Name: Candor in the Age of AI: Ethical and Practical Considerations for Lawyers and Judges
Uploaded: 2025-08-07T18:26:03.859Z
Duration: 1 h 10 min 8 s
Description: Candor in the Age of AI: Ethical and Practical Considerations for Lawyers and Judges

Transcript for "Candor in the Age of AI: Ethical and Practical Considerations for Lawyers and Judges": Good morning, everyone, or afternoon, depending on where you are, and welcome to today's webinar, A Candor in the Age of AI, Ethical and Practical Considerations for Lawyers and Judges. I'm Gina Jurva. I'm an attorney and consultant here for Everlaw. I will be moderating this discussion, and we are thrilled to dive into really one of the most pressing issues faced in the legal profession today, and that is the ethical and practical implications of generative generative AI in the law. And tools like generative AI, when we're talking about ChatGPT, Gemini, fill in the blank, they're really transforming how lawyers research, draft, and argue cases. I mean, basically, how we do our job. But with this power, of course, comes profound responsibility. So how do we uphold our duty of candor to the courts when AI can possibly hallucinate or distort facts? What happens when judges themselves use AI in their opinions? Now these aren't hypotheticals anymore. We have some real world examples, which we are going to talk about here today. You're gonna hear some cases, Mata v Avianca, Ross versus the United States. Don't worry. You don't need to know what those are right off the bat. We're gonna discuss them. You may already know, but we're gonna we're gonna talk about them a little more in-depth. To help us navigate these questions, we are truly honored to have professor Rebecca Delfino, associate dean for clinical programs at LMU Loyola Law School. Professor Delfino is a leading expert on legal ethics, civil procedure, and the intersection of law and technology. Her scholarship on AI's disruptions from deep fakes to generative AI itself have been widely cited, and she's here today to share her insights with all of us. So, professor Delfino, thank you so much for joining us. Well, thank you for having me. I'm happy to be here. Well, I'm excited to have this discussion with you. One thing I should mention before we go any further is that professor Delfino will be live with us at our Everlost Summit. It's a two day thought leadership conference in real life, in October in San Francisco, where she will be, an actual a panelist on a session called generative AI in the courts. And she will be, accompanied by Casey Newton, who's a technology journalist and founder of the publication platformer, cohost of New York Times podcast, The Hard Fork, some of you know it, Nilay Patel, who's the founder and editor of The Verge, which is, also a publication online, which is dedicated to technology, and Bridget McCormack, who is a retired chief justice of the Michigan Supreme Court and president and CEO of the American Arbitration Association International Center for Dispute Resolution. So this is a powerhouse panel. We're so excited to, to see that live in person in October. But let's let's start, first of all, in this discussion with really, like, level setting where we are. I mean, there's so much out there about generative AI. It can get confusing for so many people, myself included. And so I really wanna just ground the conversation, first of all, in how AI is actually being used in courts today. And, you know, we have seen professor, like, litigants deploy AI from everything, like I mentioned earlier, drafting motions, summarizing discovery, just overall in case preparation and and, presentation to the court. And then those judges who experiment with it for research and even drafting opinions. So can you highlight just a few cases, that you're seeing? And, of course, we'll discuss them more in-depth later in this webinar. Yes. I think that's a really important place to start because as you mentioned, we are seeing AI tools integrated at lightning speed across multiple stages of litigation, both with litigants and the courts. I like to equate it as to sort of, like, living in the wild, wild west, which none of us did live at that point in history, but it it it is really a good analogy where we are seeing people do different things in in some unsurprising and unregulated ways, in ways that they think back on it and think, well, maybe I shouldn't have done that or where, you know, they're exploring sort of the edges of what, norms and rules certainly allow and and it's an explosion. So on the litigate side, so we we are, as you mentioned, seeing people draft routine filings, motions, settlement agreements, summarizing volumes of discovery, and, you know, even even using AI to create evidence. There's a there's a there's a case that I'm sure your your, your participants have heard about. It was just in May, in Arizona where AI was used to create a deep fake of a victim impact statement in a criminal case. And that was a fascinating situation where you had a a victim who who had died as a result of the crime for which the individual was convicted, providing this, testimony about what his life was like. And this this this AI, avatar, created by his sister, gave compelling, moving testimony to the people in the courtroom. I mean, people were crying, and and and it was interesting because it it had this interesting impact. The the, the defense lawyer didn't even object because I don't think he appreciated what was about to unfold, and it did I mean, they're appealing it now. But but, you see everything from that to judges. On the other end, there's a there's a case I wanna talk a little bit about, from New York, where, an expert used AI to to create an expert opinion, and then the the court questioned him about it on the stand, and it was revealed he had used, Copilot to develop his opinion. And the court interrogated that, and then it was revealed the expert didn't really understand how Copilot worked, and, and so the court criticized him, rightfully so, but then the court flipped back and used the AI, the judge himself to to to to try and figure out the answer to this question. So we see these things happening that, six months ago, nobody would have been talking about or thinking about, and and and it does it's, it it is a range of of of uses, which which which are interesting, and powerful, but, but we need to be careful. We need to move deliberately and thoughtfully, and and and hopefully, as we become more, you know, acquainted with what it's like to live in the wild West, we will equip ourselves to be able to do that effectively. But it's it's amazing. It's an amazing time to be a lawyer with this new technology. Yeah. I mean, it's so transformative. I I I mean, in everything you said here, you know, we are the type of cases that are coming up where this is becoming an issue or at least something that we need to discuss, is, you know, everything from chat, CPT hallucinating, quote, unquote, to what you're saying, like, judges using it. So we are in, like, the the early stages of of how how to use generative AI. And so I think this is, like, a perfect time for us to actually be talking about about all of this. One question though, and this is just something that I've thought about, and I I wondered if you have as well. What happens Have there been any cases where there has been a litigate who or a party who wants to use generative AI to leverage generative AI for efficiency sake, and the other party objects and the court itself doesn't have any standing orders on it. Like, there's no rules for this particular court on generative AI, but you have two parties in conflict. Have you seen that come up yet? I haven't seen it in the published case law, but I have heard about it. And, generally, although there there although there are a number of courts that don't have standing orders, though more are coming online every day, just inherent in a court's ability both in the federal and state courts, court's ability to manage proceedings, courts are allowed to set parameters for behavior and disclosure. And I and in looking at the states and the federal law and the federal rules of civil procedure, I have not yet come across an absence of a tool that can be applied already. Now these tools about managing proceedings, managing the presentation of evidence, they may they are not necessarily specifically addressing AI and technology, but they are addressing other ways that people bring motions or present evidence. And there's enough space and discretion for the court to say, hold on. Stop. Let's take a sidebar or let's talk about this in a way that sort of preserves the status quo for now and figure out what you're doing and how you're using it and interrogate this further. So, you know, for as much as this is new, there are many old rules that exist that can currently be applied. And I think it's only a matter of time before we see this, in case law. So as as I I I know you know and and and your listeners know, when we look at case law in general, what we're looking at for the most part are appellate opinions. So it takes a moment. Right? I mean, so a lot of this we could see trickling up through some of the district court opinions, but many of those aren't published. And so there there's not an ability to really investigate what's happening yet, but it's on its way. Right? It will there are gonna be claims not long where someone is saying AI was used in this way and it was prejudicial or it wasn't proper or it violated some some some rule or some some some norm, some due process value. So we're on our way, but, so this we'll get there with respect to, people objecting. I think, actually, the closest time that I have seen this is the recent case that I referred to with with the victim impact statement Sure. Where the lawyer objected after the fact and said, your honor, you you know, the prosecution asked for a nine year sentence. You gave him ten years. And I and that could only the only way that that can be accounted for was by your allowing them to present this evidence that that was so influential on on the your decision making process. Right. And as you said, I mean, I think for those that aren't lawyers on on this webinar, it's just, you know, when when something goes to the trial court, the first court that hears evidence, it takes a while for that appeal to reach the appellate court, and whether you're in, you know, state court, federal court. And so that's why these are taking some time as you said, but we're seeing it more and more and more. And I think that's what's so cool about having you here is that you're looking at you're looking at all of these cases, all of these courts in totality, and so you can provide us sort of, like, an an overview and some specifics. So one question my last question before we really get into, like, the ethical stuff is when you begin researching, AI's impact on legal ethics or just AI's impact on the law in general, was there what surprised you the most? Like, was there anything that just came out and you were like, wow. I haven't thought of that or, oh, this is gonna be a big thing, anything at all? Yeah. I think what what has surprised me the most is, honestly, is how much faith people have placed in the output of generative AI, how quickly they came to treat it as authoritative even though we know it can hallucinate. But when I began researching AI on legal ethics, I expected to find gray areas in the ethical rules, and they're certainly there. But I didn't expect and, frankly, was quite surprised, and I have been surprised to see lawyers and even judges so readily trusting technology's results. Even their their trust they trust it more, it it seems from my observations, even more than they trust, you know, junior colleagues or their own instincts. And and I think this is because it's seductive. Right? AI has this quality because it sounds confident and because it gives you answers instantly. It gives this illusion of being accurate and competent, but we know from from data and real world cases that it that it fabricates. I mean, there's a there's a scholar, in Europe, who has been tracking every every case both internationally and in The United States where someone has been sanctioned for using AI and it's fabricated something. And there's, like, almost 300 cases, which doesn't seem like a lot, but it it seems like a lot to me as someone who teaches ethics when the ethical rules, that we're gonna talk about are are are are pretty clear. And and it's also surprising to me because precision and accountability as lawyers is so important. You know, it's it's, it's it's really surprising to me how how how confidently and how faithfully lawyers have said, well, it must be right, and judges as well. That that surprises me because I think of lawyers, I think of people that are steeped in the law as or or should be more circumspect before they leap into this is the answer. But that that's what surprised me. It's just the how much faith we put in it already. Well and it's and it's true. It's just and for those that aren't lawyers on the webinar, like, as lawyers, we have to pass the bar exam. That's one thing. But there's a separate exam we take as well on the professional responsibility, that we and part of that, which we're about to talk about, is are these duties to of candor to the court. And so I think that's a perfect way to segue into that. So when we're talking about the ethical pitfalls that lawyers can face when using generative AI, I think we should start with just some some ethics basics. And I pulled up ABA model rule three point one and three point three. Those require lawyers to ensure arguments are grounded in fact and law, certainly not fiction. And, you know, we Abianca case where we saw lawyers submit a lawyer submit chat GPT generated citations to a case that simply did not exist. And like you just said a moment ago, it's really interesting to think about the faith we put into, these large language models, generative AI. Part of it, I imagine, you said there's, like, a a sort of, like, a subduction to it because it sounds, it sounds authoritative. It's also super efficient if you're a busy lawyer, not to say that, you know, it means you don't have to you no longer have to follow ethical rules. But I can see that, you know, why this may be happening more than than we would expect. So professor, you know, how would how should lawyers, attorneys reconcile the efficiency of AI with their duty of candor to the court? Right. That's a great question. And I think, you know and and and Mata is an interesting example of that because, the the citations looked plausible. They looked perfect. And and I think where the where the the feeling is in Mata in similar cases is it's not it's not just using AI because there's nothing wrong with using AI. AI is valuable. The tools are valuable. Even the chat box that are commercially available and and are are not proprietary as in belonging and developed by law firms or Lexus or Wesar and so forth. The the ones that are readily available are certainly reasonable to use for certain tasks. So there's nothing wrong with using it. The failure, the ethical, failing, if you will, is in the failure to verify what the AI is producing for you. And that's the lesson. Right? I want the both the lawyers and the nonlawyers listening in this to understand that AI and these chat tools are not research databases. They don't even promise to be that. Right? They are helpful to generate ideas. They're help which is one of their greatest uses. In some instances, they're great in summarizing information and data that you give it to to to, assess for you, but it it it's not a legal database. It doesn't pull from a universe a closed universe of citable law. It's a prediction engine. One of the things I think that that that lawyers and nonlawyers need to understand is what it actually is doing. Is it's not giving you an answer? It's looking in its in in the in the mid it's looking through all of the the the the the billions of sources that it's been given, reference to in its training model to predict what the answer or what the next response might be. And and that's that's interesting to think about. So it gives you an answer, therefore, that in context sounds right, but it isn't necessarily right. And so how do we reconcile, AI with ethical duties? First, I think understanding that AI is a useful starting point, but it's not a final step. Right? As I mentioned, it's great for brainstorming and outlining and to flag potential cases. But everything that you submit to a court and even to a client. Right? Anything as I told my law students, anything you put your name on, that's that's your bar license, that's your credibility, that's your reputation on the line. So anything you do, you must independently verify. And that means double checking, citations even, you know, using the trusted, legal databases that have been around for thirty years, and making sure that the arguments that we make and the assessments and judgments that we give are based on precedent, not on prediction. Right? So I think that's that is the easiest way to stay on the right side of the ethical rules. Yeah. Verify everything. I mean, but it also it's great in some ways because as you as you were saying, like, it can be used for certain things. So if you are trying to condense a large you have a large amount of data, facts in a case, and you need it condensed, you it might help you summarize some of that. Of course, you would wanna go back and verify to make sure if there's any, like, important data points there that are gonna make or break your case. You probably need to make sure that you really truly exist, but it can save you a bunch of time. So I totally agree with what you're saying. I guess another question I'd have on that though is when you're thinking about, like, let's say you're a lawyer, you're using chat GPT or any of these, generative AI tools, are there any red flags in the in the output that's coming out that we should be looking for? Because, again, it does a lot of what it writes does sound, it sounds polished. It sounds it sounds well researched. Any red flags that you can give us just, like, as as general tips? Well, here's the here's to my mind, here is one of the biggest red flags, and it doesn't even come from this space, but it comes from my grandmother and probably your grandmother and everybody's grandmothers, which is if it sounds too good to be true, it probably is. Fair. Lawyers and those who work in the law, we we we we, come to our tasks representing a client, representing a position, and many of the positions, many of our clients' views and arguments are grounded in something in the law or a reason or in in some some policy perspective. But the but the notion that there is gonna be a case out there that says exactly what you wanted to say, that's often not the case. Right? That's why and with those those of us who are lawyers know that when you go to law school, you learn how to synthesize. Right? So you take a number of cases, the statutes, and other materials, and you look at them together. And from that, you can define what the rule is that might apply in the situation. And so I think the red flag is if you put it in and there's a clear easy answer, that's probably not a real case. That is the that is the AI giving you an answer because you asked for an answer. Right. That's a red flag. And the other is, you know, the other red flag, I think, and and this is more of a practice point, is that, you know, is is the tendency because everything is polished and it is it is it is fast is to want to over delegate. So this this this this notion that, you know, rather than have it help you do one targeted task, you know, tell me what the the, authorities on a particular point are saying, what are the poison views, what are counter arguments that someone might raise, those kinds of targeted tasks rather than that, delegating more to the AI than should be delegated. So, you know, having it write an entire brief, that can be efficient, but as we've already talked about, extremely dangerous. And the point is if you didn't write it, you didn't vet it, yet you're still responsible for it. And so one of the duties we haven't talked about yet is the duty of confidence, and that doesn't go away just because you're using a tool that's fast or helpful. That's another really, really important point, the duty of confidence. And if you are, again, you're submitting something that is not real, that is not that is this this case. It's not is just this unicorn case. You cannot believe it, and you're submitted to the court and then arguing for on behalf of that case, using that as your basis for your argument, that is problematic because it could also indicate that you don't know the law and then so you're citing wrong law, fake law to a court. That's a problem as well. What so when you mentioned supervision and verification, just in general, like, you know, verify what's coming out of your the, generative AI tool. So there's a case, Cole versus Ellison. You did reference it earlier, and this is where an AI misinformation expert from Stanford, actually submitted a declaration with fake citations. Clearly problematic, but that especially when it's an expert in that subject or in that field, you're submitting it. When you think about that case and then think about how we could apply it as lawyers today, whether it's a law firm, whether you work in the public sector as an attorney, anyone who's submitting something to the court, how can either a law firm or otherwise really build guardrails to prevent these mistakes? Yeah. That's a really great question, and it raises something that that as lawyers, we have taken for granted. So in prior to becoming a law professor, I was a litigator. I was a civil litigator, and I hired experts all the time. And and I had faith. Right? I'm hiring you of a PhD in some kind of discipline, and I'm hiring you and using you because you qualify as an expert. Right? You under under the the rules of federal rules of evidence or the state rules of evidence, you can provide some information that will assist the trier in fact in understanding the evidence. That's the definition of what the expert does. And I always operated and many lawyers operated, including the lawyers who were the the attorney general of the state of Minnesota in in the Cole versus Ellison case. We rely on the experts. We assume that if they're putting it in their declaration, in their statement, in their testimony, then then it's real and and it's and that we under that they they they would never because their credibility and their reputation relies upon it, they would never be fabricating something or making the evidence better than it is. I just took that on faith that that is how the world works, and, apparently, as this as Cole versus, Ellison illustrates, it's not. And so I think the takeaway for lawyers and those working, with experts is the guardrail is having conversations about, you you you prepared this declaration and you're relying on these sources and, you know, we we live in a world now where AI is a powerful and important tool. These sources are correct. Right? This is if I went to that scientific journal and looked that up, it would say what you're gonna say. So these are additional conversations I never had to have with experts because this was not a risk. It I mean, maybe it was a risk and it just I didn't appreciate it, which is probably actually true. But now we see it. It's a it's a greater risk. The good again, because I was earlier talking about how lawyers are seduced by this technology, so are experts apparently. Right? People and and and what's interesting about Cole versus Ellison, this isn't any expert. This is an expert on AI misinformation who is giving an opinion on how hard it is for people to detect what a deepfake is. So this is someone who knows better, who knows that things are that they're hallucinations and fabrications. And so the the the advice for lawyers is, you know, you're going to be responsible for what your expert says and does on behalf of your client, and so you need to verify what they're doing to the extent that that you that you can. I mean, I think that's the best guardrail that it is and have a really good expert contract with I mean, your your agreement with them should insulate and protect yourself Right. From Right. Whatever cost may may be applied at or penalty or sanction may be be applied if they if they are not relying on things that are that are real. And that's and that that really, really, like, I think drives it home for me too. Like, as you said, you have you were a litigator in a previous life and had relied on experts, and I think a lot of people who are on this webinar do that as well. I I did when I was practicing. I was in the public sector, and, you know, we would bring experts in all the time. And so there you would just you would definitely go on faith that they were an expert. You would obviously, like, you know, present them to the court as experts. They would explain their expertise, and then they would give their render their opinion based on that. Now we have a little we have an we have another factor here we have to think about. So that's going to be interesting, moving moving forward. Before we I wanna go into the next topic around around how courts are responding to the risks. But before we do, you did mention the duty of competence, and I'd love you to just talk about that for a moment and talk about how that comes into play around generative AI as well. Yeah. So the duty of confidence, is the duty that that lawyers are are are embodying when they when they hold themselves out to represent their clients on on whatever the topic or subject may be, and it's the the the idea that you have thoroughly, researched and investigated not only the facts but the law and that you are performing in a way that is, that that complies with the standards of the legal community. So it's it's, if we think about, for example, malpractice, the duty of confidence is often one of the duties that lawyers fail in a in a malpractice case, which is, professional negligence, if if you will, for lawyers. And it's the idea that that that they're that the the the way that they are representing themselves in terms of the law, in fact, in strategy, if you will, that they have not, they have not complied with that. They have not upheld that. In the criminal space, it's it's, for example, you know, we call it the effective assistance of counsel. The idea that what the what the particular community is and and the arguments that a reasonable lawyer would make, you had made. And, how that connects with with AI specifically is when you are using generative AI to, develop an idea or to find a case or to brainstorm a strategy or to draft a section of a motion, you are in in you are working with it to, as a way to uphold that duty of competence. So if you use it and it does something that is false or hallucinated, then it is if you you are not upholding that that duty because whatever it is doesn't exist or isn't quite right or is it jurisdiction specific or whatever that may be. And and so it it it it is like any other part of a lawyer's toolbox, where we are upholding the values that represent the the legal community and the issue and that are consistent with the with the norms in that particular area. And I will I will say one of the other challenges around, using AI is not is relatedly, is the duty to supervise. So lawyers have a duty to make sure that, paralegals, paralegal professionals, legal interns, summer associates, you name it. Right? Anyone who works in this space where lawyers work, that we are required as licensed lawyers to make sure that they comply with the rules of ethics also which include all of the other duties we've talked about. And one of the really interesting questions, that has come up is when you use AI in the way that some people have used it, what are you supervising? And and are are you supervising something that is the equivalent of a paralegal or are you are you using it more as a tool? Those are a lot of unanswered unanswered questions, especially because AI does people are I I I talk to lawyers all the time who say, oh my gosh. Now I didn't have to hire an extra summer associate or an extra intern because I have AI that is you know, I'm having a conversation with AI like I would a junior, you know, a junior law student or or someone a a a new associate. And I don't need that person because I'm just using it the AI that way. Well, the question becomes, are you supervising it? And what does supervision Mhmm. Your supervisor look like in this space? So it touches on our range of duties, including including those. Wow. That's it's it is fascinating, and I know that area is also something that we're going to start seeing more and more of. So there's more to come on on that, but interesting very interesting topics that are arising out of this. But let's talk about how the courts are actually responding to some of these risks, challenges, etcetera. So the judicial regulation of AI, I think we've seen, from what I've seen and and you definitely have a better overview of this, but courts are taking very different approaches around AI. So we've seen outright bans. There's a case, in Illinois. I think you'll probably mention where there was an outright ban that or two judges saying, okay. If you're going to use it, you can use it, but there's a disclosure requirement, and there's this certificate that you have to present to the court. And so very interesting and very different different ways. Do you think, professor, that disclosure is is enough if you if you follow the judge in the Pennsylvania case who said there are disclosure requirements, or should courts just ban AI filings altogether? Like, where do you really land on this from what you've seen? Yeah. That's that's an important question because as you mentioned, courts are all over the map. And, you know, I think disclosure is an important first step because it encourages transparency, it promotes accountability, it forces lawyers to own the work product they submit, which it goes to the duty of confidence we're talking about. Because if you if you know you have to disclose the use of AI, you are more likely, I say, I hope, to check those citations and to verify that out before you hit submit, and that lines up with the duties of confidence and candor, etcetera, the ones you've been talking about. But on its own, I'm not sure the disclosure captures the entire picture. It certainly doesn't regulate how we vet AI material. And that said, right, I I also wouldn't want to go too far the other direction. So I and and I don't believe an outright ban on generative AI is the way to go. It it's a quick fix especially when you have when you see high profile cases like Mada and some of the others, that are out there, but it it it's problematic. There are aspects of thinking about a ban that that truly as a lawyer and as a law professor trouble me. In first, there's the first amendment concern. Legal writing, motions, briefs, oral arguments is a form of professional expression in speech. And while courts can regulate it to a certain extent, think, you know, you can require disclosures, they I I do not believe that you they can ban these tools unless the restriction is narrowly tailored and serves a compelling government interest. And this has not been tested yet, but I think we're gonna we're gonna get there. Right? In these in these courts, and it's not just I mean, there there are a number of courts that have banned it. Some state courts, some federal districts have banned it. I think it's gonna be tested. My personal view is that a blanket ban on its use, regardless of how it's used or whether it's supervised, is probably gonna be viewed by the courts as overbroad and potentially unconstitutional. And then there's another concern, not in the civil space, but in the criminal space, and that's the sixth amendment right to effective assistance counsel that we already talked to. Right? Public defenders, appointed attorneys often are under resourced and overwhelmed. That's kind of the space that they live on. And if they can't use AI tools to help with task or legal research and drafting motions and so forth, their ability to be confident, might, and and timely, right, is going to be challenged. And so so in my view, I don't think that it would be wise to ban the use. I I think it's important to regulate. So I think where I land is on on disclosure. And I and I also think, you know, as I said with with respect to disclosure, disclosure, incentivizes the appropriate and ethical use, and we wanna do that. I am I not notwithstanding what we any impression I have given before, I'm actually pro use of these tools, but we have to do it thoughtfully and carefully. And the courts, I hope that the courts more courts don't go in the direction of an outright ban. And it seems to be that the trending is towards something more middle ground, like, whether it's disclosure or other ways that you can put checks on it versus this outright ban. Is is that correct? I think that's I think that's true. I think the initial reaction of some judges is just, I don't I don't want this. I don't I mean, everybody's hard enough. I wanna be able to trust the extent I can trust that this is a product of of of human purely human, initiative development and creation. Even some of the judges, that initially banned it in the federal court have stepped back from that now. So I think that's the direction that we're heading, and I think that's the better direction. What about so we we talked about the court's, authority to to regulate this, what what we called inherent judicial authority. So courts have affirmed their power to regulate conduct that would could be considered threatening integrity. So there was a case, the Hudson case, which you may you may want to talk about, where, you know, the court has affirmed its power to do this in the in the big broadly. Could judges sanction lawyers for failing to disclose AI use if they do not even have a standing order? Yes. I think that they can. So so I think that that the broad, you know, under Hudson, under rule 11, under, 28 USC nineteen twenty seven, there are broad rules in the federal courts, and most states have similar counterparts that give the courts the ability to sanction lawyers for behavior that undermines true seeking the true seeking function of the court. So, yes, judges can absolutely sanction a lawyer for failing to disclose that they relied on AI and drafting anything, especially if the output ends up being, you know, fabricated or misleading. I I don't think that they actually need a specific order to do that. They have they have broad authority to to to do that. K. Well, let's flip that script a little bit and talk about judges using AI tools, like JTGP, Gemini in their work, as we saw in the Ross case. Tell us a little bit about the Ross case because it's very interesting. It has an interesting fact pattern. And then I wanna talk about how you think this practice might actually impact judicial decision making moving forward. Yes. Now Ross is an interesting case. So it's out of DC. It's a their their, DC Appellate Court, and it involved a situation where an individual had left, their pet in a car, and they ended up getting a, they ended up getting cited for leaving a pet in the car, with the windows rolled up. I don't I don't know exactly how long the the the dog was in the car, but this was one of the first instances where I saw there was a there was a a majority opinion, a concurrence, and a descent. Mhmm. And, the all three use chat around this question about, what, you know, what is reasonable, what what do what what is common sense knowledge about how long you could safely leave a dog in a car. And in in it was it was interest I mean, it's it's interesting that this was done. For those who haven't read the case, it's worth it's worth a read because you can see the interplay of of the ways that the three, three opinions are using it and what they're saying about it. The the dissent, talks the most openly about its use of figuring out what's reasonable, the reasonable amount of time, and the different prompts that were given to the different, chat tools to determine the definition of reasonableness. But here here's here's my concern about the court doing this is that it you know, it's using the generative AI as if in in in a traditional sense in the way that courts would prior to the use of generative AI, prior to its existence, one might look up, other case law or even in legal dictionaries or in other spaces, where there might be research to figure out what means, you know, reasonable time, what that means. And that and and the court was using the AI in that way, in the traditional sense. But the but the problem is, and I'm not sure the court appreciates this, but the problem is is that the AI tool that was used doesn't pull directly from the authoritative legal sources Mhmm. That they would have used before. Right? It's generating predictions based on patterns of language, and that means, right, it can mistake doctrine, omit and, like, omit important context and invent things, just, as it's done for lawyers who use it. Right? So when a judge relies on the AI without whether or not they disclose it, you know, it does, raise questions about the output because of the way the AI, chat bot has been constructed. I mean, that ultimately is it's I wouldn't even say it's that's ultimately one of the failings of using it is that it is a closed universe of information. And and and it's not an answer, it's a prediction. And so what how long is reasonable to leave a dog in a car, the answer is based on the context of the number of times those three u words are used together. Right. If you think about it, that's not an answer. That's a that's a prediction. And yet, when you read the opinion, the court is saying, here is here is the, here's the answer. So that's that's a little problematic. And the best thing about that case is they were transparent about their use. And that and I remember I recall from reading parts of that that one of the judges in their opinion had said that around that ordinary use, around leaving a dog in a hot car is dangerous. Like, what is considered, like, reasonable in that situation and that went to the defendant's intent. Did she intend to do this or should she inherently have known that it is dangerous? Right? That the judge cited that that person looked it up themselves on a chat sheet to see then gave it to their law clerks and have three different law clerks look in three different places. And so I mean, it is interesting because, like you said, in the past, when we're looking at what is the common let's see if we took any terminology, what is the common use of the phrase? You would go to sources, whether it's case law that's already interpreted that, or you could go to Webster's dictionary. Right? I mean, there's just, like, a the gamut of what are considered authorities, and then you could use that for your reasoning. Here, like you said, it's a predictive model using generative AI. So Well, and and it's it's yeah. It's interesting. It's it's so I would look at Black's Law Dictionary. I would look at other legal encyclopedias. What the AI is looking at is Wikipedia. Right. And, I mean and and it's and it's it's crowdsourcing the answer, which which which is which provides some information. Right? I I wouldn't discount that and say it's, like, valueless, but it's not it it's not doesn't have the same legal value or authoritative value as the other sources that lawyers have always used. And yet it's been elevated to this level of, you know, they're not equivalent. They're not. Right. Exactly. Exactly. And we saw this in another case. I believe it was the Snell case, and there was a concurrence there. This case involved a civil dispute involving a trampoline and the ordinary use of landscaping, like someone landscaped and made a trampoline in an area that would be a pool, I think, like a swimming pool. But they used AI to help interpret the ordinary meaning of landscaping. And I mean, that's also another, like, another example very similar. But, I mean, again, is this a legitimate use of AI or are we really risking outsourcing judicial reasoning to things that are possibly crowdsourced ideas. Right? Yeah. It is a risk. I mean and and what's interesting is I would invite your listeners to run run a search, a a PGN issue, right, and run a search on a a a chat tool and and ask it a question. And, most or many of the chat tools will show you what they're looking at to get the source. And it's always amusing to me because I'll look something up and it'll show, for example, Reddit, it'll show Wikipedia, it'll show these other things. And I'll think to myself, I would never make a decision based on what was on Reddit. And then maybe maybe that sort of, like, categorizes me as as I I don't know what, but I wouldn't. I would look at it maybe. And but I wouldn't say, well and this is how I define fill in the blank. And yet, that's exactly what these tools are doing, and they're the other thing I think it's important to appreciate when you're using them is the information your output is so limited by the sources that it used to create the answer, and those sources are as imperfect and biased as the Internet is. So I think if you told a someone who's who who is, an an an you know, a college educated adult who is familiar and has used the Internet for most of their adult life to do things, you know, in Google searches or using search engines and and you you ask them, do you do you always trust what you see in here? Most of them are gonna say, no. I don't. Right. Because, I know how the the entities, the organizations, the companies that support these, how they impact, how they use the algorithms to elevate certain ideas. Right? I think that what I've just explained is it is, for the most part, accepted, right, by most people, even people who are not Internet skeptics. And yet, that same issue is built into these generated AI tools. So the same level of skepticism, is required in using that. And I I hope the courts appreciate that. Mhmm. And I hope at the very least that they disclose their use. Now there are a number of of states right now that are in in the process. I mean, Illinois, New York, California just put forward their proposal about the internal use of AI, and I encourage that. I hope more courts, create their own guardrails about how it can be used, within the, you know, the canons of judicial ethics, which bind all of the courts. So but but, you know, I can only hope for the courts, what I hope for lawyers is that they're they're thoughtful and that they adhere to their ethical obligations and its use. And and just not to put too fine a point on it, but, you know, as you said, as with lawyers in court, with judges using generative AI, if they're using it and if possibly they are citing missciting case law, fake case law, etcetera, I think the same, like, ethical rules apply here. Right? Like, we don't we don't necessarily have, like, exact remedies. Like, we I don't know if we have examples of judges who have used it improperly and then had and then been sanctioned for it or reprimanded. But it would still be the same like, they still have the same ethical duties as, you know, as to the court to as as as members of the court to the state bar, etcetera. Is that to the federal bar. Is that is that correct? Yeah. It it it it is. You know, when you look at the judicial canons, the they're not entirely clear on this, but but I I would argue in most judges, I think, are are are are willing and faithful to canons around being, forthright and, and, and grounding their decisions and the evidence in the law. So it's it's there, but it's not as clear as I think it is for lawyers. Right. There are a lot of norms around judges that are built up expectations about behavior and and so forth. So so it's there. I I would say in terms of remedy, the the the best remedy is appealing. File an appeal. A judge cites something that is not the law. Sure. Your remedy is to appeal it to the next court. That's the the clearest, way to to course correct for that. That makes sense. That absolutely makes sense. Well, I one other question just around maybe not so much the courts, but just in general, you've raised concerns about bias in training data in general, like, how these models are being trained. How can judges ensure that AI tools don't perpetuate disparities in legal outcomes? I don't think they can. Yeah. I mean, I mean, you know, if they use it and rely on it without taking a step back and comparing the output to their lived experience as judges, they can't because there is, there is no because of the way the large learning language models are built and the information they're trained on, the bias is sort of built in to the way it responds. If so the judge has to not wholesale adopt the out the whatever the outcome is, but to evaluate it in the same way it would evaluate, information that it was going to endorse and adopt as its opinion from any other source. Now it's interesting because I've done experiments with my students around asking asking these chat tools to evaluate particular cases and fact patterns and and ask the the chat to say, you know, this is a motion to to compel, give me the best arguments. And what's interesting is what the chat doesn't recognize, things around such as bias and characteristic. And then to sort of catch the chat and say, well, what about this factor that this person was a person of color or this person and the chat will honestly say, oh my gosh. That's great. That's a great point. I missed it. And then and then you say, okay. Well, now let's correct for that. And it will correct itself to a point. But but even still, there's still again, because of the the the way that they're built, the data they've been taught upon, how they've been trained, there there there are limitations. So the good news is is there's always a space for human judgment. Right. The to sort of backstop. Right? Whatever is that is coming out. And that's true of judges, that's true of lawyers, that's true of everyone. I think that's that's some really good practical advice. I think I think the the bias in training data is is certainly an issue. And I at this point, it sounds like the reasonable course of action certainly is to at least, you know, a lot of verifying. There's a lot of verifying going on, and I have noticed as well when you push back on any of these generative AI tools about something that's an omission, or something that is maybe counterintuitive to what you would believe and you tell it that, it will automatically course correct. And it will like, as you said, it'll apologize, and it'll and it'll come up with, with whatever that that reasoning is, based on what you're saying. So it also, in many ways, is trying to please the user, which can be problematic too. So I think looking at everything in totality and realizing that this is an a tool you're using in aid of whatever your task is, and just using it as that is probably the more reasonable way to go about it versus, like, blind faith. But only I know we only have a little bit of time left. And so just a few more questions for you, just on general on general topics related to this. So a couple of just, like, random questions here. How should when we're talking about training, how should law firms or even, you know, if you're a pub if you're in the public sector, if you're working at a DA's office or a a public defender, and you have junior lawyers, how should the firms or the entity train junior lawyers to use AI ethically? What are what are your thoughts there? Yeah. So this is, I think, one of the most urgent issues facing law firms and and other entities that are using young lawyers right now because these, young lawyers, recent graduates are already using generative AI whether your entity knows it or not. And if they aren't trained properly, the risks, ethical, legal, reputational fall ultimately on the organization. And a lawyer as well. So how should we train them first? I would start with the mindset, not just the mechanics. It's not enough to say don't copy and paste from CHAGBT. Junior lawyers need to understand why. And so this means grounding AI in the in the core duties of, you know, confidence, supervision, and candor to the tribunal. Every time they use AI, they be they should be asking, do I understand where this came from? Can I independently verify it? Would I be care would I be comfortable, standing behind this idea or this case in court or in front of a client? Second, I think, organizations should offer clear scenario based training, not just abstract policies Mhmm. Or real world exercises. So, you know, here's a motion drafted by AI. What's wrong with it? Right? Here's a hallucinated case. How would you how would you catch it? The goal is to teach lawyers how to treat AI as a tool, not a source of the truth because that's where we're really seeing these mistakes. And then to continue to develop the skeptical editorial instincts that we would have if the information came from another person, like a human being. Right? We are skeptical of others. We should be skeptical of this polishy shiny thing. Third, I think there should be structured oversight. Supervising attorneys need to know when and how a I I is should and will be used by their teams. This might require lawyers, you know, to flag, have them flag AI assisted work or submit both the prompt and the right the the raw output along with the final draft. Right? It's no different than supervising the work of a junior associate. You are still ultimately responsible for the final product. And finally, I think this needs to be framed really as a competency issue, which we've been talking about already. It's not just compliance. Knowing how to use AI responsible responsibly is part of being an effective lawyer right now. So, knowing when not to use it and how to spot its limits is just as important, really. And so organizations that get this right, will just avoid mistakes, I think. They'll develop a more thoughtful, agile lawyers, but it starts, with training that treats AI not as a shortcut, but as a tool, I think, that demands judgment, verification, and ethical awareness, basically, every step of the way. And that's good. And start it when there are summer associates. Start it when their law clerks have started early. And I guess even for lawyers who've been with organizations for a long time that have been practicing for years, it's important to have this training even for them at the I I would I would think. Mhmm. I agree. Okay. Couple more. You you talked about this a little bit, but just to reiterate again, like, the I know you had said what what surprised you about AI and our reliance on it when you started to research, put your put your research into this. But is there what's the biggest misconception right now in the legal practice about generative AI? Is it that what you said, that the reliance on it or something else? I think the biggest misconception about generative AI in legal practice is that it's a reliable research tool like Westlaw, but with a better personality. Right? Westlaw, we seem to have a conversation with. Right? But that's not what it is. Generative AI isn't retrieving law from a closed database. That's Westlaw. That's Lexus. Generative AI is a a predicting, you know, predicting language based on patterns in the data that it's trade up trained on. So it can sound authoritative, even brilliant, but it doesn't actually know what it's saying. It really doesn't know. So let me give you an example of this. So I had a law student who told me that they figured out the summary judgment standard. This is a civil procedure class, and it it was taught to them by Chachi and GPT. And I said, that's great. Show me the case. And they pulled up what looked like a perfectly formatted citation complete with docket numbers and lawyers and law firms and so forth. The problem is when I said to the student, now let's get on Westlaw and find the case, we couldn't find the case at all. Mhmm. But it read, you know, it read smoothly and the quest the student didn't question it. And that's the trap. Right? Generative AI is confidently wrong. And it it performs legal, you know, it performs legal authority incredibly well. It's sort of like I think of it like Tom Hanks playing a supreme court justice. He looks the part. Right? He sounds convincing, but at the end of the day, he he wouldn't be writing the real opinions. He's not a lawyer. He's never been a judge, and, you know, neither is chat GBT. Right? So when lawyers treat these tools as if they're really sources of truth instead of drafting assistance at their best. That's where the risk comes in. That's where we really get in trouble with, the the ethical rules. The right way to think about generative AI is like a very bright but completely in unsupervised intern. You know? Brainstorm, outline, even tighten up language. It writes beautiful emails. Right? No question. But you still have to do the thinking. Right? That misconception is you can just put it in there and solve it, and you still have to check the law. Otherwise, your beautifully drafted brief might turn out to be legal fan fiction, essentially, is is what it could be. And you're you know? And and you never know what that will lead to in your next, court appearance. I mean, I think when I think about the lawyers in MADA, when I think about the Stanford expert, I think what's going on with them now because what they essentially did is they put themselves out there, representing clients or a level of expertise, and they gave over their their personal professional credibility to a tool that, you know, is, is is always confident, always certain, but seldom or sometimes not right. Some yeah. I oh, exactly. And it just I think about it, you know, when I was practicing and I had cases I was practicing, in in the criminal realm, and so I would have to make arguments. And so one of the things I would do, just thinking about how I used, how I would how if I was practicing today, how I would use it. When I didn't have chat two p t, I would bounce ideas and theories of a case and presentation of how I present the evidence off of my colleagues. And that was great when I could get their time. We could sit around and bounce cases around. But now if I had something like a chat GPT, you know, I have, you know, tested it out and tried to do that. And it does give you some really interesting theories. But theories that if I were gonna present today, like you said, I wouldn't absolutely need to make sure it's grounded in the facts of my case, that it's not making up facts that make it sound better, and that any case while cited certainly jeopardize what was called jeopardizing, but I think it's still used. But, you know, so but checking the citations, and be ready for any case in the brief that you're going to submit to the court for that judge to call you out in court and ask you about it. Like, that's that's a key, and that was something that even way before CHAT GPT when I was a law clerk and then later a prosecutor, it was very important that I was anything in my brief, I knew I would be I would have to stand behind it and and argue and say explain why I put it there versus Right. Not. So I think those are some of the biggest biggest things on my end as a former practitioner as well. Last question for you. We talked about training and law clerks, and and some are associates, but just a few if you could just leave us with a few actionable strategies, you know, a AI is here to stay, generative AI. But how can we truly adopt it responsibly, use it in our legal practice, and confidently just sort of lead our profession forward, which we are clearly headed forward with this. It's going to be used. So how can we do this confidently and responsibly? Right. That's a great question. And and and and my thoughts to end on this is is within ourselves, we have to develop a sort of a a culture of verification, a mindset of verification. I think these tools are fabulous. I use them all the time. But make no mistake, it's not a shortcut. It's not the shortcut you think it is. It's just an additional tool. And so I my my thought is use them, but use them in tandem with other things, like your own, professional, experience as a practitioner and a lawyer. Test it against what you know, having developed, you know, no matter how many years in practice you have lived and experienced different scenarios. Test the output against that, and then also test the output against the the the legal the the sources, the various sources that we've used for years, Westlaw and Lexus and other internal sources. Test it as well. Use it when you can. It's great for brainstorming. It's great for bringing out, complex and interesting ideas, but it can't be the only thing. It's a partner. It's an assistant, but there's still it's there's still you. Right? And I think going forward, sort of integrate it into our practices rather than have it be a substitute for what we used to use. It's just not it's not an acceptable reliable substitute. If I were a futurist, I would say that someday it might it's getting better and better, and it does get better and better, but there's there's always gonna be limitations to it because of some of the things we talked about earlier. So it's it's a subs it's not a substitute. It's a helper. It's a helper. That's a that's a great way to to put it. Well, professor Delfino, thank you so much. This was truly invaluable, all of your insights. I can't wait to hear them in person, in October, but the key takeaway is, like, AI is not necessarily replacing lawyers here, but lawyers who use AI responsibly may replace those who don't. Like, you have to have some confidence here and and all of the tools and the takeaways you just told us are how we can do that. And like I said, if you want to see professor Delfino live, at Everlaw Summit in October, this October in San Francisco on her special session of generative AI in the courts, go to Everlaw forward slash summit. Get more information there. And then for future reading, we will share some of professor Delfino's materials with this webinar, and some of the the white papers and some of the ADA rules. We'll send you our opinions as well. You'll get those. You can reach professor Delfino at rebecca.delfino@ls.edu for any further follow-up questions. And from all of us here at Everlaw, thank you for joining us. Appreciate all of your time listening to our webinar today. And, professor Delfino, many thanks to you. Appreciate your time. Thank you for having me. This has been fun.