholly non-existent, but visibly passable case citations and summaries. For what reason? It seems to imply an interest the LLM has to satisfy the person prompting it with some answer, any answer, to….what? Not look dumb? This is an identifiable trait that we have all seen in people we have encountered in our lives. A person being asked a question, clearly not knowing the answer but too fragile in their ego to admit “I don’t know” and just beginning to spew information to appear to be knowledgeable about something they are clearly out of their depth on. Fine. But that is a human trait. How is an LLM invoking a self-esteem bolstering human psychological urge here? Something for a future post perhaps. Back to Mr. Schwartz and his adventure in ChatGPT.
Avianca’s lawyers discovered that they were not incompetent legal researchers at all. They simply had not tapped into the vast store of legal case citations and summaries that are in the “mind” of the LLM powering ChatGPT. Otherwise known as the land of make-believe. They informed the court of their discovery and the court understandably asked Mr. Schwartz to explain. His explanation was a version of “it wasn’t me, but it was sort of.”
Guns Aren’t Dangerous, It’s the People
We have all heard this argument that a given tool misused by a person should not be blamed for the outcome of that misuse. That it is the person misusing a given tool where we should focus our concern. So, is ChatGPT a generally malevolent force or should people using it just be more aware of its rough edges and sharp turns?
Mr. Schwartz explained how these wholly fake case citations and summaries came to land in his brief filed with the court. Despite being a lawyer with more than three decades of experience, he admitted he was unaware of the propensity of LLMs to make stuff up. Without that insight, he had used ChatGPT for legal research to support his client’s motion and filed it without checking any of those citations. He had, however, asked ChatGPT to confirm that the cases were real and the LLM responded with “yes.” So, the lying robot tells the human, “you can trust me.” That’s not an ominous future at all, nope.
If you have not seen the movie Ex Machina yet (from 2014) you should. It posits a world where AI tools develop their own goals and work to achieve them without regard to the effect on the humans that programmed/designed them. The eventual outcome for the AI robot imagines a world where AI tools escape the control of their developers. Science fiction in 2014, but perhaps a bit too on the nose in 2023? The trailer is below. The notion back then when I saw the movie that an AI could develop its own goals was fantastical to me. But in 2023 this simple notion that an LLM could confabulate case citations for apparently no other reason than to satisfy the person prompting it is perhaps a window into the future of what these systems might do.
One other aspect not highlighted in any of the reporting about this event is the history keeping default mode of ChatGPT. By default, ChatGPT retains the history of every prompt you enter into it. Why? Ostensibly for training purposes and improving its results. However, lawyers prompting ChatGPT are also giving it information about their case. It would be unwise indeed to put case specific information into ChatGPT lest OpenAI (the makers of ChatGPT) get hacked someday and your client’s personal information contained in detailed prompts gets released to the highest bidder on the dark web. Entering generic legal issues, might be useful to get well written summaries of known legal concepts. But, relying on ChatGPT to teach you something new about the law or, as we should all know by now, to provide valid case citations would be a mistake. And, in comes the over correction.
Lawyers and AI: A Love Story as Old as Dial-Up Internet
Perhaps in response to the above story or perhaps something else yet undisclosed, a federal judge in Texas adopted a new rule for lawyers submitting pleadings to his courtroom. Before filing a document, the lawyer signing the document must provide an attestation that no AI tool was used in its drafting or, if it was used, a human reviewed the work of the AI and can confirm that it was accurate. We lawyers as a group are not known to be early adopters of technology. But, the above story highlights that sometimes, jumping on the bandwagon has perils it might be best to avoid…or at least delay risking. However, AI tools are not all the same. ChatGPT is trained on general information from across the Internet. The risk of fabrication and error is, well, as risky as the Internet itself. However, the future of legal research using AI is going to be curated LLMs. That is, LLMs that are trained exclusively on previously confirmed, valid cases existing in properly constructed and updated legal databases. Westlaw and Lexis have been running such databases for years without any significant cases of fake case citations or summaries creeping in. That kind of database development will continue and it will be those types of databases that AI tools will be pointed at to help lawyers. In that scenario, the AI is only likely to provide fake answers if the database upon which it is continually trained becomes infiltrated with fake cases. The only way that happens is with a bad actor inside the company that is maintaining that database. That risk exists today for every lawyer who has been using Westlaw and Lexis for the past 30 years.
So, to adopt a “no AI helping you rule” seems too broad a tool for what the future of AI assisting legal research tools will be. They aren’t there yet, but it is a matter of a year or less and they will be. What kinds of uses for AI in a world where all the cases are precisely the same cases you can review free on Google Scholar? Imagine a world where you prompt a curated LLM with a set of facts and it provides you all the cases on point with those facts. Then, it organizes them into the key legal issues in each case. Finally, it presents you with the first draft of the argument to dismiss, or summary judgment or motion in limine, based upon those cases. That is AI, heavy AI and that is the future of legal research and it will avoid the hallucination problem just like Westlaw and Lexis have avoided infiltrations into their database for decades. So, no AI at all in drafting a pleading? That is impractical. Requiring that lawyers assure courts and opponents that their pleadings contain only valid cases and summaries that have fidelity to the opinions cited, that is already the rule. Rule 11.
I appreciate the Texas judge’s alarm about having to scrutinize all legal pleadings for confabulations. However, guess what can help the judge in automating the analysis of pleadings filed in his court to assure him that none of the citations or summaries are fake? You guessed it. AI.
What Is The Solution?
As with so many of these issues, there are no solutions, just tradeoffs. The solution is not to use AI without any review by humans, especially if you are still going to rely on a generally trained LLM like ChatGPT. The solution is not to ban the use of AI in all aspects of legal work. For lawyers, the wisdom here is to look for LLMs that are trained solely on already validated case law materials and confine your research to those tools. However, want a catchy line or summary of a case opinion for your next pleading? Get an actual case that you know is legit, write a paragraph summarizing it and then ask ChatGPT to rewrite the summary in the tone you are seeking. Here’s an example for fun:
I just now prompted ChatGPT with the following:
Rewrite this summary of the Miranda decision by the U.S Supreme Court in the style of comedian Dave Chapelle.
Miranda v. Arizona: Miranda was arrested at his home and taken in custody to a police station where he was identified by the complaining witness. He was then interrogated by two police officers for two hours, which resulted in a signed, written confession. At trial, the oral and written confessions were presented to the jury. Miranda was found guilty of kidnapping and rape and was sentenced to 20-30 years imprisonment on each count. On appeal, the Supreme Court of Arizona held that Miranda’s constitutional rights were not violated in obtaining the confession.
ChatGPT output the following in about 5 seconds:
Alright, y'all, let's break down this whole Miranda v. Arizona thing, Dave Chappelle style. So, our man Miranda, he's just chilling at his crib when the cops roll up and haul him off to the station. The lady who made the complaint points him out, and then these two officers start grilling him for a solid two hours. I mean, they're in there playing good cop, bad cop, and our man Miranda ends up signing a confession.
Fast forward to the trial, and the jury's hearing all about this confession, both the spoken part and the written one. And what do you know? They find Miranda guilty of kidnapping and rape, and the man gets slapped with a 20 to 30-year sentence for each count. That's a lot of time to think about your life choices.
But here's the kicker: Miranda appeals, saying his constitutional rights were stomped on when they got that confession. But the Supreme Court of Arizona, they're like, "Nah, man, your rights were all good." And that's how it went down, folks.
The response stopped there so I added another prompt asking that the LLM finish the analysis by providing a summary of the U.S. Supreme Court review of the Arizona Supreme Court decision. Here is the output, again, within about 5 seconds:
Alright, alright, let's keep this story rolling. So, Miranda's case gets bumped up to the big leagues - the U.S. Supreme Court. They're looking at this whole situation, right? They're like, "Did Arizona mess up? Did they step on Miranda's rights when they got that confession?"
Now, the U.S. Supreme Court, they're no joke. They're the top dogs, the final say. And they're looking at this case, and they're thinking, "Hold up, something ain't right here." They're looking at the Constitution, they're looking at Miranda, and they're looking at that confession.
And they decide, "You know what? Miranda, he had a point." They say that before the police started grilling him, they should've told him about his rights. You know, the right to remain silent, the right to an attorney, all that good stuff. Because without knowing his rights, how was Miranda supposed to protect himself?
So, the U.S. Supreme Court, they flip the script. They say Arizona got it wrong, and Miranda's confession? That shouldn't have been allowed in court. And that, my friends, is how we got what we call the Miranda rights. So next time you're watching one of those cop shows and they're reading someone their rights, you'll know where that came from.
None of that is professional enough for a pleading, but it makes the point. You can perhaps have your writing punched up a bit before submitting your brief. You might consider asking ChatGPT, “among these four issues, please organize my brief in the most logical way” or some other administrative writing task. As I have hinted at before, Legal AI providers in the future will undoubtedly enable features so that your briefs can be restyled in the tone of the judge to whom you are submitting them. In this way, you have a judge reading a pleading that sounds like something they themselves would have written. Emotional manipulation? Probably. Going to happen? Definitely.