Law-Backed AI: Aligning Artificial Intelligence With Human Values – Damien Riehl

Episode Notes

Transcript

Rocky Dhir:

Hi, and welcome to the State Bar of Texas podcast. So June of 2026, I got to do something kind of fun. I spent a few days in Houston, Texas at the State Bar of Texas annual meeting. It was a great experience, but I got to tell you, on Friday of the meeting, we had a very special keynote speaker. Damien Real is one of the foremost experts on AI and the law. And I know we talk a lot about AI. We’ve been doing a lot of AI episodes and you hear about AI everywhere. It’s everything the lawyers want to talk about. And I get it. Look, it’s a big deal. We’re always talking about, is it going to take our jobs? Is it going to replace us? AI hallucinations. We talk about all things that could go wrong. Damien’s keynote speech was really focused on how to do AI responsibly and how it can be used to actually further our practices, how to enhance us as lawyers.

I thought it was pretty compelling. So we’re going to do something a little different this month. I thought what better way to talk to you about Damien Real than to actually let you listen to Damien’s keynote address? So we’re going to replay that for you right here. Take you right back to that room so you can hear everything Damien had to say. I also had the opportunity to sit down with Damien afterwards and we had a little video session, a little video podcast session. You can catch that on the State Bar of Texas YouTube channel. But for now, sit back, relax and hear a little bit from Damien Real about the responsible use and the potential of AI. Enjoy.

Announcer:

Now is my privilege to introduce our keynote speaker. Damien Riehlis a lawyer and technologist with experience in complex litigation, digital forensics and software development. A lawyer since 2002 encoder since 1985, Damien clerk for the chief judges of state and federal courts, practicing complex litigation for over a decade, has led teams of cybersecurity and world spanning digital forensics investigations and for the past decade has built AI backed legal software. At Clio, Damien aids in the expansion of various products, including AI backed technologies into a billion document dataset from a hundred plus countries all to improve legal workflows. Among his many professional activities, he chairs the Minnesota State Bar Association’s AI committee, which is establishing an AI sandbox to promote access to legal services. Damien is also working with the Colorado and Arizona courts on how AI can approve access to justice initiatives. He has a wealth of information to share, so please help me welcome Damien Reel to the stage.

Damien Riehl:

Thanks. These are exciting times. I speak fast. Why listen to me? I’ve been a lawyer since 2002. I clerked for the chief judges of the state and federal district court and appellate court. I worked for Robins Kaplan for 15 years representing Best Buy in much of the commercial litigation victims of Bernie Madoff. It also helped sue JP Morgan over the mortgage-backed security crisis. I’ve also been a coder since 1985. This is me and this is my commodore 128 I used to code. My mom loves this photo. So the number of lawyers that are also big law refugees that are also judicial clerks that also know technology, there aren’t a lot of us in that Venn diagram. After about 10 years of practice, I said to Thomson Reuters, “Here’s legal tech that I think could change the world. You should build it and hire me. ” And they were dumb enough to do that.

So I led a team of a hundred programmers and 50 lawyers building this really big thing for TR. I left that really cool job to go into cybersecurity where the biggest thing I did was that Facebook hired me and my company to investigate Cambridge Analytica. So I spent 14 hour workdays working with their data scientists and my former FBI, CIA, NSA guys, and we figured out how bad guys use Facebook data. At the end of a 14-hour workday, I said to my buddy, Do you want to break music? He said, “Yeah, let’s break music.” So we’ve now copyrighted 471 billion melodies, every melody that’s ever been and ever can be and put everything in the public domain to protect you stole my melody lawsuit defendants. This guy says that I rock. My buddy says that Elon Musk has horrible, horrible judgment in all areas, including this one.

So I was on CBS Sunday morning. You made a year old enough that you know what that is. The Atlantic did a thing about me. I was able to be Jeff Hinton just a few months after this and he got the Nobel Prize for his work at artificial intelligence. These days, my day job is with Clio. My playground is a billion legal documents. All the cases, statutes, regulations, motions, briefs, pleadings where we get to be able to do extract data science from those and run AI across all those things. I also help lead legal data standards. The Financial Times of London named me one of six champions of generative AI innovation. I lead the MSBA’s AI committee. I also work with bar associations around the country, including Spring Courts of Colorado, Arizona, and actually Louisiana too, and hopefully Texas if you’d like to. This is a way that we do AI right and name me in it.

And I speak a lot. I was just in last week, I was in New York, Boston, Washington DC just came from yesterday, San Francisco. Tomorrow I’m going to be in Minneapolis, going to Madrid by London, back to Madrid and back to Minneapolis. Turns out 251 talks over the last bunch of years, talking to judiciaries, talking to the state. I have firms, corporate law departments, and I talked to the Vatican last November. So the idea is that we need to align AI with human values and it turns out that the Vatican has a lot of human values locked behind the archives. So the idea is that maybe you scan a lot of those documents to be able to put those human values into AI and it turns out that doing that. I’m working with my favorite vibe coding priest. His name is Father John Daracio and we’re doing data science on the archives to be able to ingest those into AI.

So I speak a lot and here I am talking to you. Of course, AI is everywhere in the news these days and what AI is matters, but it matters less than the fact that we’ve gone a remarkable amount of time in just a few years. GBT three and a half in December of 22 beat about 10% of humans on the bar exam, but that’s not nearly as impressive as GPT-4 beating 90% of humans on the bar exam. And that’s worth repeating. GPT-4 beat 90% of humans on the bar exam. And a lot of people say, “Well, that’s fine because the practice of law is not the bar exam.” And I agree. But the question is, how much of what we do every day as lawyers and that we bill our clients for every day is actually far easier than the bar exam. And that’s why I’ve been speaking 250 times around the world.

This is game changing. And the fact is that AI didn’t stop at 90%. This is my friend Matthew who kept tracking it. AI now gets 100% on the bar exam. It aces the bar exam. How many humans ace the bar exam? AI will continue going better and better. This is another view of the passing the bar exam. You can see that ChatGPT, Gemini, Claude, even DeepSeq, free and open source. And Grok does it too. And by the way, Grock, not only is it beating the bar exam, but it’s also 50 cents per million tokens. They might say, “What’s a token?” And our brains think about 10 million tokens each year. So our year’s worth of thinking is about $5. Years worth of thinking is about $5. This is human baseline. AI already beat us in things like reading comprehension, unreasonable answerable questions, lawyer things. GBT2 was about a level of a preschooler.

GPT3 was the level of an elementary schooler, which turned a high schooler, which turned into beating PhDs in physics, which turned to beating some of the best mathematicians in the world. And the question is, where’s each of us on this and how soon before AI just blows past even the best of us? This is the ArcAGI Prize. People didn’t think that it would be beat until 2039 and even then it would just be about 50%. But it turns out it didn’t take until 2039. It literally took three months. And now we’re on to ARCAGI2 and that got beat and now we’re onto AGI3 and we just keep going and going and going. This is human PhDs in their field and it used to do very poorly, but in 25, it beat human PhDs in the field. And not only did it do better, but you can see there used to be $50 per million tokens.

Now it’s 50 cents. So the intelligence is going increasingly up while the cost goes decreasingly down. And what does society do with something where the intelligence goes up and the cost goes down? They use more of it. This is Jebin’s paradox. As the tokens prices fall, look at the demand just going through the roof. This is human lawyers against machines. This is a double-blind study where they gave a human lawyer task, gave the machine a task, and then mixed them up, gave them to human lawyers as review. And the lawyers that reviewed it said that humans got beat by the machines. Not only did it get beat, but ChatGPT was both more reliable and more useful than the humans and Gemini out of the box was more reliable and more useful according to these judges that are human judges that were partners in law firms.

So the question is how much of what we do every day is truly judgment or how much is it actually intelligence for which we’re paid judgment rates? That is the question of our time. That is we have now this machine that is maybe better than me, maybe better than you in intelligence. If you think about the dumbest person on the left and the smartest person on the right, there’s a bell curve. And when you think about that, where’s AI in that bell curve? And right now, people estimate AI to be 130. That’s 98% of humans are not as smart as the AI. That’s today, and of course, how soon tomorrow, before it goes to 99%. Then that’s for humans. Now think about lawyers and how much legal work actually requires the 99th percentile and how much of paralegal work, first year associate work, how much access to justice can be done with that less than 99%.

And by the way, what I said was actually not truthful, that was truthful two months ago. Now it’s at 137, which is now it’s at the 99th percentile and it’s just going to keep going up and up and up. So how are agents being deployed? AI agents? It turns out that coding is the top one and look at where legal is. Turns out this is a lot of opportunity that people in this room could be able to say, “Hey, maybe I can use AI to help access to justice.” Where is the venture capital and private equity money going? Turns out number one is coding. Number two is legal. Why is number two legal? Because we have a lot of value in this room that the venture capitalists and the private equity say, “Hey, I want to capture some of that value with AI.” Turns out that there are 1,000 legal AI companies.

Let me repeat that. There are 1,000 legal artificial intelligence companies that venture capital and private equity is flowing into. And they’re saying we need to help access to justice and we are going to do this whether you like it or not. This guy thinks that maybe we should use this to be able to help access to justice. To be able to smooth up the mismatch between how much legal need there is and how expensive we lawyers are, 92% of legal needs are unmet because we’re too expensive. I don’t agree with him on everything, but I certainly agree with him on this. If we as lawyers are not ashamed of that 92%, we should be. And if we’re not going to use AI to be able to help that 92%, shame on us. Let’s listen to our chief justice. So we’re going to be talking today about what litigants are using, that is both the lawyers that are sitting in this room and your clients.

That is pro se litigants are using this all over the place. And of course, decision makers are using this. I’m working with courts around the country that are saying we are getting a deluge of these cases. We need to use AI to do this. The American Arbitration Association has arbitration bots. You upload your documents, upload mine. 20 seconds later, the bot gives a decision. AI is a tool that solves problems, but we often think of them in that order. We should be saying, “What are the problems we want to solve? What are the tools to solve that problem?” And it turns out that AI can solve a lot of those problems. I want AI to blank. What fits in the blank? Usually an AI could be able to answer that question. When I was a litigator, Robins Kaplan, the blank I wanted to say is I wanted to be able to create a bunch of counterarguments in good facts.

So what I did is I took an actual lawsuit. There’s a case where a bunch of coders sued Microsoft and OpenAI and the lawsuit was over a bunch of code that they uploaded to GitHub. They ingested that into GitHub Copilot and the coder said, “Wait, all of that is MIT licensed. So by ingesting it, you breached our contract. There’s a breach of contract laws here.” So I thought it’d be fun to have ChatGPT argue against its owner, OpenAI. So what I did is I took OpenAI’s motion to dismiss and I uploaded it to ChatGPT and I said, “Create a bunch of counterarguments saying how we’re going to win as coders.” And it created a bunch of counterarguments and it said OpenAI argued, “Hey, you as coders lack standing.” And ChatGPT said, “No, no, no. We have standing that are directly linked to defendant’s actions.” And then I said, “For each of the claims, give me the elements needed to prove those claims.” And it gave me the actual non-holistic elements saying that it needs to be injured, that the defendants caused the injury and a court can redress that injury.

So cool, give me some facts that if true, would be able to prove each of those elements and do that for every single claim. And it said you could argue that your injury in fact was, your tool was so widely used that it decreased the demand in my coding. Injury, in fact, yes. Causation, yes. Redressability? Yes. You could argue that the coders lost a big contract with a big client who used OpenAI’s product instead. Injury in fact? Yes. Causation? Yes. Redressability? Sure. I said, now create a questionnaire asking questions that if true would help the plaintiffs win. And here’s the output. Hey, coders, have you suffered economic loss? Can you quantify that loss? Lost revenue, lost business opportunities. Give me documentation of these things. Can you show me a link between the dates that you lost your business and when OpenAI started using your code?

These are all good questions that will help me win and everything here took me less than one minute. So the question is, how long would that have taken your associate? And if your associate charges 500 an hour, how are you going to bill for 45 seconds with a prompting? And what if you don’t have to be a prompt engineer like I am? Everything I just showed you and about 30 or 40 other things is just baked into the software. As soon as you upload the motion, no prompting necessary, it gives you all those things and about 40 other things, no prompting necessary. And that’s true for you as lawyers. It’s also true for your clients, that as we work for the biggest corporations in the world use our software. We’ve talked about the death of the billable hour for decades and I think that’s a mischaracterization.

The billable hour is not going to die. But what’s going to happen is the 90% of billable hour and 10% flat fee that most firms have, that 10% flat fee is probably going to grow to 20 to 30 to 50 to maybe hover around that 60 to 70% that the big four accounting firms hover at flat fee. Because the big four accounting firms say, “Wait, if I shrink my cost, I increase my profit margins.” So they do 60 to 70% flat fee, but they still do 30 to 40% billable hour because there’d be dragons here that we don’t know about. So I think that the death of the billable hour has been greatly overstated, but we’re just going to say the billable hour is going to shrink as a percentage of revenue and flat fees, contingency fees are going to increase. Another thing I wanted to do is I want to create logical inconsistencies in the other side.

So where you said, find logical flaws in OpenAI’s arguments. And where OpenAI said, “Hey, you have no legal interest in restricting the study of the code that you put on the internet.” And ChatGPT said, “Hey, that’s inconsistent because it implies by making the code freely available, I gave away my legal interest, but I put terms and conditions, the MIT license.” That is a good legal argument. And this one still blows my mind. It said you could argue that my Article three injury and fact was with the breach of contract itself because it was at that point that I lost control of my code. That is an amazing argument that I wish I could say after 15 years of litigation that I thought of, but I’m ashamed to say I don’t think I would’ve thought of it. In fact, I wondered, is this hallucinated, made up?

And so what I did is I went to my tool that has actual non-hallucinated cases, statutes, and regulations. I said, “Can the breach itself be an Article three injury in fact?” And it gave me actual non-hallucinated cases saying, yes, the breach of contract itself can be the injury in fact, but I didn’t stop there because I went to the text itself of Marsha McClellan and it says, yes, it can be a concrete injury. And then I followed up with other cases that then cited other cases saying, yes, breach of contract can be Article three standing. So the risk of hallucination with what you see on your screen here is zero because this is the actual non-hallucinated text from the actual non-hallucinated cases and there are only three companies in the world that have that data. You need the data and then the hallucinations are zero.

When I was at William Mitchell College of Law, one of my professors said, “If this thing happens, if you want to understand a statute, put it into if then statements. If this thing happens, here’s the output.” So I took a random statute, happens to be a New York statute for falsifying business records. Maybe a former president was convicted of this and I said, “Put this into if then statements. And here’s the output. If someone falsifies business records and intends to commit or aid and conceal another crime, then class E felony.” Now look at the text on the left and look at the text on the right. I’m going to ask you three questions, which is easier for you as lawyers to understand? Question number two, which is easier for your clients to understand. And question number three, which do you think is easier for computers to understand?

Because computers work in if then statements. So what if you have access to, say, a billion legal documents, all the statutes, all the regulations, not just in the United States, but in a hundred countries worldwide. And what if you were to run this through all of those statutes and regulations? We’re on our way to computable law. I said now argued that Donald Trump violated the statute and it says if he falsifies business records and intended to commit or aid and conceal another crime, then Class E felony. This is a reasoning machine. I’m a copy lawyer.That’s one of the things I did. I’m from Minnesota. So I said, “When is purple rain going to be in the public domain?” But I didn’t start with nothing. I started with the duration of copyright statute. So with that retrieval augmented generation, I said, “Well, he died in April of 2016.

Statute says life of the author was 70 years, 2016 plus seven equals 2086, but it also says that it goes to the end of the calendar year, so it doesn’t actually expire until January one of 2087. That is the correct legal answer. Now, if you’re a copyright client, option number one is to ask your friendly copyright lawyer. Option number two is to ask your friendly bot with yesterday’s case, yesterday’s statute and yesterday’s regulation. And that option one lawyer is not going to know why the phone stopped ringing. People say, well, AI is only as good as what it trained on. It’s not as good on the novel legal theories that I come up with. And to that, I say that’s a mischaracterization of the technology. A friend of mine works for an insurance company, and she said, Damien, we’re thinking about using effective computing in our call center.

What do you know about effective computing? And I said, I’m a tech lawyer, but I don’t know what it is, but let’s ask our friends ChatGPT. And it says effective computing is how computers understand emotions. And then I said, okay, how could it be used in a call center? And I said, you could use it to be able to see what are the emotional state of the callers. Are they highly emotional because of their accident? I said, cool. Give me a list of legal issues. And here’s the output. Have you thought about privacy? Have you thought about whether the callers consented to you analyzing their emotions? Have you thought about if you get hacked? What if all that data is in the hacker’s hands? These are good legal issues. And by the way, this is nowhere in the training data. Nowhere in the training data has anyone asked about effective computing and legal issues.

Even though it’s nowhere in the training data, what AI does is connects two things in latent space. Imagine the left-hand side is effective computing. The right-hand side is legal issues and what it does is AI connects those two to be able to do issue spotting for this fact. And that’s what lawyers do. We connect facts with the law. And if the law has yesterday’s case, yesterday’s statute and yesterday’s regulation, it does quite well, including beating 99% of humans on the bar exam. So everything that we do is based on words. As lawyers, we ingest words, we analyze words and we output words and it turns out that AI does all three of those things at superhuman speed and at a post-graduate level. So a friend of mine has a post-it note on her monitor saying, “Can AI help with this? ” And she says, “The answer is always yes because everything we do is based on words.

So is there anything AI can’t do? ” And I would say, “Yeah, maybe some things.” If your client asked you, “What are my odds of winning this motion for this case type in this court before this judge?” I would wage you that 99% of these people in this room would say the odds are pretty good or not bad. But what if you could say, “Well, the odds of winning summary judgment in this, all the courts are 54%, but winning in this court is about 45%, but winning for this judge, it goes up to 56%, but for trademark cases like mine, the odds are about one and three. Now, if you’re a client, would you rather hear pretty good or not bad? Or would you rather hear before I spend $60,000 on this motion, the odds of winning are about one and three. Corporations are doing this right now and they’re doing it without your help.

And by the way, ask a human what the odds are and the odds the human would say pretty good or not bad, ask an AI and it would hallucinate all over the place. But because we’ve done the hard work of extracting all the motion types and extracting all of the grants, denials, et cetera, we’ve done that with good old-fashioned AI, not large language model AI, but good old fashioned AI. We can now do this with a high percentage of accuracy. By the way, this isn’t new. We’ve had this for five years. The future is here. It’s not evenly distributed. Now what if you were to say, I want to be this judge, Susan Richard Elson, to grant my motion for summary judgment. What if you could say, show me winning motions like mine for this case type in this court before this judge? What if you were to take all of Citizens Richard Nelson’s motions for summary judgment saying, Your Honor, this is just like the case you decided last month, or just like the case you decided yesterday.

This is a universal brief bank where the smallest lawyers can beat with the biggest firms in the world. And now what if you were to take that and say, now draft me a new motion that is statistically likely to win. That is statistically likely to win for this claim in this court for this judge. What if you were to take all of Citizen Richard Nelson’s motions for summary judgment that she’s granted? Take the text of those, put those into a large language model and say, “Here are my new facts. Put this into here and give me a motion that is statistically likely to win.” This is why I left the practice of law back in 2015. I said someday we’re going to be able to do this and here we are 10 years later and we can do this. Tickle this judge’s brain and the way that this judge’s brain loves to be tickled.

And what if you bake all this into legal software? What if you don’t have to be a prompt engineer like I am? What if, again, this is the reason I left and practice of law in 2015. So what if you upload a complaint and it does everything I did when I was a litigator, Robins Kaplan? Here are all the claims. Here are all the facts supporting those claims. Here are the parties. What are they looking for? Money or an injunction? What are the defenses to those claims? What questions should I ask my client to help them win? No prompting necessary, you can get these answers and your clients can leave these answers. Fortune 20 companies are using this every day and the question is how often do they need to call their external counsel or not? Here’s copyright infringement law, all the facts, good defenses to the copyright infringement.

Here’s questions to ask the client. “Hey, open AI. Did you ingest the New York Times content and do the same thing for contracts? What are definitions, overlapping clauses? What are languages? The futures here, people are using this right now. It’s just not evenly distributed. So it’s used today by law firms. It’s used today by clients. And where do you stand in this? I talk to corporate law departments and they say this is like having a really smart colleague sitting right next to me. It doesn’t do things I couldn’t do, but it does things that it would have time to do. I use this many times a day, things that I would have called my lawyer about it and that lawyer doesn’t know why the phone stopped ringing. And how many hours would it take to be able to do a 200 page memo? And our AI can do it in about five minutes.

And because of that, the internal counsel that’s using this says, now I don’t have to defer to extended internal counsel anymore. Because in the past I had to say,” Oh, you spent the 200 hours. Now I have to defer to you. “But I get the 200 hours worth of work on day one. So when the lawyer says something that doesn’t make sense, I say,” No, look at page 120 that there’s a case that actually goes against that theory. I don’t have to defer to external counsel anymore. We wanted for patent drafting, we wanted to save $2 million and we did in year one. By the way, year one was 2023. Contract review, we saved 4,000 hours in year one. Contract summarization used to take six months, now it takes a day. Per matter, hoping to save 50,000. Turns out we saved 100,000. We want to bring most 51% of our work in – house.

By the way, this is a Fortune 20 company said we’ve already built 160 workflows that used to go to external counsel that we brought in – house. And we just don’t have to call our external counsel anymore. And they’re not going to tell their external counsel why. They’re just going to say, no, thanks. We don’t really need you anymore. So why do they hire you? And after AI, why will they hire you? You have judgment. I’ve seen this before. You work at your company, but I have worked with a bunch of companies like you. Here’s the judgment I know. Companies want to transfer their risk to you. I want your letterhead to be able to say it’s okay to do this. I want to be able to see what’s happening in the industry, what’s happening to regulators, what are you hearing on the market? And you know that judge, you know the counterparties.

These things will persist post AI. Internal accounts appear how many hours you spend on a thing. And if an internal council gets deliverables, 200 pages, no prompting necessary, how important are your deliverables? They’re building, not buying. Now everybody in this room, maybe some people are guilty of saying, AI have used it. It’s not quite there yet. I say, okay, which tool did you use? Did you use the pre-version of ChatGPT? Of course it’s not there yet. And how are you using it? That’s like saying this piano is awful. Because I banged on it for about five, 10 minutes and it didn’t work and it sounded awful. It’s not the piano folks. Maestros make masterpieces and lawyers are making masterpieces all over the place. Excel spreadsheets, it’s not the problem with Excel. It’s not Excel, it’s you. I haven’t used it. It’s not quite there yet.

It’s you. This is a funny and maybe the most profound cartoon you’ll see about AI is certainly the one I’ve seen. Person on the left said, “Look, I took a bullet point, turned it into an email I pretend I wrote.” The recipient said, “Look, I took an email, turned it into a bullet point. I pretend I read.” It’s funny but profound because if it started as a bullet point and if it ended as a bullet point, what’s the point of the email in the middle? There could be a thousand versions of that email. There could be a million versions of that email. The bullet points are the most important things. And by the way, the bullet points are ideas that are uncopyrightable and the email in the middle is also uncopyrightable according to the US copyrighted trademark offs. So we have uncopyriable turtles all the way down.

People worry about hallucinations. That’s one of the biggest reasons why they don’t use AI. Turns out humans do that hallucination too. They have a double-blind study where they said humans summarize a document, machine summarize it, mix them up, gave them to human reviewers and the human reviewers said that the humans hallucinated more than the machines. Humans got facts wrong more than the machines. They do that. By the way, that was September 23. Hallucinations gotten better since then. Same thing with doctors. And really that’s because we as humans get facts wrong. Watergate, before he knew there were tapes, John Dean got a lot of things wrong. He hallucinated, but he got some gist of things right. And the reason that he got things wrong is because we don’t actually access our brains through database. We actually reconstitute the memories every single time. So that’s why the fist starts is this big and then the next time is this big and this big, because we’re reconstituting those memories.

But it turns out that’s what AI does too. Every time it’s reconstituting the memories, sometimes getting it right, sometimes getting it wrong. So how do you fix that? Well, you do, just like you would ask Nomchansky, “Hey, explain to me the book that you wrote back in 1960.” And he’s going to hallucinate all over the place. He’s not going to remember. But then if you say, “Hey, Noam, here’s three pages from that book. Summarize those.” He’s going to get things right. That’s because he has the book right in front of him. So treat your AIs much like you treat a human. If you don’t want them to lie, don’t give them a closed book exam. Give them an open book exam. Give them the cases, the statutes, the regulations, your client’s facts. And if you upload that with context into the system, it is the rate of hallucination that’s going to be less than humans.

Giving them a few documents, that’s called retrieval augmented generation. And it turns out that the text you see on your screen here, that case will be non-hallucinated 100% of the time and that text will be non-hallucinated 100% of the time. And you might say, “Well, how can you say 100%, not 99%?” And I can say 100% because this text wasn’t created by AI. This text was a database search like we’ve been doing database searches for the last 40 years. 100% of the time, this text will be non-hallucinated and 100% of the time that case will be non-hallucinated. You need the ground truth to be able to do things. This is my friend and colleague, Ed Walters. He said, “There’s an easy way to avoid those hallucinations? Read the cases.” Because it turns out I clerked for judges and if I could tell you how many briefs I read where I read a proposition, then looked at the case and that case didn’t support that proposition at all.

Lawyers have been hallucinating the propositions for decades, but now it’s just painfully obvious that bad lawyers will bad lawyer because the cases don’t exist. If you want that hallucination problem to be solved, read the damn cases. And the A for retrieval augmented generation matters. Because if your client asks, “What’s the average case length for civil rights cases that settle in the Central District of California?” If your client asks you this, how would you answer? Could you answer? This is an impossible thing to answer, right? Because especially that settlement part, how do you see whether it’s settled or not? But if you have the data, if you have the Central District of California cases and if you have the fact that it’s a civil rights case and you have the file date and you have the terminated date, if you also have the fact that a stipulation to dismiss usually means the case has been settled, if you have all that data, you can then say that from 25 cases, some settlements in those 12 days, but most of them take about a year and a half.

That is answering an unanswerable question because you have the data, because we have 975 million dockets and documents across all the federal courts and 38 state courts where we can pull that data and give the answer to an unanswerable question. And that’s only because we’ve structured that unstructured data. So what if we had all of Texas’s cases, statutes of regulations and judicial opinions and are able to give that kind of number for any question that you give? And writing is thinking. A lot of people say, “What’s going to happen to our lawyers? Are they going to atrophy because they just can’t think anymore?” And writing is thinking, but the question is how much of our thinking is actually original? Because it turns out that everything we do is based on precedent, or at least it should be, cases, statutes, and regulations. And so when I litigated, if I told this judge I’ve got this great idea, that judge is going to say, cool, show me the case.

And if I can’t, good luck. And when I was a baby lawyer, one of my senior partners said, “If Damien, my most valuable billable hours is where I lean back and I think really hard.” And the question is in that hour, how many good ideas can that lawyer have? Maybe three, maybe four. And in that same hour, how many good ideas can AI have? And what if those ideas are based on non-hallucinated cases, non-hallucinated statutes, and non-hallucinated regulations? And what if you can then curate, because maybe those three or four ideas are actually part of that thousand. And now you can sort to say, here are the most likely to win in front of this judge or in front of this court. And how much of this is recreating the wheel or creating wheels that will never stand up in court because the cases don’t support it.

So between going with something that’s worked and cases in the past and something that hasn’t worked, why wouldn’t you go with a thing that has worked? And so of course, maybe we as humans are going to be taking those thousand and curating and saying, “I like this argument and I like that argument.” And that’s maybe going forward what we’re going to be doing. We’re going to be doing that with agentic law firms. In November of 22, that was a shot heard around the world, but that was just a chatbot. One prompt equals one output and that’s what everyone here mostly thinks is AI. But I would say that we’re far beyond auto. We’re far more transformative because in December of this past year, we had what’s called cloud code and open claw, which has agents. That is a whole agentic AI where swarms of agents are doing this work.

It’s not just a chatbot. And so this is humans on the loop where you’re overseeing the agents working rather than being in the loop. I’m a vibe coder, that is I’ve been coding since age 10 and I now have right now six bots working and just waiting for me to do the next thing. The bots are building code. So in our agentic law firm, you have an associate agent that then intakes all of the documents, that is all of the emails, all the Slack messages, all the documents. And then it says, “Oh, that looks like a breach of contract issue. That seems like a trade secret issue.” Oh, and it looks like you mentioned Texas and California and New York. And then you are able to say, “Okay, what are the elements of breach of contract or what are the elements of trade secret in Texas?

What are the elements of trade secret in New York? What are the elements of trade secret in California?” And maybe there’s fourth elements in California, three elements here. And then what it does is it finds the gaps and says, these facts go into each of these buckets. And then it says, “Oh, you’re missing a bucket number three. Client give me facts that can fill in bucket number three.” And then what you could do, by the way, this is zero billable hours at this point. And then the associate agent throws it up to the partner agent that then throws raps at the associate saying, “No dummy, do it this way. Oh, do it this way.” And they can do it as many times until you reach diminishing returns and then you throw it to opposing counsel agent, which then says, “No, no, no, I’m going to win because of this, ” which further bolsters the associate agents work throwing polished rocks because it’s doing this work on that side too, exposing every single one of your weaknesses and then you throw it to the judge bot and the judge bot says, “I’m going to find for plaintiff for this thing and then defend it for this thing and then throw it to the jury bot and be able to say which one has the facts that are most likely to resonate with the jury.” And you don’t just do that one time, maybe you do it 10 times or a thousand times.

And maybe you get the distribution curve of what is the most likely to win arguments that are going to sway this judge or sway this spot to improve my pleadings and only then do you hand it to a human Only then after everything I just told you does a human get it. This is not ChatGPT. People do not see this train coming down the tracks and it turns out that this is not science fiction. This is happening today. Helen Fan, I just saw her yesterday in San Francisco when I was giving a talk. She has built a bunch of bots, associate bots, partner bots, paralegal bots, opposing counsel bots, and they all argue with each other as to this thing and then they gets outputs. And this is my friend Kyle Barr, another good litigator lawyer that he did the Elon Musk OpenAI case where he went through the jurors, put their profiles into it and said, “After each day, which way is the jury swinging?

This is happening today. The future is here, it’s just not evenly distributed.” And the question is, “Well, Damien, how do I do this? ” And that’s kind of like saying, “Is there a book to read to learn how to swim?” The answer is no. You just swim. If you’re not dipping your toes into the water right now, you’re falling behind because I guarantee that there are plaintiff’s lawyers out there that are using this all over the place. And if you’re not using this, you’re bringing a knife to a gunfight. So law backed AIs. What if you were to align AI with human values? AI scraped the internet, some of the grossest parts of the internet, including assess pools like Twitter and Reddit and even with then it beat the bar exam with that. But what it got wrong is something like the rule against perpetuities, right?

I got that wrong. But what if you have a law back to AI? What if it has all the law, including the rule against perpetuities for all of the cases, statutes, and regulations and you have that worldwide? Wouldn’t that be a better system than general AI? Wouldn’t that be more accurate? And wouldn’t that also help Sam Altman when he stood in front of Congress saying, “We need to align AI with human values.” My response to Sam is, “Hey, we’ve kind of had that for about 800 years.” It’s called law because your human values are diferent than my human values, which is diferent than her human values, which is diferent than his human values. The way that we’ve aligned human values societally for 800 years is through the law. So maybe the AIs should align with the law that you and I follow. We’ve also had, even longer than that, we’ve had other values through faith traditions, Christian faith traditions, Jewish faith traditions, et cetera.

We’ve already had these alignment with human values, but the question is whose values are we going to follow? Whose laws are we going to follow? A friend of mine works for a large cloud company whose name you know that I’m not going to say. She’s a lawyer that says, “I need to take down objectionable material from the cloud.” And if it’s Nazi material under German law, it has to come down. Under German values, it has to come down. But under Texas law, it has to stay up as free speech. So she says, “Whose values do I follow? German values, German laws, or Texas values, Texas laws, because I can’t do both.” So we as lawyers navigate this every day. What is the least worst option? What’s the penalty under German law? What is the penalty under Texas law and what is the least worst option?

What if our bots did that too? What if we could follow and find the least worst option for each of these? And what if you hypothetically had access to 110 countries worldwide and you could access European Union law and Texas law and Mexican law and Brazilian law and Singapore law and what if you could then say, “This is the decision that aligns with most of the countries with the largest populations and maybe that aligns with my faith.” I know of a venture capitalist that says people that follow Christian values, companies that follow it make more moneyThat is they outperform the S&P 500. Love your customer, love your employees. Outperforms the S&P 500. So what if you were to say, “Here’s a 50 state survey on whether you should lay off 40% of your workforce. Here’s a multi-country survey and maybe look to the faith traditions, maybe what’s been scanned from the Vatican and be able to say, Will this maybe violate my faith tradition because I want to make more money.

I want to outperform the S&P 500. And what if as agents are doing more of our work, if you tell the agent, Hey, go start a business. And by the way, there is now a two person billion dollar company doing just this. Where they said, create a business plan, create an LLC, create a product, create a bunch of ads, create a bank account, bring in some money, achieve my goals. And if you have any questions, let me know. Think of all the laws that could be broken here. Well, what if you were to say, Hey, before each of these actions, consult the legal Oracle, consult Texas law and German law and my faith’s traditions. And if this violates any of those, let a human know. Human in the loop. How much better would our society be? Because this is preventive law. This is not having an ambulance at the bottom of a cliff.

This is putting a fence at the top of the cliff before the bots do things wrong and maybe before your clients do things wrong. Maybe we should take the ambulance from the bottom of the cliff and put fences like this at the top of the cliff. And it turns out that we already have AI that’s foundational model is based on actual non-hallucinated cases, statutes, and regulations. And we have other data sets with this worldwide data sets. In the five minutes we have less, let’s talk about business to block. And from a client’s perspective, I could call a partner, ask them a question, assign an associate, take a couple of days. They’re pretty smart. I could also do the same. Option two is to ask an AI tool. It’s not going to take two days and it’s not going to cost $5,000. And option one lawyer is not going to know why the phone stopped ringing.

And I gave this talk in Manhattan and this AmLaw 20 firm, they had a couple billion dollars in revenue last year. They said, “Damien, I don’t want $5,000 matters. I want $5 million matters.” I said, “I worked $5 million matters at Robin’s capital and you and I know that those are filled with $5,000 tasks. So how many clients are going to put up with you spending $5,000 on the thing said, No, I used Damien’s tool and I got it for free. There’s a study where they said for contract review, how accurate are the humans? Humans were 93, the machines were 95. Humans took 200 minutes, the machine took four. Cost per document, 76 for humans, buck 25 for robots. Already back in 23, this is a 300x cheaper for a machine that only was it cheaper, but was also better, better, faster, cheaper, all three of those.

So then the question is, okay, what happens next? Do I just fire all my associates? Do I just get rid of them? This is world number one that I hope we don’t do. Please do not fire your associates. Instead, keep your associates and let their agents underneath them do the work and serve the 92% of legal needs that are unmet because of access to justice. 92%, a lot of people call that a latent market that’s just waiting to be tapped. They need your health and individually they can’t pay much, but in the aggregate, 92% is a great number. So don’t fire your associates. Keep your associates to steal work from your competitors and go down market to serve that 92% because God knows we all need it. And the slowest lawyer wins the rates. That is, if you spend 10 hours on the thing that I spend one hour on, you make more money than I do.

But that’s also not true for everybody. It’s not true for contingency fee lawyers, not true for flat fee lawyers, not true for subscription fee or success fee. Also not true for corporate law departments. Also not true for government lawyers. For all of them, they love efficiency because they have more work than time. So maybe people should move from that hourly billing to one of the green because they can make more money. You want to incentivize hours. Say the associate, if you bill $3,000, you get more money. Cool. I’ll give you more hours. If you want to incentivize efficiency and quality, I’ll give you more efficiency in quality. We as lawyers often think about how much money we brought in last quarter, last half, et cetera. When we think less about how much that cost, what is our overhead? Because revenue minus cost equals profit margin.

But if you go over to the flat fee side or contingency fee side, you realize if I shrink my cost, maybe using AI, I make more money. Go to this way because not only can you then say, “Hey, I’ll charge you less than my competitors.” The client says, “Wait, wait, not only do I pay less, but I also get the certainty of a flat fee, sign me up.” But the thing is the client doesn’t need to know that they need to push that number up and even make more money and make your higher profit margins. And then it just becomes a debate as to you and your clients as to what that top line number is going to be. This is how lawyers make more money with AI than before AI, shrinking your cost to increase profit margins. When I was a litigator, I would do plaintiff side antitrust phases and bid rigging was something that often happened, but almost inevitably somebody would drop their prices.

And what would happen then is they would steal the customers from all their competitors. And what would happen next is that all their competitors’ prices would come down. So this is how price fixing cabals went away. But it turns out that lawyers don’t have price fixing balls. We maybe have a hourly billing cabal. But lawyers all over the place are realizing, wait, if I shrink my costs, I make more money if I jump over to the flat fee side. And these flat fee lawyers are stealing the customer client base from all their competitors. And what happens next is obvious. The question is, do you want to be lawyer number one or do you want to be lawyer number 20 that doesn’t know why your phone stopped ringing? So between three potential worlds, there’s a world where the old world where you pay people to do work.

So one of my friends works two full-time jobs and gives 100% productivity even though he works two full-time jobs. Another world is that you could work 40 hours a week and then your employer makes way more money using AI. Another third world is you lay off a bunch of people and then they, the people remaining spend 40 hours a week and deal with 5X returns. Societally, we are going to have to decide where this is because this is a don’t ask, don’t tell world. This is a world of abundance where we make more money than ever. There’s a world of scarcity where we fire a bunch of associates and we fire a bunch of employees. I am on team abundance. I think we can make more money than we ever have societally and I think we law firms can make more money if we realize that we just can’t practice in the same way as we have in the past.

Let’s move forward to our world abundance. Let’s not do scarcity. Let’s not fire our associates because there’s something called Jevon’s paradox. Jevin’s paradox, we thought LED light bulbs were going to save us a lot of energy, but it turns out they’re so cheap we just leave them on all the time. Electricity with LED light bulbs are too cheap to measure. So what if legal work were too cheap to measure? What if every time, gosh, every time I call my lawyer, it’s a thousand bucks. Oh, forget it. I’m going to risk it. What if that thousand bucks shrinks to a hundred bucks? Man, I’m going to call you every day. And in the aggregate, maybe I spend more than $1,000, maybe $1,500, $2,000. This is called Jevin’s paradox, that if we shrink our costs using AI, we’ll increase societal use of lawyers and increase access to justice.

So in closing, I would say let’s not go to the world of scarcity folks. We can do so much good in this world. We can do good for access to justice. We can do good for our clients if we just think differently about our business models. That 92% of legal needs, again, if we’re not ashamed of that 92%, we should be. Let’s do something about being able to help that 92%. Thank you very much.