Latest / CLEs You Actually Want to Hear / GenAI Case Law: A Survey of 50 Judicial Opinions Addressing Use (and Misuse) of GenAI in Court Documents with Professor Heidi K. Brown
Transcript
- 0:03Welcome to the Gen .A .I. case law, a survey
- 0:06of 50 judicial opinions addressing the use and
- 0:09misuse of Gen .A .I. in court documents. I'm
- 0:12Rume Damran, chair of the Beverly Hills Bar Association's
- 0:16barrister section, and I'm delighted to introduce
- 0:18today's speaker, Professor Heidi Brown. Professor
- 0:22Brown is a graduate of the University of Virginia
- 0:24School of Law, associate dean for upper level
- 0:27writing at New York Law School, and a former
- 0:30litigator in the construction industry. She earned
- 0:33her master's degree in applied positive psychology
- 0:36from the University of Pennsylvania in 2021.
- 0:39Having grappled with extreme performance anxiety
- 0:42as a law student, attorney, and new law professor,
- 0:45Heidi ultimately untangled her fears and embraced
- 0:49authenticity as a powerful asset in teaching
- 0:52and practicing law. Professor Brown has more
- 0:54than two decades of experience in practicing
- 0:57law, and she's worked as a counsel with Moran
- 0:59Lee LLP, a boutique litigation firm with the
- 1:02offices in the DC area and New York City, where
- 1:06she handled all aspects of litigation and arbitration
- 1:09of complex construction contract and building
- 1:12development disputes. She specialized in brief
- 1:15writing for state and federal cases throughout
- 1:17the United States and has mentored junior associates
- 1:20in civil procedure, litigation strategy and legal
- 1:23writing. And now let's get started with the program.
- 1:26Thank you so much for having me. Greetings from
- 1:29the East Coast. So I have a ton of material today,
- 1:33so I'm going to jump right in. But just, yes,
- 1:35I am a law professor now. I teach legal writing,
- 1:38and I am an admitted, very late adopter of technology
- 1:43usually. But when ChatGPT sort of rolled into
- 1:46my legal writing classroom uninvited, and it
- 1:49was really like January of 2023, I thought, oh
- 1:52no. I need to get up to speed on this technology.
- 1:56So I've kind of thrown myself all in to really
- 1:59embracing the technology, but the kind of slowly
- 2:01introducing myself to it, letting it introduce
- 2:04itself to me. But at the same time, trying to
- 2:07stay on top of how these tools are really being
- 2:10used in our profession, sometimes really well,
- 2:13and then sometimes to a great detriment. So I'm
- 2:17going to. kind of cover the case law that's been
- 2:20happening. And as you can tell from my prior
- 2:22slide, I changed the title of this literally
- 2:25today, because when I first submitted the title
- 2:28for this, I had only counted up to 50 cases.
- 2:31And now we are up to, by my count, 64. And I
- 2:35don't even think I've caught them all. So today
- 2:37I'm going to focus on, since this is the Beverly
- 2:39Hills Bar Association, I'm going to focus on
- 2:42California state bar rules and judges rules regarding
- 2:46generative AI just really briefly. And then I'm
- 2:49going to dive into the case law. I'm going to
- 2:51start with the positive because that's how I
- 2:53roll. And I'm going to talk about how one judge,
- 2:55at least one judge, has been very transparently
- 2:58using AI in his own practice. I'll explain that
- 3:01in a bit to interpret language. That's how he's
- 3:04using it. And then we're going to dive into the
- 3:06case law involving lawyers and pro se litigants
- 3:09and some experts using and actually misusing
- 3:14generative AI and not checking their sites before
- 3:16they submit documents to the court. So I want
- 3:19to flag all of those for us today. Then because
- 3:22again, we want to focus on positive, what can
- 3:24we learn out of this? I'll talk about ethical
- 3:27considerations going forward, our need to get
- 3:30up to speed on these tools so we can avoid those
- 3:32same pitfalls, and then how we might adjust our
- 3:35workflow. Let's dive right in. The premise for
- 3:38this is to invite us to have a lifelong learning
- 3:42mindset, and we're required to. Comment 8 to
- 3:46the ABA rule of competence says that members
- 3:49of the legal profession have a to quote, keep
- 3:52abreast of changes in the law and its practice,
- 3:55including both the benefits and the risks associated
- 3:58with relevant technology. So we kind of don't
- 4:00have the luxury of sticking our head in the sand
- 4:02regarding generative AI, even though it might
- 4:05feel awkward or unfamiliar to a lot of us. California's
- 4:09rules are similar. Comment one to rule 1 .1 of
- 4:12the competence rule in California says the same
- 4:14thing. We got to get up to speed on the benefits
- 4:17and risks of relevant technology. And then the
- 4:20state actually the committee on professional
- 4:22responsibility and conduct on regulation of use
- 4:25of generative AI. Well, actually those recommendations
- 4:29came out from that committee on regulation of
- 4:32use of generative AI. California was one of the
- 4:36early issuers of state bar rules. Then I just
- 4:41checked, if you're curious about what judges
- 4:44are doing across the country, there are two great
- 4:47trackers. One is on Lexis and one is on Bloomberg.
- 4:50I double -checked the Lexis tracker this morning
- 4:53to make sure I didn't miss any. There's one superior
- 4:56court judge in California who has issued standing
- 4:59rules regarding generative AI. And then nine
- 5:03federal judges have issued standing orders regarding
- 5:06generative AI. Now, interestingly, you can check
- 5:10the jurisdictions across the country. Most judges
- 5:12are not saying don't use it. They're saying be
- 5:15careful when you use it. And if you do use it,
- 5:18most judges, at least all the standing orders
- 5:20that I've seen, are requiring us to attach a
- 5:23certification to our filings with the court.
- 5:27to disclose that we used it, how we used it,
- 5:30but then also that we have checked every single
- 5:32citation to make sure that the tool didn't accidentally
- 5:37hallucinate law. These are the concerns that
- 5:43the ethics opinions across the country and judges
- 5:45have raised. Obviously, the term hallucination
- 5:49is the tech term that the AI gurus have dubbed
- 5:53the things that these tools create out of thin
- 5:56air, but they do it so assertively that we think
- 6:00that these results are real. We got to be careful.
- 6:04about that. Of course, other judges are concerned
- 6:07about us disclosing confidential information
- 6:10to publicly available tools. There's all kinds
- 6:13of discussion about disclosing to our clients
- 6:15that we're using them. Lots of literature going
- 6:18back and forth on whether this will change our
- 6:20billing practices, perhaps if it does make us
- 6:22more efficient. I can tell you having incorporated
- 6:25these tools into my legal writing classes, so
- 6:28far they are not making me or my students more
- 6:33efficient because because we are adding extra
- 6:35time to check all the tools. But there are discussions
- 6:38about maybe eventually we'll be more efficient.
- 6:40And then you can see on the screen other issues
- 6:42that are being raised in these ethics opinions.
- 6:45Our duty to supervise, our more junior employees
- 6:47using these tools, checking for output bias,
- 6:51bias in the output, because these tools have
- 6:53been trained on trillions of words written by
- 6:57humans, which unfortunately also have bias baked
- 7:00into language. And so we need to check and make
- 7:02sure that we're just not using output that has
- 7:07those same concerns. Then my students keep raising
- 7:10the issue of environmental concerns because these
- 7:12tools do require the use of a ton of energy and
- 7:16interestingly, even water to cool the computer
- 7:19power. But let's start with the positive use
- 7:23cases. It's been really interesting hearing about
- 7:27how some judges and then some law professors
- 7:30have been experimenting with this too. have been
- 7:33testing out whether AI can assist in predicting
- 7:37and resolving ambiguities in contract and language
- 7:40interpretation. So I wanted to flag for you two
- 7:44really interesting cases that came out last year
- 7:46in 2024 out of the 11th Circuit, so the federal
- 7:51court, 11th Circuit Court of Appeals. So Judge
- 7:55Kevin Newsom, he wrote two concurring opinions,
- 7:59one in May 24 and one in September of 24. And
- 8:03I just found these really interesting. So both
- 8:05of these cases were dealing with the interpretation
- 8:09of the plain language or plain meaning or the
- 8:12ordinary meaning of words. So the first case
- 8:15was an insurance dispute. And this was Snell
- 8:19versus United Specialty Insurance Company. And
- 8:21in this dispute, there was an insurance coverage
- 8:24question of whether an in -ground trampoline
- 8:27and a retaining wall that an insured party had
- 8:30installed in his yard constituted landscaping
- 8:35in the context of the insurance policy. The majority
- 8:39court ended up rendering its decision, I think
- 8:41actually based on the insurance application and
- 8:44I believe denied coverage. But Judge Kevin Newsom
- 8:48or upheld the denial of coverage. Judge Kevin
- 8:52Newsom though wrote a really long, I think it's
- 8:54like 26 page concurring opinion. very transparently
- 8:59discussing how he got curious about, you know,
- 9:02in addition to what judges normally do to interpret
- 9:05words or determine the ordinary meaning of words
- 9:08that are in dispute in a lawsuit, judges consult
- 9:12dictionaries, you know, they go through the canons
- 9:14of of interpretation, but he thought since Since
- 9:18generative AI tools like ChatGPT and Anthropics
- 9:23version Claude and Google Gemini, since these
- 9:26things are called large language models, they
- 9:29actually process language and predict the next
- 9:33most logical word and sentences, etc. I won't
- 9:36go into all the technical mumbo -jumbo now because
- 9:39I will mess it up. But he decided to conduct
- 9:42an experiment to see if He asked these generative
- 9:45AI tools the same question, what is the ordinary
- 9:48meaning of the term landscaping, if it would
- 9:50jive with his judicial interpretation. And then
- 9:55he did the same thing a couple of months later
- 9:57in a criminal sentencing enhancement case, where
- 10:02the court had to analyze the ordinary meaning
- 10:04of a multi -word phrase physically restrained
- 10:08in the context of an armed robbery sentence enhancement.
- 10:13go to kind of what Judge Newsom said about his
- 10:16insights in asking large language models like
- 10:19chat GPT, Google Gemini, and anthropics Claude,
- 10:23the ordinary meaning of these particular terms
- 10:25that were at issue in these cases. He wrote in
- 10:28both of those concurring opinions that large
- 10:31language models are trained on vast amounts of
- 10:33data, words, that aim to reflect and capture
- 10:37how we as individuals use language in our everyday
- 10:40lives. He definitely acknowledged all the potential
- 10:43upsides of AI as well as the hazards of these
- 10:46tools, which is what the rest of our conversation
- 10:48today will talk about. But he called these generative
- 10:51AI models high octane language prediction machines
- 10:55capable of probabilistically mapping, among other
- 10:58things, how ordinary people use words and phrases
- 11:01in context. And so you can go read the opinions
- 11:03to determine how he interpreted landscaping and
- 11:06also physically restrained. But he highlighted
- 11:09in the Deleon case, the physical restraint case,
- 11:14he highlighted the benefit of generative AI tools'
- 11:16ability to help interpret a composite multi -word
- 11:21phrase that a regular dictionary or thesaurus
- 11:24can't help us do. Because if you type in physically
- 11:27in the dictionary, it only gives us physically.
- 11:30It won't give us physically restrained. I actually
- 11:33wrote a little article on the platform Medium
- 11:36about how I love using ChatGPT's paid version
- 11:404 .0 as a superthesaurus for the same reason.
- 11:43I like plugging in multi -word phrases, sometimes
- 11:46even sentences, to get a whole list. I'll ask
- 11:49it to give me 10 words that mean, and then I'll
- 11:52type in a really long phrase because I'm trying
- 11:55to capture a particular tone. In case you're
- 11:58interested in what other countries are doing
- 12:00also on the judges level, I haven't found too
- 12:05much yet in multiple countries, but at least
- 12:08three countries. people are using generative
- 12:12AI to predict how courts will decide disputes.
- 12:15And then also judges in those countries are using
- 12:18generative AI, not as a substitute for their
- 12:21own judgment, but as a kind of a gut check. So
- 12:24for instance, in the UK, there is an employment
- 12:27tribunal where some analysts are using AI to
- 12:31predict how a court will decide disputes in that
- 12:34employment tribunal. One judge in Cartagena,
- 12:37Colombia, the country Colombia used GPT to kind
- 12:41of solidify his decision in a health care coverage
- 12:44case and then in Argentina. courts in Buenos
- 12:48Aires are using chat GPT to draft rulings and
- 12:52really complex administrative and tax matters.
- 12:55I put a citation on here too that an American
- 12:58judge has a sub stack blog in which he talks
- 13:00about how judges in the United States might get
- 13:03more involved in generative AI as well. If you're
- 13:06interested in how these LLMs might help us. kind
- 13:10of game out whether contracts will have ambiguities
- 13:14or disputes over particular words, the way things
- 13:17are phrased. Here's a really interesting law
- 13:20review article by two professors called Generative
- 13:23Interpretation. And in this article, they give
- 13:25examples of how they have used large language
- 13:29models to, again, sort of game out or predict
- 13:33whether phrasing in contracts like construction
- 13:36contracts. I was a construction litigator for
- 13:3820 years. So they're 80 page contracts. So in
- 13:42situations like construction contracts or insurance
- 13:45contracts, I think another example in this article
- 13:48was matrimonial settlement agreements. Really
- 13:52interesting just to see how we might use these
- 13:54models to help us prevent ambiguity or conflicts
- 13:58in contract interpretation. So the real crux
- 14:02of our conversation today, let's get into the
- 14:05cautionary tales. I actually wanted to also point
- 14:08out that all of my images in this PowerPoint
- 14:11were generated using AI. I have a very creative
- 14:16brain, but I have zero artistic talent. But I'm
- 14:19also really tired of just using stock art in
- 14:22my PowerPoints because I teach through PowerPoint.
- 14:25So I just decided I would experiment with AI
- 14:28to create really interesting images. And I've
- 14:30really had a lot of fun doing that to kind of
- 14:32jazz up my presentation material. So this is
- 14:35one of my favorite images on this screen for
- 14:37the art cautionary tales. OK, so let's get into
- 14:40it. When I first pitched the CLE program idea,
- 14:45I called it 50 plus, I think it was 50 cases.
- 14:48As of today, I'm recording this on April 25th,
- 14:53I think today is the 25th, yes, of 2025. We are
- 14:58up to 64 total cases involving lawyers, pro se
- 15:02litigants, and experts who have gotten in trouble
- 15:05for misusing generative AI in submitting court
- 15:09filings. 30 of those cases, at least by my count.
- 15:13Of course, I might be missing some that maybe
- 15:15haven't been published, etc. But 30 cases have
- 15:18involved lawyers. So I'm going to break that
- 15:20down for us today. Where did it all begin? All
- 15:24right, so you might have read about the very
- 15:26first case called Matta versus Avianca, which
- 15:29took place in my hometown, the US District Court
- 15:34for the Southern District of New York. So this
- 15:36was the first case that came out. It was June
- 15:38of 23. And in this particular situation, I feel
- 15:42kind of bad for these lawyers because they were
- 15:44the first, but they used Chachi BT to draft a
- 15:50brief because you know we were all hearing about
- 15:51it in the news they they tried it out. Unfortunately,
- 15:54they cited AI -hallucinated cases in an opposition
- 15:58to a motion to dismiss. Here's where things went
- 16:00really wrong, though. When called out by opposing
- 16:03counsel, these lawyers failed to come clean and
- 16:06disclose that that's what they did. And instead,
- 16:09they, quote, doubled down, to use the court's
- 16:11language, and avoided addressing the issue until
- 16:14six weeks after the court ordered them to produce
- 16:17the phantom cases, the cases that didn't exist.
- 16:20So this was the very first sanctioned case the
- 16:23court ordered. the attorneys to send notice of
- 16:25the court's reprimand to the client, write a
- 16:29letter to each of the judges mentioned in the
- 16:32hallucinated cases. Actually, I can't see my
- 16:35screen because the - The transcript is popping
- 16:39up. I think it was a $5 ,000 penalty in this
- 16:41case, which was a pretty hefty sum for the very,
- 16:44very first case. Okay. Since then though, I've
- 16:48tried to categorize the cases because there've
- 16:51been a range of lawyer mindsets about this. None
- 16:55of them are good. Well, one of them is good.
- 16:58One category of lawyers have immediately admitted
- 17:02fault and expressed regret. Because again, these
- 17:05tools are new to a lot of us. We should know
- 17:07better by now, absolutely, in 2025. But the first
- 17:11category I'm going to focus on is the lawyers
- 17:13who immediately admitted fault and expressed
- 17:16regret. The next category I'll focus on is those
- 17:20lawyers, kind of like the Mata versus Avianca.
- 17:23case in which the lawyers at least initially
- 17:26failed to take responsibility or explain themselves.
- 17:30And then there are lawyers who are getting in
- 17:32trouble for willful failure. And then this is
- 17:35my least favorite category because I teach interns
- 17:39and law clerks and junior associates. Some senior
- 17:43lawyers have shifted blame or at least tried
- 17:45to shift blame to the more junior folks in the
- 17:48law firm, and we do not want to be doing that.
- 17:51Courts don't like that either. I'm going to start
- 17:53with the first category, the better category,
- 17:56taking responsibility and expressing regret.
- 17:59Early in February, I have a Canada case on here
- 18:02because it was one of the earliest situations
- 18:04as well. In a Massachusetts court, again, filed
- 18:11a brief or certain filings with AI -generated
- 18:15citations that did not exist. This lawyer, when
- 18:19called out on it, immediately acknowledged responsibility
- 18:22and expressed regret to the court, apologized,
- 18:25etc. This person, this lawyer did talk about
- 18:28the fact that he had delegated the research and
- 18:31the writing to recent law school graduates and
- 18:34a trusted associate, but he didn't blame them.
- 18:37He said he erroneously did not review the citations
- 18:40provided by his more junior team as their supervisor.
- 18:44He knew that he had to take responsibility as
- 18:46their supervisor to check if they were using
- 18:50AI. The court did impose a monetary sanction
- 18:55of $2 ,000, but the court took into consideration
- 18:58the sincerity of the lawyer's contrition. I just
- 19:02wanted to point out the Canadian case because
- 19:05it happened around the same time. Similarly,
- 19:08a lawyer submitted a brief with an AI hallucinated
- 19:10case, but he also conceded the error and apologized
- 19:14and expressed remorse that he was embarrassed.
- 19:18The court found no intent to deceive and just
- 19:22ordered him to pay opposing counsel's costs incurred
- 19:25in researching the non -existent cases. Kind
- 19:29of early on in this whole journey of ours in
- 19:32generative AI, another New York case, a disbarred
- 19:36attorney used a generative AI tool called Google
- 19:39Bard at the time, which has subsequently changed
- 19:42to Google Gemini. This disbarred attorney supplied
- 19:46these cases to his lawyer. The lawyer did not
- 19:49independently review them, but just stuck him
- 19:52in the brief and filed the motion. But again,
- 19:55neither the opposing counsel or the court could
- 19:57find the cases because they were elucidated by
- 20:00Google Bard. This lawyer also, like the others
- 20:03I just mentioned, accepted responsibility for
- 20:05the error and apologized. I found this a little
- 20:09bit shocking. Since the court found no evidence
- 20:12of bad faith, there was no sanction at all in
- 20:14this case. This one was more the public embarrassment.
- 20:17The court actually wrote in the opinion, the
- 20:20court deemed the attorney's acts embarrassing,
- 20:22certainly negligent, and constituting extreme
- 20:25carelessness. Early on, some judges were feeling
- 20:29like the public embarrassment was enough to deter
- 20:32this from happening again. Let's shift now to
- 20:36the unfortunate situation of lawyers failing
- 20:39to take responsibility. In another New York case,
- 20:42I don't know why a lot of these were coming out
- 20:44of New York in the initial era of this, but a
- 20:48lawyer filed a reply brief. This was in surrogate's
- 20:50court, filed a reply brief containing fictitious
- 20:53cases and failed to acknowledge the error. The
- 20:57court struck the pleading. That was the result
- 20:59here. The court struck the pleading and scheduled
- 21:01a hearing to determine the appropriateness of
- 21:04sanctions. I really like the quote that I've
- 21:07called out here, because the judge said, you
- 21:09know, a simple Lexis search of the cases cited,
- 21:12which takes mere seconds, would have shown this
- 21:16lawyer that the cases that they got through AI
- 21:19were incorrect or fake or nonexistent. And so
- 21:23I liked that this judge kind of spelled it out
- 21:25in the opinion that counsel, we have a duty to
- 21:28take minimal time and at least the bare minimum
- 21:31of time and effort to cross check this information
- 21:33to realize this is happening. More failure to
- 21:37take responsibility. These cases are still from
- 21:40last year. And don't worry, I'm going to mention
- 21:42the cases that are literally happening up until
- 21:45as recent as last week or this past week. Another
- 21:49case in this was in the Western District of Virginia,
- 21:53where I took the bar exam ages ago. This council
- 21:57filed a brief citing AI hallucinated cases again
- 22:00and when the court The court deemed this gross
- 22:04error and asked the party to explain themselves.
- 22:07She failed to explain herself. The court noted
- 22:10this silence is deafening. I thought that language
- 22:14was pretty powerful. Again, the court issued
- 22:17an order to show cause as to why the attorneys
- 22:19should not be sanctioned. For some of these,
- 22:22I haven't gone back to check to see what the
- 22:24ultimate sanction is because they happened so
- 22:26long ago and some of the opinions aren't published,
- 22:29but I can follow up on this one. Around the same
- 22:33time in the summer of 2024, in Florida, a different
- 22:37lawyer was subject to a very robust disciplinary
- 22:41proceeding for multiple reasons. But one of the
- 22:44ethical infractions was filing papers with, again,
- 22:48hallucinated cases, failing to take responsibility.
- 22:52And this one tried deflecting blame to his client.
- 22:54I think the client had possibly participated
- 22:57in some of the research using generative AI.
- 23:00All right, so now we'll get to my least favorite
- 23:02category, the attempts to shift blame to junior
- 23:05staff. In California, this was an early case,
- 23:09October 20th, actually both of these are early
- 23:11cases, October 2023. This was an eviction case
- 23:15in which a law firm submitted court filings containing
- 23:18fake case law and the lead attorney on the case
- 23:21sort of tried to shift blame to the first year
- 23:24lawyer in his office who used the generative
- 23:26AI to draft the papers. Interestingly, the court
- 23:30did rule against the law firm and dismissed the
- 23:32eviction suit, but the financial penalty was
- 23:35$999, which a news article about this pointed
- 23:40out said that it was $1 below the threshold for
- 23:43reporting the matter to the state bar. Months
- 23:47later in Colorado, This one had a much more egregious
- 23:52result. The lawyer did the same thing, filed
- 23:55a motion containing cases he found through ChatGBT
- 23:58without verifying their accuracy. He discovered
- 24:02the cases were fake but did not alert the court.
- 24:09Also, this particular lawyer lied and blamed
- 24:13a legal intern. The court issued a stipulation
- 24:16to discipline and suspended the attorney's bar
- 24:19license for one year and one day. That was a
- 24:22much more dramatic outcome. Let's talk about
- 24:26one case in which an attorney was deemed to have
- 24:30willfully failed to comply with the rules. This
- 24:34was a case in the Second Circuit, so again, back
- 24:40on the East Coast. This attorney cited fake cases
- 24:44and after investigation, the court said that
- 24:48they made no inquiry much less the reasonable
- 24:51inquiry required into the validity of arguments
- 24:54she presented. This was beyond just fake cases.
- 24:57This was using generative AI to generate arguments
- 24:59that made no sense. The lawyer ultimately admitted
- 25:04she used chat CPT. I think in this case too there
- 25:07were some other non -compliance issues with the
- 25:11court rules and the court directed the attorney
- 25:14to notify her client of the reprimand and this
- 25:16particular attorney was referred to the Second
- 25:19Circuit's grievance panel. So we're seeing more
- 25:22of this kind of disciplinary action. As I mentioned
- 25:26at the beginning, when we jumped on today, many
- 25:29judges across the country have issued standing
- 25:31orders related to generative AI. Again, most
- 25:34are not saying we shouldn't use it, but are saying
- 25:37we must be ethical and methodical when we use
- 25:41it and make sure we check everything. No matter
- 25:44what court you're practicing in, definitely check
- 25:46the standing orders because they all vary in
- 25:49language and in their requirements. Many are
- 25:52requiring a certification or a declaration, like
- 25:55we add certificates of service or we add different
- 25:58things to briefs depending on the jurisdiction.
- 26:02This particular case was, oh, it's California,
- 26:05September of 24. The judge, I believe, was concerned
- 26:10that the brief was It was a motion that had a
- 26:14fake case and the court was concerned that the
- 26:16brief was generated using generative AI, but
- 26:18there was no declaration about such use attached
- 26:21to the filing, even though this particular judge's
- 26:24standing order required such a declaration. Council,
- 26:29when asked about it, revealed she did use generative
- 26:31AI to draft the motion, but still didn't explain
- 26:34her non -compliance with the standing order.
- 26:36And so that resulted in a show cause order as
- 26:38well. Oh, I put on this read the syllabus. As
- 26:42a law professor, we're asked so many times questions.
- 26:46I'm like, read the syllabus. So this is the judged
- 26:48version of read the syllabus. Read the standing
- 26:50rules. Some cases have simply resulted in adverse
- 26:55rulings, not an additional sanction. In a Texas
- 26:59court case, this one was early on as well in
- 27:022023, a lawyer submitted an appellate brief citing
- 27:06non -existent cases. And the court deemed the
- 27:11brief deficient and illogical because of the
- 27:14the non -existent cases and also I believe even
- 27:18just the logic of the arguments made no sense
- 27:20here. This court couldn't determine with certainty
- 27:23whether the attorney had used generative AI and
- 27:26the court actually resisted the temptation to
- 27:29issue a show cause order or report the attorney
- 27:32to the state bar. But this is one of the very
- 27:34earliest cases that happened in Texas. Back to
- 27:382024, similarly, well, this was also an appellate
- 27:42case. The court struck an appellate brief that
- 27:45failed to comply with the court rules and also
- 27:47contained the fabricated case law, and the appeal
- 27:50was dismissed. So again, adverse outcomes based
- 27:54on citing fake cases and not following the rules.
- 27:57All right, so I will bring us up to date by my
- 28:01count. There are 12 additional cases already
- 28:05in 2025. And again, I might not be catching all
- 28:08these. My librarian at the law school where I
- 28:11teach, New York Law School, has been so awesome
- 28:13at helping me kind of stay on top of which cases
- 28:16are popping up. And again, by our count, we've
- 28:20counted 12 more cases involving lawyers. I'm
- 28:23going to get to the self -represented litigants
- 28:25in just a moment. You can see on the screen,
- 28:30these are only five of the 12. I've got a couple
- 28:32slides on the additional 12 cases. I won't belabor
- 28:36the point. All of these lawyers on the screen
- 28:39today cited filed motions or briefs containing
- 28:43non -existent. I like how the language is. the
- 28:48thesaurus or the synonyms for fake cases. We
- 28:50can call them non -existent, fictitious, fake,
- 28:53one court called them phantom cases, and you
- 28:57can see the the penalties kind of run the gamut.
- 29:00The the United States versus Hayes case was a
- 29:03$1 ,500 sanction. Another what oh the one in
- 29:09the middle the highest sanction that I've seen
- 29:11to date was fifteen thousand dollars because
- 29:14the court issued a five thousand dollar penalty
- 29:17because there were three for each of three separate
- 29:20briefs so that added up to fifteen thousand dollars.
- 29:24Another one was $3 ,000 for one particular lawyer,
- 29:28and he had his Pro Hac Vice status yanked, and
- 29:31$1 ,000 sanction for the two other lawyers on
- 29:33that case. And then the $2 ,500 penalty. So you
- 29:37can kind of see that there's really no, it's
- 29:41all over the map in terms of what financial penalties
- 29:45courts are issuing. Okay, so that's the first
- 29:48five of the 12 additional cases I've found in
- 29:5125. I wanted to particularly highlight this one
- 29:55that happened earlier this year in the federal
- 29:59court in Puerto Rico. Again, a party submitted
- 30:02filings with rampant citation errors. I recommend
- 30:05you read the briefs on this show cause order
- 30:09because the one party who did not cite hallucinated
- 30:13cases wrote a really well -written but somewhat
- 30:16scathing brief about how frustrating it was to
- 30:22see these fake citations and also that the party
- 30:25really didn't take responsibility. The lawyer
- 30:27for the other side did not take responsibility
- 30:29for this. In response to the order to show cause,
- 30:34the party who used generative AI absolutely did
- 30:37asserted it didn't use AI, but that the errors
- 30:40were due to human oversight under, this is the
- 30:43quote from the case, under significant time constraints.
- 30:47But the court said it is not persuaded. And the
- 30:51court deemed sanctions appropriate given the
- 30:54extent of the errors in the motions. And this
- 30:57is what I wanted to highlight for you, the subsequent
- 30:59time and energy required by both defendants and
- 31:02the court to check the plaintiff's filing. I
- 31:06wanted to really flag this language for us because
- 31:09the court said, rather than showing contrition,
- 31:12the memorandum in compliance, which is what the
- 31:14party called their brief, strikes a defiant and
- 31:17deflective tone, and it contained more erroneous
- 31:21citations. So it was like doubling down on doubling
- 31:24down, which really did not bode well. The sanction
- 31:28was in order to pay the defendant's attorney's
- 31:31fees in relation to the filings. And so the attorneys
- 31:35who did not err are required to submit a fee
- 31:40request. And I think that is still pending because
- 31:42this case came out pretty recently. These are
- 31:45the remaining five cases of the 12 new ones.
- 31:50Again, in every single one of these cases, lawyers
- 31:53submitted briefs citing case law that didn't
- 31:56exist. You can see the one on the far right,
- 31:59Kummer versus Lindell in the United States District
- 32:02Court for the District of Colorado, 30 defective
- 32:05citations in that case. Not all of those were
- 32:08fake cases, but miss citings and things like
- 32:11that. In many occasions, the court suspects generative
- 32:18AI use isn't 100 percent certain, but you can
- 32:22see in two of the cases, the lawyers were ordered
- 32:25to serve a copy of the judge's decision on their
- 32:29various state bars. One lawyer, he was not allowed
- 32:34to appear in the case anymore. I think that may
- 32:36have been another Pro Havice situation. Then
- 32:39again, in the Kumher versus Lindell case, there
- 32:42was another order to show cause why this lawyer
- 32:45should not be sanctioned and subject to discipline.
- 32:49The last case for the lawyers, and I'm going
- 32:53to switch the pro se litigants, the self -represented
- 32:56ones in 10 seconds. This one I found really interesting.
- 33:02In the federal court in New Mexico, a lawyer
- 33:06paid a freelance lawyer through a company called
- 33:09Law Clerk, and I capitalized it because that's
- 33:11how it is in the court's opinion. This council
- 33:16paid a freelance lawyer through this particular
- 33:19company to write a brief, and the brief contains
- 33:21six non -existent cases. And it was shown that
- 33:25those cases came from the use of generative AI.
- 33:28So the court fined the lawyer. who hired the
- 33:32freelance lawyer, fined that lawyer $1 ,500 and
- 33:36ordered that lawyer to send a copy of the sanctions
- 33:39order to the company law clerk, require the lawyer
- 33:43to complete a CLE regarding legal ethics and
- 33:46writing or the use of AI and writing, which is
- 33:49what we're doing right now, self -report to the
- 33:52state bar disciplinary boards, and report the
- 33:56freelancer to the state bar disciplinary board.
- 33:59a big deal in that case as well. As you can tell
- 34:04from what I've described, the results have, again,
- 34:07been all over the map. Financial penalties ranging
- 34:10from that California case of $999 all the way
- 34:13up to $15 ,000. That's at least the highest one
- 34:16that I've seen and then the whole range of other
- 34:19actions. All right, so let's now get into the
- 34:23self -represented litigants. So the way the rest
- 34:25of this is going to go for the next half hour,
- 34:27self -represented litigants, I'm going to get
- 34:29into, I think, two experts situations. One very
- 34:37persistent law firm who tried to use AI for a
- 34:40purpose for eight times and was not successful.
- 34:42Then one super recent situation in which a court
- 34:46cited to AI research and a pleading, which is
- 34:48a different category, but I'm going to explain
- 34:50that too. As of today, again, April 25th, 2025,
- 34:56we have at least potentially more. but at least,
- 35:00by my count, 23 cases in which pro se litigants
- 35:04have submitted case filings with fabricated case
- 35:07law, which obviously adds burden to the work
- 35:10of opposing counsel and the court staff. So I'll
- 35:13start with the earlier ones, where the court
- 35:16really emphasized the burden on opposing counseling
- 35:19courts when average citizens who are representing
- 35:22themselves are probably using these public models,
- 35:26most likely a lot of the unpaid models too. Some
- 35:29of the paid versions of these models are getting
- 35:33better and better, obviously. But anyway, so
- 35:37let me just dive into these cases. So in these
- 35:40cases, judges either notice that the cases don't
- 35:44exist. Sorry, in these particular cases, judges
- 35:47are noticing citations to case law that does
- 35:50not exist, or they're noticing really inartful
- 35:54pleadings or just really poorly written pleadings.
- 35:58This is interesting. I don't know how many of
- 36:01you've tried using these tools, but they can
- 36:03actually write clearly and logically. but not
- 36:08so much necessarily in a legal argument. And
- 36:10so the courts are seeing, it's not that they
- 36:12have typos and mistakes in them like that are
- 36:14grammatical mistakes, it's that the logic doesn't
- 36:17make sense or that it's making legal arguments
- 36:20that have no basis in law. And so courts are
- 36:23noticing that and wondering whether these parties
- 36:25have used generative AI. So as you can see on
- 36:28the screen, a couple situations like that where
- 36:31pro se litigants either did submit brief citing
- 36:33imaginary cases, one cited 24 cases, and only
- 36:38two existed and did not stand for the propositions
- 36:41asserted. I will tell you, though, in the pro
- 36:46se litigants cases, the courts have been much
- 36:49more reluctant to impose financial sanctions
- 36:52or any other penalty, but I'll get to some adverse
- 36:57rulings in a moment. I wanted to quote the language
- 37:00from these judges because I found this language
- 37:03interesting as a person who always. went up against
- 37:07opposing counsel in my construction cases and
- 37:10never encountered a pro se litigant in my construction
- 37:13cases. I thought this language was informative
- 37:15to me at least. Magistrate Judge Susan Rodriguez
- 37:20acknowledged the quote latitude that is afforded
- 37:23pro se litigants, but did indicate reliance on
- 37:26generative AI tools may result in sanctions or
- 37:29penalties when used inappropriately. Similarly,
- 37:33another judge called the custom of affording
- 37:37pro se litigants special solicitude, but also
- 37:41said the court won't look kindly on similar infractions
- 37:44in the future. And then another judge, Angela
- 37:47Quiglis, said litigants who use generative AI
- 37:50to draft their brief should not rely on our continued
- 37:53magnanimity. So there is a statement that we
- 37:56afford them latitude, pro se litigants, special
- 37:59solicitude. There's been an effort to be a little
- 38:02bit magnanimous, but that is not going to last.
- 38:06More judges have said the same thing. Any further
- 38:09filings by this particular pro se litigant will
- 38:12result in sanctions or may result in sanctions.
- 38:16Judge Eric Schumacher said, we caution plaintiff
- 38:20that his pro se status does not excuse his failure
- 38:22to check the legal citations he offers to a court.
- 38:25And then Judge Roslyn Silver told the litigant,
- 38:29you must comply with, we have specific rules
- 38:32about this and please check the rules before
- 38:34you file in our court. Something for our profession
- 38:38to think about, Magistrate Judge Gary Stein noted
- 38:41the possibility that pro se plaintiffs might
- 38:45not be aware of the risk that CHAT GPT and other
- 38:49models are capable of generating this fake case
- 38:53law, but also they also might not have access
- 38:55to our paid versions of Westlaw or Lexis so they
- 39:00can check the citations. So that's something
- 39:02that we as a profession should think about. How
- 39:07do we make sure that pro se litigants can get
- 39:09access to tools where they can check the citations
- 39:11and not get in trouble for just relying on public
- 39:15chat GPT and other tools? Some of the results
- 39:19did result in dismissals of these self -represented
- 39:25litigants cases. You can see in the Western District
- 39:29of Washington, the court dismissed a matter when
- 39:31there was case law that was hallucinated. Oh,
- 39:35the same lawyer, remember the case I mentioned
- 39:37about the willful failure. So this was the same
- 39:40lawyer appearing pro se, representing herself
- 39:43in a lawsuit against Delta Airlines back in New
- 39:46York. And the court expressed serious concern
- 39:49regarding this Lydian citation of a fake case,
- 39:52acknowledged that special solicitude custom,
- 39:56but noted this is a lawyer. This is not your
- 40:00non -lawyer self -represented person. She had
- 40:04been practicing law for two decades, so the court
- 40:06dismissed the complaint. Interestingly, in the
- 40:10third case on this screen, Missouri Court of
- 40:12Appeals, this is the case in which there were
- 40:1424 case citations, but only two existed and still
- 40:18didn't stand for the propositions asserted. The
- 40:21court dismissed the matter and imposed a $10
- 40:24,000 penalty on this party for filing a frivolous
- 40:26appeal. There may have been some other stuff
- 40:29going on there I'm not aware of, but that was
- 40:31a pretty significant financial penalty. Again,
- 40:36four more cases in December. These cases pro
- 40:40se parties continuing to file briefs containing
- 40:42hallucinated sanctions. Again, a couple of these
- 40:46courts declined to impose monetary sanctions,
- 40:49but issued adverse judgments against these parties.
- 40:53One court presuming AI used dismissed the party's
- 40:57claims. You can see another court dismissed the
- 41:00consolidated appeal and actually This particular
- 41:04Gutierrez case involved a lot of transgressions
- 41:07not following the rules so that that lawyer was
- 41:10barred from further pro se filings. Okay and
- 41:14then two more cases bringing us up to almost
- 41:16the present two more cases in March just last
- 41:19month. This One of these parties, again, we're
- 41:24still in the self -represented litigant category,
- 41:27submitted a brief containing what the court said
- 41:30were alarming miscitations. The court suspected
- 41:34AI misuse and emphasized, again, this is a burden.
- 41:38This is a burden on the court. It's a burden
- 41:41on opposing counsel. This was, again, back in
- 41:44the district court of Puerto Rico, acknowledged
- 41:47this is becoming an increasing problem among
- 41:50pro se parties. In the Court of Federal Claims
- 41:54in March, similarly, self -represented litigant
- 41:57cited hallucinated cases. And I kind of liked
- 42:00the way this court framed it. The court said
- 42:03plaintiff's briefing built on AI -generated cases
- 42:06that stand for legal propositions in direct contravention
- 42:09of actual case law is the epitome of baseless.
- 42:13But once again, the court issued a warning because
- 42:16it was a self -represented litigant instead of
- 42:19imposing sanctions. Two more recent cases in
- 42:22April. I think I have these two and then one
- 42:24more and then we'll shift to experts. Two more
- 42:27recent cases in April. We're back in Texas. This
- 42:31litigant filed objections to a magistrate judge's
- 42:35report, but cited a non -existent case which
- 42:38made the court suspect generative AI misuse.
- 42:42just reminded the prostate litigant that they,
- 42:45just like an attorney, are bound by the court's
- 42:48rules. There's no getting around being bound
- 42:51by the court's rules. So the court overruled
- 42:54the objections to the magistrate judge's report
- 42:56and dismissed the case with prejudice. Same thing
- 42:59in the on versus arch block case. Oh, this was
- 43:03interesting. I put on here that this judge suspected
- 43:06AI use because there were problematic citations
- 43:09in the motion that the self -represented party
- 43:12filed. Oh, and they failed to comply with the
- 43:15certification requirement of AI use. In this
- 43:18judge's standing rules, there was a certification
- 43:21requirement if the party used AI. Suspecting
- 43:25AI use, though, the judge tried to find the cases
- 43:28themselves using Westlaw and ChatGBT and couldn't
- 43:33find a case. The court does a really nice job
- 43:36in this case of discussing both upsides and risks
- 43:40to generative AI. So again, not saying that these
- 43:43tools are terrible for everything, they're not.
- 43:46I find them really helpful personally in my own.
- 43:50I love writing, so I don't outsource my writing
- 43:52to generative AI, but I really love how it boosts
- 43:55my creativity and word choice, like I mentioned,
- 43:58using it as a super thesaurus. The court in this,
- 44:01in this on versus arch block case, highlights
- 44:04some of the upsides that it's, you know, efficient,
- 44:06it's efficient in summarizing documents, etc.
- 44:09You can read the opinion if you want to know
- 44:11the other upsides discussed in that case, but
- 44:14also the court identified the risks. And because
- 44:18this particular litigant like one of the cases
- 44:21I mentioned at the very beginning, doubled down
- 44:23and was not forthright in their use of AI, the
- 44:27court denied the person's motion. Last case involving
- 44:32a self -represented litigant happened just on
- 44:36April 15th, 2025, so 10 short days ago, Eastern
- 44:41District of Michigan. This was a self -represented
- 44:46litigant I don't know how to pronounce the Latin
- 44:49for the proceeding. I forget how to say it. But
- 44:52you can look it up. An indigent litigant hand
- 44:55wrote at the bottom of a typewritten brief, this
- 44:59motion was done by AI. They wrote the little
- 45:02i. And then the court wrote, that is a bad idea
- 45:06in one very real and important sense. This plaintiff
- 45:09now will be held responsible for the content
- 45:11of his filings under federal rule 11, which we
- 45:16all know requires us to verify that we're signing
- 45:20based on the accuracy of all facts and law in
- 45:23our brief, whether we use artificial intelligence
- 45:25or not. And then the court emphasized they will
- 45:28not accept AI as an excuse for inaccuracies.
- 45:32Okay, we've already talked about the Gutierrez
- 45:34case, but I thought it's interesting that this
- 45:37was the first time I had seen the word phantom
- 45:40case law in the case and it was because the opposing
- 45:43counsel wrote a notice to the court of phantom
- 45:47case law pointing out that they suspected that
- 45:51the case law cited in the pro se plaintiff's
- 45:54documents did not exist. This was the case in
- 45:57which the court ended up barring that particular
- 46:00plaintiff from further appellate pro se filings
- 46:03in the particular probate action because he had
- 46:06violated a lot of the court rules. All right.
- 46:09I forgot that I included a slide on two more
- 46:12cases in which judges suspected generative AI
- 46:15in the context of poor briefing, but because
- 46:17they had really no proof, they just noted the
- 46:21possibility. Again, at SDNY, I believe this is
- 46:26Magistrate Judge Ona Wang. I took my students
- 46:28to observe court because she was holding a status
- 46:32conference in an interesting AI case. And I believe
- 46:35this is Judge Wang said that she has no general
- 46:38prohibition on the use of AI, but this particular
- 46:40party's brief was not helpful to the court. OK,
- 46:44so let's switch gears a little bit and talk about
- 46:46experts. And I wanted to talk about two cases,
- 46:51to my knowledge, in which experts have stumbled
- 46:54or fumbled by trying to rely upon, or actually
- 46:59the first one is intentionally trying to rely
- 47:02upon AI to support an expert opinion. Back in
- 47:06New York, in the surrogate's court, actually
- 47:10I said this was a recent case, but now it's already
- 47:13been about six months, but An expert witness,
- 47:16this was in, let me go to the next slide. An
- 47:19objector, so a party who is objecting to the
- 47:23accounting of an estate in an estate administration
- 47:26scenario hired an expert witness to render an
- 47:29opinion about damages calculations related to
- 47:32real estate property. There were some allegations
- 47:34that the trustee had mismanaged. either the timing
- 47:38of the real estate sale, I believe, or the value
- 47:40of the property. This particular expert had used
- 47:45another generative AI tool, Microsoft Copilot,
- 47:49not to do his calculations for him, but to cross
- 47:53-check his calculations. He submitted his conversations
- 47:58with Microsoft Copilot as part of his expert
- 48:00report, or I believe maybe as part of his expert
- 48:03testimony. The judge, was interested, but deemed
- 48:09the damages calculation done by Microsoft Copilot
- 48:12as unreliable, and here's why. because the expert
- 48:16didn't take the extra steps or couldn't take
- 48:20these steps perhaps of explaining how Copilot
- 48:23functions, how it produces its output, what sources
- 48:26it relies upon, or how the resulting calculations
- 48:30in this case met the reliability standards applicable
- 48:33to the expert opinions. I think the court even
- 48:37tried to experiment with Microsoft Copilot to
- 48:41see if it was repeatable and therefore more reliable,
- 48:45and the court deemed it not reliable. So this
- 48:49particular judge accentuated, like that verb,
- 48:52I mean, I wrote that verb, and then quote, council
- 48:55has an affirmative duty to disclose the use of
- 48:58artificial intelligence in the context of expert
- 49:01reports. And then any proposed evidence would
- 49:04be subject to a fry hearing. The standard in
- 49:06New York state court here was the fry standard
- 49:09prior to its admissions. I wanted to take just
- 49:12a second to refresh our recollection about expert.
- 49:16testimony admissibility standards based on three
- 49:19different rules, and then I'll explain the other
- 49:21expert scenario that went awry because of generative
- 49:25AI. For a standard, you might be familiar with
- 49:28this, but scientific evidence, expert evidence
- 49:32must be based on a principal procedure or methodology
- 49:35which has, quote, gained general acceptance in
- 49:37its particular specified field. So unless we
- 49:40can show that generative AI generated damages
- 49:44calculations have gained general acceptance,
- 49:47they're not going to meet the FRI standard. The
- 49:50Frye rule says it doesn't have to be universally
- 49:52accepted or unanimously endorsed in the field,
- 49:55but it must be reasonably relied upon and generally
- 49:58accepted as reliable, which the judge in that
- 50:01prior case I mentioned implied that the Microsoft
- 50:04Copilot damages calculations would probably not
- 50:07meet that standard. The court will, under the
- 50:10Frye standard, will consider judicial opinions,
- 50:13scientific or legal writings, or experts other
- 50:16than the proffered expert and analyzing the reliability
- 50:19question. As we all know, Federal of Evidence
- 50:25703 talks about reliability as well. Under FRE
- 50:30703, so in federal court, an expert's opinion
- 50:33must first be helpful to the trier of fact, must
- 50:37be based on sufficient facts or data. must be
- 50:41the product, here's the reliability, must be
- 50:43the product of reliable principles and methods
- 50:46and reflect a reliable application of those principles
- 50:50and methods to the facts of the case. And then
- 50:53our trusty old Daubert standard or Daubert standard,
- 50:56we have to show if we're submitting an expert
- 50:59report or expert testimony in federal court or
- 51:01in a Daubert jurisdiction, The court will analyze
- 51:06whether the expert's technique, theory, or methodology
- 51:09can be and has been tested, whether it's been
- 51:12subjected to publication and peer review, whether
- 51:16it has a known or potential error rate. The court
- 51:20will look at whether standards exist and whether,
- 51:23again, the technique, theory, or methodology
- 51:25has attracted widespread acceptance within a
- 51:28relevant scientific community or within the relevant
- 51:31community of them. industry or a subject matter
- 51:35of that expert. Okay, so getting to the next
- 51:39expert situation. You may have read about this
- 51:42in the news. This one is, again, just frustrating.
- 51:47So here's the situation. So in Minnesota, there
- 51:52was new legislation proposed to ban what's called
- 51:56deep fake technology to influence an election.
- 51:59I don't know the technical, technical definition
- 52:01of deepfake, but basically it's using technology,
- 52:05likely AI, to make us think that a particular
- 52:08person is saying something about a particular
- 52:12thing, a public opinion, etc. So Minnesota's
- 52:18attorney general hired or engaged an expert.
- 52:22to opine in favor of this new legislation to
- 52:25ban deep fake technology to influence election.
- 52:28And this expert had massive technical expertise.
- 52:35This person was the founding director of the
- 52:37Stanford Social Media Lab. This particular expert
- 52:40submitted a declaration in support of the law
- 52:44preventing deep fakes. deepfake technology to
- 52:47influence an election. But unfortunately, the
- 52:49declaration from this expert cited academic sources
- 52:53that do not exist. So the lawyers challenging
- 52:56the law, challenging the anti -deepfake legislation,
- 53:01suggested the sources were generated by an LLM.
- 53:04And as it turns out, the expert, he was very
- 53:07forthcoming. He admitted that he used an AI tool
- 53:10to help him as part of his drafting process,
- 53:13which many people do. Like I just, I said that
- 53:15I use AI to help me come up with different vocabulary,
- 53:19et cetera, check my grammar sometimes. Problem
- 53:22was this particular expert had put placeholders
- 53:25in his declaration where he was going to go back
- 53:29and fill in citations later. But as it turned
- 53:32out, if I'm reading the opinion correctly, ChatGBT,
- 53:35I forget which tool he used, I believe it was
- 53:37ChatGBT, but anyway the AI tool filled in the
- 53:41citations for him and he didn't notice that the
- 53:45the LLM had done that for him and he didn't go
- 53:47back and check those new newly inserted citations
- 53:52using the technology he usually uses for academic
- 53:55citation checking. And so those fake citations
- 53:58slipped into the declaration and went got in
- 54:01front of the court. And so the court excluded
- 54:03this expert's testimony and saying that the citation
- 54:08of fake academic sources, quote, shattered his
- 54:11credibility. The shatter verb was the court's
- 54:14language. So not good. Okay, as recently as three
- 54:19days ago, three days ago, April 22nd, 2025. Actually,
- 54:24I don't know why I'm being so dramatic. This
- 54:26case is not that dramatic. This is just an interesting
- 54:29recent case that happened three days ago. Yelp
- 54:31versus Google. So two kind of technology related
- 54:35companies. Yelp is asserting an antitrust action
- 54:41against Google. Yelp is claiming that Google
- 54:44has a monopoly on local search services online.
- 54:50Procedurally, Google filed a motion to dismiss
- 54:52Yelp's antitrust action. And in the pleading,
- 54:55in the complaint, I believe, Yelp pleads this
- 54:59quote. While exact figures can vary, it's generally
- 55:02estimated that Google holds well over 90 % of
- 55:05the market share in local search. And then it
- 55:08cites, or sorry, I haven't looked at the exact
- 55:12complaint, but the court opinion on this issue
- 55:14says this fact was generated by a query to Google's
- 55:18Gemini AI tool. Google, in filing a motion to
- 55:23dismiss, argues that citation to an artificial
- 55:26intelligence tool is not a plausible allegation.
- 55:29Interesting. The court said Google's concerns
- 55:33In the court's ruling on the motion dismissed,
- 55:35the court said Google's concerns about, quote,
- 55:38pleading by bot are legitimate, but neither the
- 55:41federal rules of civil procedure nor the local
- 55:44rules prohibit such use of AI, which is true.
- 55:46We're not prohibited from using AI. So at this
- 55:49point in the case, the court declined to address
- 55:52any perceived to be negative implications of
- 55:57Yelp's use of AI in the pleading. So we can stay
- 56:00tuned and see what happens with that case. Our
- 56:03final category before I get to kind of what we
- 56:06should do about all this in our last few minutes
- 56:08together, one particular law firm, again back
- 56:12in New York, has it represents parties in cases
- 56:17alleging individuals with Disability Act, IDEA
- 56:21claims against the New York City Department of
- 56:23Education. So eight times, so basically eight
- 56:27times this particular law firm won on behalf
- 56:30of their client, they won the IDEA claim against
- 56:33the New York City Department of Education and
- 56:35under the statute they're entitled to recover
- 56:37attorney's fees. But eight different times it
- 56:41attached dialogue with CHAT CPT prompts and responses
- 56:44to support assertions of alleged prevailing party,
- 56:49sorry, prevailing market rates in its attorney's
- 56:51fees request. So the CHAT CPT was an attempt
- 56:55to groove up the prevailing market rates of the
- 56:58attorney's fees amounts. Eight times this law
- 57:02firm did this and eight times all the judges,
- 57:05each judge and eight. All eight of the cases
- 57:07rejected, disregarded, and declined to consider
- 57:10or credit the CHAT GPT, quote, evidence in support
- 57:14of the law firm's fee requests. Based on going
- 57:17back to our discussion about evidentiary standards,
- 57:21CHAT GPT inventing false information, the tool's
- 57:24propensity to fabricate information, and its
- 57:26current reputation, this was all in 2024, as
- 57:30an unreliable source. I think the last case I
- 57:33found on this case was maybe October of 24. Language
- 57:38in two of the judicial opinions, though, acknowledged
- 57:40these decisions were based on the existing state
- 57:43of generative AI. Judge Paul Englemeyer in one
- 57:47of these cases said, barring a paradigm shift
- 57:49in the reliability of this tool, the firm is
- 57:53well advised to stop doing this, stop referring
- 57:56to CHAT GPT in future fee applications. But he
- 58:00was talking about maybe eventually, I mean, I
- 58:02interpreted this as maybe eventually there'll
- 58:04be a paradigm shift. And Magistrate Judge James
- 58:07Cot says, chat to BT at least in its current
- 58:10iteration and that was back again in 2024 should
- 58:14not be utilized to support a fee application.
- 58:17I know we're coming up against time so I will
- 58:20make some recommendations and then I'll see if
- 58:22there are any questions and people can stick
- 58:24around for a few extra minutes if you're curious
- 58:26about the questions but otherwise everybody else
- 58:29in two minutes I think can log off. How can we
- 58:32be part of the solution? Basically what we're
- 58:34doing right now, what we've just been doing for
- 58:36the last hour, let's understand generative AI
- 58:39tools current limitations related to legal research.
- 58:42I mean, even our legal professions tools are
- 58:46not perfect. So we really need to understand
- 58:49what these tools are good at and what they still
- 58:52need to grow on. We have to incorporate site
- 58:56verification protocols into our workflow. We
- 58:58don't have the luxury of relying on these tools.
- 59:02We have to check every single site. We can use
- 59:05them to generate citations, but we have got to
- 59:07check every single site. We also, as we're doing
- 59:10today, should provide generative AI literacy
- 59:13training for members of our institutions, offices,
- 59:16and communities from entry level all the way
- 59:18up to senior leadership. I don't know about you,
- 59:22but if you've read Dr. Ethan Malik is a professor
- 59:26at Wharton, and he's is really linking education
- 59:31and AI and he sets down what he calls sort of
- 59:34the 10 hour challenge. He said all of us should
- 59:37play with these tools for 10 hours to understand
- 59:40for our own purposes what they are good at and
- 59:43what they are not good at because we learn a
- 59:45lot in 10 hours. So I want to kind of re -up
- 59:48the 10 hour challenge and use it for things that
- 59:50are even just fun or interesting in your life
- 59:53that have nothing to do with your work and you'll
- 59:55learn a lot. Continuing on with how we can be
- 59:58part of the solution, let's enhance open and
- 1:00:00transparent communication up and down our staff
- 1:00:05and supervisory systems to make sure that people
- 1:00:09aren't using this without our knowledge. But
- 1:00:11if they are, we're all in communication about
- 1:00:13it. Let's commit time to making sure that we
- 1:00:16edit any generative AI generated legal writing
- 1:00:19to ensure our work product is helpful to judges
- 1:00:21and our legal readers. You saw the cases I flagged
- 1:00:25in which judges said these briefs just are not
- 1:00:26helpful. Let's make them helpful. And then absolutely
- 1:00:30we need to comply with judge's standing orders
- 1:00:32related to certifying. On the pro se part, I
- 1:00:36mean, our profession can probably do more to
- 1:00:38educate the average citizen about the existing
- 1:00:41perils and relying upon public generative AI
- 1:00:44tools and somehow maybe we can help provide free
- 1:00:48or cost effective methods for pro se litigants
- 1:00:50to verify cases that they have found as well.
- 1:00:53And then let's remind our experts about the evidentiary
- 1:00:55standards that will absolutely be applied to
- 1:00:58opinions and calculations derived with the assistance
- 1:01:01of generative AI tools. So I think That is it.
- 1:01:05I've written some articles about these that are
- 1:01:07in the materials, and please feel free to reach
- 1:01:09out to me if you have questions or want follow
- 1:01:14up or you want me to come talk to you and your
- 1:01:17law firm more about these particular topics.
- 1:01:20Okay, so let me see if I can find the Q &A in
- 1:01:24my toolbar, and I will see if you have any questions.
- 1:01:28I'll stop the share. Q &A. Okay. All right, great
- 1:01:35question. How can one reasonably comply with
- 1:01:38ABA rule 1 .1, comment 8, and similar rules when
- 1:01:41technology is advancing so exponentially that
- 1:01:43once one is versed enough to comply that the
- 1:01:45next thing is already past its shelf life? Great
- 1:01:48question. I mean, I think these things are evolving
- 1:01:51on a daily basis, but we can't just. wait until
- 1:01:55they're, quote, perfect. We really do have to
- 1:01:57try. That's why I like Professor Malik's 10 -hour
- 1:02:00rule, because if we just commit to experimenting
- 1:02:04with these tools, both the public -facing ones,
- 1:02:07and definitely don't put any confidential information
- 1:02:10into those, but also the legal professions ones
- 1:02:14and testing them out, that, in my opinion, I'm
- 1:02:17not an expert on professional ethics compliance,
- 1:02:20but I would say As long as we are staying abreast
- 1:02:24of benefits and risks associated with these tools,
- 1:02:27we don't have to be experts in these tools. And
- 1:02:30I would go ahead and read the standing rules
- 1:02:32in your jurisdictions, because that's what the
- 1:02:33judges are saying too, and the ethics opinions
- 1:02:36from various jurisdictions. We're not supposed
- 1:02:38to be tech experts, but we also can't avoid getting
- 1:02:42to know these tools until they're quote perfect,
- 1:02:45because we just don't have that choice. So I
- 1:02:47would recommend playing around with them. Okay,
- 1:02:51and second, is AI and contract interpretation
- 1:02:55capable of taking into account canons of construction,
- 1:02:58some of which are codified in California? So
- 1:03:01I have a whole other seminar on prompt engineering,
- 1:03:04but if you give instructions to the AI tools,
- 1:03:08it will apply canons of construction. There are
- 1:03:15some great tools. I'm not going to say they're
- 1:03:17great because I haven't really tried them. There
- 1:03:18are a plethora of contract drafting tools and
- 1:03:21contract interpretation tools that you can experiment
- 1:03:24with, and I think that that would be a good experiment
- 1:03:27to undertake as well. But I do have a whole ton
- 1:03:30of stuff on prompt engineering if you're interested
- 1:03:32in learning how to teach the tool the canons
- 1:03:35of construction. And then, okay, is there an
- 1:03:39AI app that can be used to verify the existence
- 1:03:42of actual cases behind citations in an AI -generated
- 1:03:45legal writing? I do not know, but I think you
- 1:03:51can probably test out whether you can upload
- 1:03:55a document containing citations and run it through
- 1:03:59the site checking tools on Westar Lexis. I have
- 1:04:02not... tried that, but I would recommend that
- 1:04:05we experiment again as part of our 10 -hour challenge.
- 1:04:09All right. My last question, I think, by checking
- 1:04:11citations even if generated by AI, what type
- 1:04:15of checking are you suggesting, jeopardizing
- 1:04:17through digital means, wondering if case citation
- 1:04:20check embraces AI? What I mean by this is if
- 1:04:25you've used any AI tool, whether it's public,
- 1:04:28whether it's Westlaw or Lexis, whether anything,
- 1:04:32we have to do our due diligence as lawyers, no
- 1:04:35matter what. So I would say, take the citation,
- 1:04:39type it into a tool that we know has polls from
- 1:04:44real cases, like Westlaw or Lexis and other ones
- 1:04:47that are out there, I suppose, and read the case.
- 1:04:51We can't just rely on the citation or the AI
- 1:04:55tools to also verify that we can't rely on the
- 1:05:01AI generated content as the case saying what
- 1:05:05it's proposed to be saying. We still need to
- 1:05:07go back and read these cases. That's why I know
- 1:05:10there's a lot of hype out there right now about
- 1:05:12how these tools are supposed to make us super
- 1:05:14efficient. My personal opinion, I don't speak
- 1:05:17on behalf of anybody except myself, is that initially,
- 1:05:20The research part is not making us faster because
- 1:05:23we still have to check everything. Eventually
- 1:05:26it will get better and better and we might get
- 1:05:27more efficient at it, but now I would say don't
- 1:05:31err on the side, just do it. Check every case
- 1:05:34citation, still read the cases, make sure these
- 1:05:37are actual cases coming from real. publications
- 1:05:41or the sources that we trust with like a Westlar
- 1:05:44-Alexis citation, even if it's an unpublished
- 1:05:47case on Westlar -Alexis. So that's my answer
- 1:05:50to that. Keep, keep checking, keep checking,
- 1:05:53keep checking. If you have any questions additionally,
- 1:05:57please feel free to reach out. You have my contact
- 1:05:59information. I love kind of introducing people
- 1:06:02to generative AI literacy. I love teaching prompt
- 1:06:06engineering and Again, I would just want to put
- 1:06:09out there like I love writing. I identify as
- 1:06:12a writer before my fancier titles of professor
- 1:06:15or lawyer. And so I'm never going to totally
- 1:06:19outsource my writing to AI, but I do enjoy using
- 1:06:22it to boost turbo boost my creativity and just
- 1:06:27as kind of a supplement or a companion, but not
- 1:06:31a replacement. I will leave it at that. Thank
- 1:06:34you very much for your attention and I know I've
- 1:06:37thrown a lot of information at you today, but
- 1:06:38hopefully we all learn something.