LexisNexis CEO: the AI law era is already here

Right this moment, I’m speaking with Sean Fitzpatrick, the CEO of LexisNexis, one of the vital vital firms in the whole authorized system. For years — together with once I was in regulation faculty — LexisNexis was principally the library. It’s the place you went to lookup case regulation, do authorized analysis, and discover the legal guidelines and precedents you’ll must be an efficient lawyer in your shoppers. There isn’t a lawyer at present who hasn’t used it — it’s basic infrastructure for the authorized career, similar to e mail or a phrase processor.

However enterprise firms with enormous databases of proprietary info in 2025 can’t resist the siren name of AI, and LexisNexis is not any completely different. You’ll hear it: once I requested Sean to explain LexisNexis to me, the primary phrase he stated wasn’t “regulation” or “information,” it was “AI.” The purpose is for the LexisNexis AI software, referred to as Protégé, to transcend easy analysis, and assist attorneys draft the precise authorized writing they undergo the courtroom in help of their arguments.

That’s a giant deal, as a result of up to now AI has created simply as a lot chaos and slop within the courts as wherever else. There’s a constant drumbeat of tales about attorneys getting caught and sanctioned for counting on AI instruments that cite hallucinated case regulation that doesn’t exist, and there have even been two courtroom rulings retracted as a result of the judges appeared to make use of AI instruments that hallucinated the names of the plaintiffs and cited details and and quoted circumstances that didn’t exist. Sean thinks it’s solely a matter of time earlier than an lawyer someplace loses their license due to sloppy use of AI.

So the large promise LexisNexis is making about Protégé is just accuracy — that every part it produces shall be primarily based on the true regulation, and way more reliable than a basic goal AI software. You’ll hear Sean clarify how LexisNexis constructed their AI instruments and groups in order that they will make that promise — LexisNexis has employed many extra attorneys to overview AI work than he anticipated, for instance.

However I additionally wished to know what Sean thinks instruments like Protégé will do to the career of regulation itself, to the job of being a lawyer. If AI is doing all of the authorized analysis and writing you’d usually have junior associates doing, how will these junior associates study the craft? How will we develop new senior folks and not using a pipeline of junior folks within the weeds of the work? And if I’m submitting AI authorized writing to a decide utilizing AI to learn it, aren’t we getting near automating a little bit an excessive amount of of the judicial system? These are large questions, and so they’re coming actual quick for the authorized business.

I additionally pressed Sean fairly arduous on how judges, notably conservative judges, are utilizing computer systems and know-how in service of a judicial idea referred to as originalism, which states that legal guidelines can solely imply what they meant on the time they had been enacted. We’ve run tales at The Verge about judges letting automated linguistics techniques attempt to perceive the originalist intent of assorted statutes to succeed in their most well-liked outcomes, and AI is just accelerating that development — particularly in an period the place actually each a part of the Structure seems to be up for grabs earlier than an extremely partisan Supreme Court docket.

So I requested Sean to demo Protégé doing a little authorized analysis for me, on questions that look like settled however are newly up for grabs within the Trump administration, like birthright citizenship. To his credit score, he was sport — however you may see how taking the corporate from one that gives easy analysis instruments to at least one that gives precise authorized reasoning with AI may have large implications throughout the board.

This one is weedsy, but it surely’s vital.

Okay: LexisNexis CEO Sean Fitzpatrick. Right here we go.

This interview has been frivolously edited for size and readability.

Sean Fitzpatrick, you’re the CEO of LexisNexis. Welcome to Decoder.

Thanks. Nice to be right here.

Thanks for becoming a member of me. That is my first interview again from parental go away. Apologies to the viewers if I’m rusty, however apologies to you if I’m simply completely crazy.

I’m very excited to speak to you. I’m very a lot a failed lawyer, my spouse is a lawyer, there’s plenty of attorneys on The Verge group. The authorized career in America is at a second of absolute change, plenty of chaos, and an unlimited quantity of uncertainty. And LexisNexis, if the viewers is aware of, tends to take a seat on the coronary heart of what attorneys do all day. Most attorneys are utilizing LexisNexis each minute of on daily basis. What that product is, what it may well do, and the way it helps attorneys do their job connects to plenty of themes that we’re seeing each within the authorized career and in know-how and AI usually.

So, begin at first. What’s LexisNexis? How would you clarify it to the layperson?

LexisNexis is an AI-powered supplier of knowledge, analytics, and drafting options for attorneys that work in regulation corporations, companies, and authorities entities.

That’s a brand new conception of LexisNexis. Once I was in regulation faculty within the early 2000s, it was simply the factor I searched to search out case regulation.

Sure, we’ve reworked over time. We had been simply that analysis supplier. Over time, we’ve acquired and built-in extra companies. In 2020, once we launched our Lexis+ product, we built-in all these issues collectively, turning into an built-in ecosystem of options. Then, in 2023, we launched Lexis+ AI, and that’s once we actually grew to become an AI-powered supplier of knowledge analytics, determination instruments, and drafting options. AI capabilities have actually allowed us to do extra issues than what we’ve historically completed previously

That soar from being the gold customary database of authorized opinions, reasonings, case notes, and all that to “we’re going to do the be just right for you or allow you to do the work” is a giant one. That’s a cultural soar. Clearly, there have been some acquisitions alongside the way in which that helped you to make that soar, which you’ll be able to discuss. What drove you to make that soar, to say, “Really, the attorneys need assistance drafting the motions, the proposed opinions they may give to a decide?” What made you say, “Okay, we’ve acquired to step into truly doing the work?”

I believe it’s been a pure evolution. As know-how has advanced, it’s opened up new avenues of issues we will do. We are inclined to take the most recent know-how, introduce it to our clients, and spend time speaking to them about how they suppose that know-how may be finest utilized within the authorized atmosphere. Then, we translate the concepts they got here up with into merchandise that resolve or tackle these alternatives.

Let me ask you a fairly philosophical query. It’s one which I wrestle with on a regular basis. It’s one which I speak to our viewers about on a regular basis. I believe most individuals who encounter the authorized system suppose it’s fairly deterministic. Our viewers is fairly technically targeted. They’re used to computer systems. Computer systems are, till just lately, fairly deterministic. You place in some inputs, you get some outputs. Most individuals who encounter the authorized system suppose it’s fairly deterministic. You place in some inputs and also you get some predictable outputs.

And what I’m at all times saying is, “That’s not the way it works in any respect.” You present as much as courtroom, the decide is in a foul temper, you don’t have any concept what’s going to occur. You’re a giant firm with an antitrust enchantment, you present as much as the three-judge appellate overview board, and you don’t have any concept what’s going to occur. Actually something may occur at any time. The judicial system is essentially not deterministic. Though it’s structured like a pc, attempting to consider it like one can get you in all types of hassle. Possibly the very best instance of that is folks on Fb placing the phrases “no copyright meant” on the underside of films. They suppose they will problem these magic phrases and the authorized system is solved, and so they simply can’t.

AI is that drawback in a nutshell. We’re going to take a pc, make it higher at pure language. We’re going to make the pc essentially not deterministic — you may’t actually predict what an AI goes to do — after which we’re going to use that to the essentially non-deterministic, human nature of the courtroom system. Someplace in there’s a large philosophical drawback about making use of computer systems to the justice system. How do you consider that?

To begin with, you’ve got these large investments occurring with the foundational fashions. Every of those hyperscalers — Microsoft, Amazon, Google — are placing in near $100 billion. So these fashions simply proceed to get higher and higher over time. That’s on the foundational mannequin degree. We don’t actually function at that degree. We construct purposes that make the most of these foundational fashions. And at that degree, we see costs are dropping. We used to pay $20 for 1 million tokens two years in the past, and at present we would pay 10 cents for 1 million tokens. That permits us to do issues at velocity and at scale that we’ve by no means been in a position to do earlier than.

And there are plenty of issues concerning the regulation that make these fashions engaging. A lot of the regulation is language-based, and these fashions are actually nice with language issues. The regulation is precedent-based, and so —

Properly, that’s up for grabs. We’ll come again to it.

I’ll grant you that. You take a look at the actions that attorneys do: they draft paperwork, they do analysis, they summarize issues. The fashions are all actually good at some of these issues. So, you’ve got this good storm, with this know-how and the issues attorneys do coming collectively.

But, when folks attempt to use these consumer-grade fashions, there are all types of issues with them. Such as you stated, it’s not deterministic. You’ll be able to’t simply put info into a pc and get a solution out. If that had been the case, we wouldn’t want a courtroom system. These fashions are simply not constructed for the authorized system. You’ll be able to’t go into courtroom and say, “I discovered this on the web.” It’s a must to have authoritative content material.

The deadline for GPT-4o was 2023, I consider. You could have info that’s continuously up to date. Your viewers in all probability doesn’t know this,, however there’s the citator, which historically has stated, “That is good regulation” or “It’s not good regulation, it’s been overturned.” Now, it’ll let you know if it’s the regulation in any respect or if some system simply made it up. These techniques are probabilistic. They need to put collectively a solution that’s in all probability proper. Properly, that’s not the usual we have now in authorized. You’ll be able to’t go in with one thing that’s in all probability proper. So, you’ve got this complete record of points that these fashions don’t tackle.

What we’ve tried to do is tackle these with a courtroom-grade answer. Our system is backed by 160 billion paperwork and information. Our curated assortment is our grounding information. So can’t go into courtroom and say, “I discovered this on the web,” however you may consult with a selected case. We even have what we name a citator agent that’ll examine that case to guarantee that it wasn’t fabricated by the system and is definitely nonetheless good regulation. You may also take a look at the case regulation abstract so you recognize what the case is about. You’ll be able to take a look at the headnotes so you may see the actual factors of regulation that had been addressed in that case and see if it’s nonetheless a sound case.

Privateness is one other problem. There’s a particular relationship that exists between attorneys and their shoppers in that attorney-client privilege, so there’s some privateness necessities that you simply want as a way to keep that. When you’re utilizing one in every of these simply consumer-grade fashions, you don’t have the extent of privateness and safety that you simply want. Transparency is one other problem. You place a query in, you get a solution again. Properly, primarily based on what? What was the logic that the system used? We open up the black field so you may see the logic that’s being utilized. We give the attorneys the power to go in and really change that. If this mannequin is getting one thing unsuitable, the lawyer has the chance to vary it in order that they get the result that they’re driving for. However, as you stated, the regulation just isn’t deterministic. There are many various factors that go into this, however it’s good to have a system that’s legally pushed, that’s goal constructed for authorized conditions as a way to actually function in a courtroom-type atmosphere.

There’s two issues I actually need to push on. Once more, I used to be not an excellent lawyer. I don’t need to ever faux on the present, to you, or to anybody else that I used to be any good at this. However you study a selected mind-set in regulation faculty, which is a fairly rigorous, structured means of approaching an issue, going to search out the related circumstances and precedents, after which attempting to style some answer primarily based on that. That seems like we’re simply shifting phrases round, but it surely’s truly a mind-set. Earlier than AI confirmed up, we’d mash collectively utilizing a phrase processor and considering a sure means. Now, we’re pulling them aside. We’re saying that the pc can transfer the phrases round and generate some considering. So, that’s one factor I need to push on. I’m very interested by that as a result of it feels just like the lawyering a part of being a lawyer is being subsumed right into a system, and that may change how we lawyer.

The opposite half is that if anybody goes to have a look at the work being completed. We’re already seeing attorneys get sanctioned for submitting briefs with hallucinated case citations in them. There was only a case the place, I consider, a courtroom needed to rescind an opinion as a result of it had a hallucinated case quotation in it. That is unhealthy. That is simply straightforwardly a risk to the system and the way we would take into consideration attorneys, judges, and courts. It’s not clear to me that anybody’s going to make use of the instruments as rigorously as you need.

So on the one hand there’s “We’ve made the considering simpler.” Then again, it’s, “Oh, boy, everybody’s going to get actually lazy.” They’re each in your reply. They’re each saying, “We’re making it simpler to have a look at these items. We’re making it quicker to do the analysis.” I’m simply questioning the place you suppose the considering is available in.

I don’t suppose that these fashions substitute the attorneys. I believe they assist the lawyer and increase what the lawyer does. So, if you consider an exercise {that a} lawyer may do — let’s say they had been making ready for a deposition. They should give you a listing of questions that they’re going to ask the person who’s being deposed. You’ll be able to take the details round that exact case, load them right into a vault, and level the system to that vault and say, “Based mostly on the details of this specific matter, develop a listing of deposition questions.” That’s one thing {that a} lawyer would’ve completed on their very own. Up to now, they could have referred to a listing of questions that that they had beforehand or one thing —

Really, can I simply seize that instance? Possibly a lawyer would’ve completed that, however extra typically a lawyer would’ve informed a bunch of junior associates to take a seat within the basement and try this. That was how these junior associates discovered do their job. That’s what I imply. We’re farming out the considering, and a few folks may by no means truly try this considering. That may change the career down the road in actually substantive methods.

Proper. It’s an apprentice system. So, for those who begin to take among the layers out of the underside, how does everybody skip the underside layer and nonetheless make it to the second with the identical capabilities and abilities? That’s an actual problem. I believe the techniques are permitting attorneys to not have the affiliate try this work. Now they will say, “Generate me 300 or 700 questions.” It doesn’t take that lengthy to undergo 700 questions, and the fashions by no means get drained. From our expertise, they’ll undergo that record of questions and say, “First query? Yep, that’s an excellent query. I might’ve considered that.” The system made it a little bit bit quicker, but it surely didn’t actually assist them. Second query, identical factor. Third query, identical factor. Fourth query doesn’t even make any sense, scratch it off the record. With the fifth query they’ll say, “Oh, that’s fascinating, I wouldn’t have thought to ask that however that’s in all probability vital, so I’m going so as to add that to my record.” So, there’s an effectivity part to it, however I believe there’s additionally a greater final result part.

By way of the apprenticeship piece, I believe persons are struggling proper now to determine how that’s going to impression the apprenticeship mannequin. Somebody was describing to me that that they had labored on a scenario the place they had been taking a look at securitized property. After they had been an affiliate, they did this challenge for an organization that had 50 states price of protection, and they also grew to become the knowledgeable within the agency on asset securitization in all 50 states. For 4 or 5 years, anytime someone had a query, they got here to that particular person. It was a good way to make a profession. Now, the system can do all that info for you. So his query was: “How is that ever going to occur now on this new world?”

I believe corporations are going to wrestle with that, however I additionally suppose they’re going to determine it out. We are inclined to get among the smartest and brightest folks going into the authorized career, and up to now, they appear to have found out each problem that’s confronted the business. I believe they’ll determine this one out as effectively.

What are some options you’ve seen as folks attempt to determine this out?

I don’t know that folk have give you plenty of options across the apprenticeship mannequin. What we’re for positive seeing is that persons are embracing AI. It’s right here, it’s within the courtroom, it’s within the regulation agency. Two-thirds of attorneys are utilizing AI of their work, in keeping with our surveys, and our survey’s in all probability a little bit outdated. I’d say the quantity’s in all probability greater. I don’t find out about you, however I exploit AI on daily basis. It’s now in my private and work lives. I believe the authorized career is completely fitted to it, so it’s solely going to broaden.

If you see the attorneys getting sanctioned and the courts having to rescind opinions, is there an answer that includes utilizing LexisNexis so it gained’t occur to you? Or do you suppose that’s a symptom of one thing else? Everybody’s simply utilizing AI, I get it. In all probability the largest cut up for our viewers proper now’s between the information that claims everybody’s utilizing these items on a regular basis and the hostility our viewers expresses concerning the instruments, their high quality, and the truth that plenty of that utilization is pushed by large firms simply placing it in entrance of them. There’s one thing occurring the place, to justify these huge investments, the instruments are displaying up whether or not the customers are asking for them or not, after which we’re mentioning that everybody’s utilizing the instruments.

What I hear from our viewers is, “Properly, I can’t flip off the AI overview. After all I’m utilizing the software as a result of it’s simply in entrance of me on a regular basis. I can’t make Microsoft Workplace cease telling me that it’s going to assist me. It’s simply in entrance of me on a regular basis.” So, if you see the errors being made within the authorized system at present — the attorneys getting sanctioned, the lazy AI use, the shortage of apprenticeship that’s going to impression the whole subsequent technology of attorneys and the way rigorous they’re — how do you make your product tackle that? Or are you simply not eager about that proper now?

No, we’re positively eager about it, and we’ve included issues into our product. This stuff at all times make the headlines once they occur, however I believe it’s a small share of attorneys which might be inflicting these issues. Simply taking one thing and bringing it into courtroom has by no means been the usual. You’ve at all times had the duty as a lawyer to examine the fabric and guarantee that it’s legitimate earlier than going into courtroom. And a few people aren’t doing that. We definitely noticed that within the [Tim] Burke case the place some attorneys submitted a doc to the courtroom that I believe had eight citations in it and 7 of them had been simply utterly —

However that was inevitable. The day ChatGPT confirmed up, half of the authorized pundits I do know had been like, “That is inevitable. This final result will occur,” after which it occurred. There wasn’t even a stutter step. It simply occurred instantly. That’s what I’m attempting to push on. Is the answer simply that LexisNexis has a software that’s higher and you need to pay for it, or is the career going to should construct some new guardrails as we take the rigor away from the youthful associates?

Properly, you may by no means cease an lawyer from taking it into courtroom and never doing the right work. That’s going to proceed to occur. I believe someone’s going to lose their license over this in some unspecified time in the future. We’re seeing the sanctions begin to ratchet up. So, a pair attorneys acquired fined $5,000 a chunk, after which some attorneys in federal courtroom down in Alabama acquired referred to the state bar affiliation for disciplinary motion. I believe the stakes are growing and growing. What we do with our system is present a hyperlink to a quotation if we have now it, so you may click on on it and see it in our system.

And there’s no fabricated circumstances inside our system. We’ve a group mechanism that ensures that each case in there may be legitimate. It’s shepardized and has headnotes and completely different instruments that attorneys can use. So, we make it very easy so that you can use our system to examine and guarantee that the citations you’re bringing into courtroom are usually not solely legitimate and nonetheless good regulation however are additionally in the appropriate format. Format is vital. We examine for all these items and make it very easy for the lawyer to do the work they should do. They should guarantee that case is on level, that the case remains to be legitimate.

One of many many causes I used to be a horrible lawyer was due to that second if you get your first regulation agency job and also you notice your boss simply has a library of their favourite motions on file. They’re simply going to tug from the cardboard catalog, change some names and dates, and file the movement. The decide will acknowledge the movement and the lawyer, and that is all only a bizarre formality to get by means of the subsequent stage of the method. Possibly we’ll by no means get to the substantive a part of the case as a result of we’re simply going to settle it, however we have to file this movement we had and elaborate. This really was demoralizing. I used to be like, “I’m simply doing paperwork. There’s nothing about this that’s actual.”

I’m in all probability describing what each first 12 months affiliate goes by means of till the examine hits, and it simply didn’t work for me. How shut are you to having Lexus AI simply try this factor, have it acknowledge the second and say, “We’ve the banked movement and we’re simply going to file it to a system?”

Properly, we will join right into a doc administration system (DMS) that has an lawyer’s prior motions. We’ve our vault capabilities, to allow them to load their motions up. They will nonetheless use the motions they’ve already developed. And that’s a wonderfully nice solution to do issues as a result of —

Properly, I’m saying, from scratch.

Proper. We’ve the power to do it from scratch too, however plenty of attorneys don’t need to do it from scratch as a result of they’ve reviewed each single phrase in that movement and so they know that it’s good. In the event that they do it from scratch, then they should overview each single phrase. But when they need to do it from scratch, we will try this for them at present, and if they need, we will use their prior work product because the grounding content material to create a brand new movement, or we will use our authoritative materials. They will select the supply and the grounding content material.

I suppose I’m asking what degree of automation is there. So, you’re an lawyer. You’ve acquired a doc administration system, you’ve acquired a brand new shopper, and it’s good to file some standardized movement that you simply at all times file for no matter factor it’s good to do, like a continuance. At what level does Lexis [AI] say, “I’m watching this case. I’m going to file this for you. I’m simply going to hit the buttons for you. Don’t fear about it,” in the way in which that a terrific authorized assistant may do?

We’re at all times going to provide the lawyer the chance. We don’t need to simply be doing issues on their behalf unsupervised, so we’re going to provide them the chance. We may get to the purpose the place we are saying, “It appears such as you want a continuance. Right here’s a draft of a continuance push,” and it’ll robotically file it. We’re not at that time at present, however for those who want a continuance, we may draft it for you. Our imaginative and prescient is that each lawyer goes to have their very own personalised AI assistant, and it’s going to know their apply space and their jurisdiction, together with getting access to their prior work product. The techniques are solely nearly as good because the content material behind them, so it’s going to have entry to our 160 billion paperwork and information, and it’s going to have the ability to automate duties that they do at present. If you consider all of the various kinds of attorneys and all of the completely different duties that they carry out, there’s in all probability 10,000 duties that could possibly be automated.

So, we’re working with our clients to know what an important duties are, and we’re working with them to automate these duties at present. We’ve the biggest and most sturdy backlog of tasks that we’ve ever had in our firm’s historical past as a result of there are such a lot of of issues that may nonetheless be automated, and we’re working with our clients to do this. In the event that they inform us, “What we actually need is so that you can robotically file this” or for us to offer them with an alert that claims, “Hey, this deadline is arising and it’s good to file this. Right here’s a draft. Do you need to file it?” I’m positive we will develop it.

We’re not at that time at present, however we’re within the drafting part. That imaginative and prescient just isn’t a five-year imaginative and prescient or a three-year imaginative and prescient, that’s accessible at present. That’s Protégé. That’s what Protégé does at present. There are duties that it may well do, however we haven’t completed that large backlog but.

When you take a look at the sweep of different CEOs who’ve been on Decoder, they’re going to let you know, “You simply combine our laptop imaginative and prescient system and we’ll use [electronic case files] so that you can file this movement.” They’ll all be very blissful to promote you that product, I’m positive.

The rationale I’m asking it this fashion is as a result of once I get the patron AI CEOs on the present, they love to inform me that they’re going to jot down my emails for me with AI, after which the subsequent sentence they are saying is, “Then, we’ll kind your inbox with AI.” In some unspecified time in the future, the robots are simply writing emails to one another and I’m studying summaries. One thing essential has been misplaced in that chain. One of many funniest outcomes of AI is my iPhone all of the sudden simply summarizing emails and producing emails for different iPhones to summarize, and I don’t know what’s happening.

That’s unhealthy within the authorized context. We’re automating doc technology to make the case for our shoppers. On the opposite facet, the judges and clerks could be utilizing these identical instruments to ingest the circumstances, summarize them, perceive the arguments, and write the opinions which might be the outcomes. Culturally, I believe it’s vital so that you can have a perspective on the place that ought to cease as a result of in any other case we’re simply going to have a completely automated justice system of LLMs speaking to one another. Possibly there’ll be some guardrails that different folks don’t have, however we’ve taken an unlimited quantity of people out of the loop.

I believe it’s important to have the human within the loop. It’s an vital a part of the method. I may see the bots going forwards and backwards on issues like if somebody says, “Hey, are you able to meet at 9 o’clock?” and your system opens up the calendar, says you’re accessible to satisfy this individual in your excessive precedence record, and units up the assembly. If you’re speaking about substantive authorized issues, the stakes are too excessive. You’re speaking a couple of disabled veteran getting or not getting their advantages. You’re speaking a couple of sufferer of a pure catastrophe getting or not getting insurance coverage proceeds. You’re speaking a couple of single mom getting or not getting welfare advantages. These are all authorized issues, and so they actually have a big impact on folks’s lives. The stakes are means too excessive for bots to be going forwards and backwards and sharing info.

Do you suppose that clerks and judges needs to be utilizing AI the identical means attorneys needs to be? That’s the place I might draw the road. I believe the clerk needs to be made to learn and interpret every part as people, and the judges needs to be made to jot down every part as people, but it surely doesn’t appear to be that line has been formalized wherever.

I don’t suppose a decide ought to write each line. I believe that they might use AI. It’s nice if you put ideas in, put the phrases round that idea and construction them in an orderly means. I believe that there’s a part of the work that could possibly be completed by AI, but it surely shouldn’t be a bot speaking to a bot. I don’t suppose it needs to be totally outsourced to AI. You’ve acquired a duty as a decide, as a regulation clerk, as a lawyer to overview that doc and ensure it’s truly saying what you propose it to say.I believe most attorneys are utilizing it that means. It would create a terrific draft, perhaps at 80 p.c, which lets you do 20 p.c of the work. However that 20 p.c is the deep, analytical thought work, the belongings you truly went to regulation faculty to do versus what you had been describing earlier. It’s going to permit attorneys to do extra of that sort of labor.

I’m curious to see how completely different jurisdictions and circuits strategy the query of what the judges and clerks needs to be doing. I sense that that stress goes to specific itself in numerous methods throughout the sector.

Judges have gotten forensic auditors. They’re reviewing this info on the lookout for pretend circumstances. We don’t need them doing that. That shouldn’t be their job. I believe issues do want to vary in a few of these areas.

Utilizing AI to catch AI is one other theme that comes up on Decoder on a regular basis.

I’ve totally forgotten to ask you the Decoder questions. So let me try this, after which I need to zoom out a little bit bit farther. These are my very own questions. You’ll be able to inform, I’m a little bit rusty.

I’m trying on the LexisNexis management construction, it’s very difficult. There’s a CEO who’s not you, Mike Walsh, however then you definately’re the CEO of the US and the UK. There’s a bunch of different VPs in all places. You’ve acquired a mum or dad firm referred to as RELX. Clarify how LexisNexis is structured and the way your half suits into it.

RELX is the mum or dad firm, and it’s publicly traded. It has 4 divisions. Authorized and Skilled is a type of divisions, and its CEO is Mike Walsh. I report back to Mike. I’m the CEO of our North America, UK, and Eire companies. The way in which that we’re organized, it’s a matrix. We go to market primarily based on buyer segments. So, we have now a big regulation enterprise, a small regulation enterprise, a company authorized enterprise, a federal authorities enterprise, a state and native authorities enterprise, a Canadian enterprise, and a UK enterprise.

Then, we have now practical teams that help that. So, we have now product administration, and so they’re answerable for our product improvement roadmap and the product technique. We’ve an engineering group, and so they take the route from product administration however truly construct the merchandise. We’ve practical teams that help that, finance, HR, authorized, and world operations that does issues like accumulate content material for us. When you get used to it, it’s not that difficult of a construction. It’s very well built-in and seamlessly built-in collectively, which permits us to function actually rapidly. We will get issues completed rapidly and effectively. And I might say that the entire course of is buyer pushed.

I’m actually within the construction, notably the truth that you’ve got the UK, Eire, and North America. I’m fascinated by company buildings, and one of many issues that strikes me is that you’re not answerable for the taxonomy of your product, proper? These international locations’ governments are answerable for the taxonomy of their authorized techniques. The English authorized system and the American authorized system have commonalities however wildly completely different buildings. The Canadian authorized system and the US authorized system have wildly completely different buildings. Canada truly has extra in frequent with the UK given their shared historical past. How do you consider that? Are these completely different groups? Have they got completely different database buildings? How does all that work?

We do have completely different groups and completely different database buildings, however we’re truly attempting to consolidate to the extent that we will as a result of when we have now related issues, we shouldn’t have them marked up otherwise in numerous databases. Getting them marked up in a constant means will permit us to do what we name “excessive reuse,” which is to principally use that very same know-how we develop in a number of jurisdictions with restricted adjustments to that system. What that enables us to do is admittedly give attention to that core system and roll it out rapidly, so that everybody the world over will get the advantages of all these adjustments. However you’ve got civil regulation in some jurisdictions and customary regulation in others, and the legal guidelines are structured in numerous methods. So, you do have issues that make that more difficult, however that’s the overall concept behind what we’re attempting to do.

Are you able to apply the identical AI techniques to those completely different authorized techniques in the identical means, or are you actively localizing them otherwise?

I might say that we actively localize them, however we attempt to reduce the quantity of labor that we do as a result of plenty of it may be completed in an analogous means.

Usually, there’s plenty of concern about American authorized precedents touring throughout the ocean, notably within the UK. You’ll be able to see the American tradition conflict will get exported and reveals up in plenty of other ways. Do you suppose your software will make that higher or worse? When you’re not pulling them aside and are literally attempting to attenuate the variations, you may see repeat arguments or repeat buildings simply primarily based on the way in which the AI works.

Each is predicated on the content material of the person jurisdiction. So, we don’t combine the content material, however we do attempt to make the most of the identical know-how. For instance, there’s search relevance know-how to search out the case that’s most intently related to the matter that somebody is engaged on. We will take that and construct it for the US market or the UK market, after which we will transfer it to a different market and it’ll work fairly effectively. Then, we have to do some modifications to make it work very well for that exact jurisdiction. We get 80 p.c of the DNA transferred over in that core mannequin.

I used to be just lately speaking to Mike Krieger, who’s the chief product officer of Anthropic — only a completely completely different dialog on a special factor — however he stated this factor to me, which is caught in my thoughts. He stated, “I acknowledge Claude, I can see Claude’s writing.” He stated, “That’s my boy,” which is cute. Does your AI have a character? Can I acknowledge its writing in all these completely different jurisdictions?

We use a multi-model strategy, and so it’s in all probability rather less clear which specific mannequin drove one thing. After all with agentic AI, issues have actually modified. I believe that was in all probability true a 12 months and a half in the past, however now with agentic AI, when somebody places in a question… let’s say they wished to draft a doc. Possibly a shopper is sending in a request and he or she’s fascinated about a premises legal responsibility problem across the responsibility to tell a trespasser a couple of harmful situation on a chunk of land. The question will go right into a planning agent, who will then allocate that question out to different brokers.

It must do some deep analysis, so perhaps it makes use of OpenAI o3 as a result of it’s actually good at deep analysis. On the finish, it must draft a doc, so perhaps it makes use of Claude 3 Opus, which is admittedly good at drafting. We’re mannequin agnostic, and we’ll use no matter mannequin is finest in a selected activity. So, the outcome you get again was truly doubtlessly completed by a number of completely different fashions, which in all probability makes it a little bit bit tougher simply to see if it was drafted by OpenAI.

Is that mirrored in your construction? You describe engineering, product, and your localization, however you’ve acquired to construct that agentic orchestration layer and determine which fashions are finest for every goal. You possibly can design an engineering group round that drawback particularly. Is that the way you’ve completed it or is that completed otherwise?

We’ve an engineering group that focuses on the planner agent and the task of the duties to completely different brokers.

Is that the place the majority of your funding is or is it paying the token charges?

I haven’t truly damaged it out that means, so I couldn’t let you know. The token charges are definitely an vital a part of the funding. Engineering is a large portion of the funding. The attorneys that we rent to overview the output and inform us if it’s good or not good are a large piece of the funding. So, it’s unfold out over many various issues, however we’ve definitely spent some huge cash on that exact problem.

Inform me about these attorneys. You rent attorneys to principally do doc overview of the AI? Are they very senior attorneys? Are they moonlighting from large corporations? Are they a bunch of junior associates in a basement?

It’s primarily based on the duty. What we attempt to do is get attorneys which have expertise in a selected matter. So, if we’re taking a look at paperwork associated to a mergers and acquisitions transaction, we wish these to be checked out by somebody who has some expertise in mergers and acquisitions. They will inform us that the doc appears nice, or inform us if it’s lacking specific issues. Then, we will return and say, “Why did we miss these specific issues and what adjustments do we have to make to how we’re coaching and directing these fashions to appropriate that scenario going ahead?”

What’s the largest factor you’ve discovered from that course of?

The largest factor I’ve discovered is how vital it’s to have attorneys doing that work. I anticipated to rent plenty of technical folks and information scientists to do that work. I didn’t actually anticipate to rent a military of attorneys. However I believe one of many secret sauce elements of our answer is that our outputs are lawyer reviewed. That’s how we preserve getting extra related outcomes.

The place had been you finest at to start out with and the place had been you worst at to start out with?

We weren’t actually good at something to start with, and I believe we’re constructing issues out. Typically it’s a apply space, generally it’s a activity. When you take a look at all of the completely different duties attorneys try this we had been speaking about earlier, in lots of circumstances the duty’s output is a few form of a doc. So, we’re actually targeted proper now on enhance our doc drafting.

Is all this income optimistic but? Are you getting cash on all this funding or do you see that on the horizon?

Our progress charge is unquestionably accelerated because of this. The principle factor that we’re targeted on is the shopper final result. What we’re seeing is that the purchasers are getting happier and happier with the answer, so I might say that it’s been very profitable in that regard. It’s the quickest rising product that we’ve ever had.

Rising quick however shedding cash with each question is unhealthy, proper?

We’re not there. We’re not shedding cash with each question.

Are you breaking even or are you getting cash?

Particularly on AI instruments, or total?

Most of our funding is in AI instruments.

Let me take the final bit right here and zoom out much more broadly. I discussed that I might deliver up precedent once more on this dialog.

I believe for those who’re being attentive to the authorized system of America proper now, you recognize that it’s just about in a state of pure upheaval. You’ve acquired district courtroom judges calling out the Supreme Court docket, which isn’t a factor that often occurs. You’ve a Supreme Court docket that’s overturning precedents in a means that makes me really feel like I discovered nothing in regulation faculty. Chevron deference is out the door. Humphrey’s Executor, the regulation that retains the president from firing FTC commissioners is, I’m guessing, out the door. Roe v. Wade was out the door. Simply these foundational precedents of American regulation out the door.

A whole lot of that’s primarily based on what conservative judges would name originalism. I’ve plenty of emotions about originalism, however a giant development inside originalism is utilizing AI, or what they name “corpus linguistics,” to find out what folks meant previously. Then, you are taking the AI and also you say, “Properly, it did the job for me. That is the reply.” Are you anxious that your instruments shall be used for that form of effort? As a result of it actually places plenty of stress on the AI software to know plenty of issues.

I’m not that anxious. I don’t suppose the Supreme Court docket is asking LexisNexis what we expect it ought to do.

However definitely courts up and down the chain are.

They’re asking authorized questions, they’re getting solutions again, after which they’re deciphering these solutions. We’re offering them with the uncooked content material that they should make the determinations, however we’re not training regulation. We’re not making these selections for them.

I’m going to spring this on you, however right here it’s: John Bush is a Trump-appointed decide. He cited the emergence of corpus linguistics within the authorized subject, and he stated, “To do originalism, I need to undertake the extremely laborious and time-consuming strategy of sifting by means of all this. However what if AI had been employed to do all of the overview of the hits and compile statistics on phrase assembly and utilization? If the AI could possibly be trusted, that may make the job a lot simpler.”

That’s him saying, “I can outsource originalist considering to an AI.” This can be a development. I see this notably with the originalist judges, that the job they suppose they’re meant to do is decide what a phrase meant previously. And AI is nice at being like, “Statistically, that is what that phrase meant previously, and we’re going to outsource some authorized reasoning.”

That is, I believe, very odd. My ideas about originalism and stare decisis in America in 2025 apart, saying, “I’ll use an AI to succeed in into the previous and decide this which means” appears very odd. I’m questioning how you’re feeling about your software being utilized in that means.

I positively perceive your level there. I take into consideration the analogy of a brick. You should utilize a brick to construct a hospital and maintain sick youngsters, or you would take a brick and throw it by means of a window. One use is admittedly nice and one other is fairly damaging, however in both case, it’s a brick. I take into consideration our software as being neither good nor unhealthy. I believe it could possibly be used for good. I believe it could possibly be used for any sort of exercise that attorneys [need]. I wouldn’t need to say originalism is a foul factor. I believe it could possibly be used for a lot of various things. I believe it could possibly be used for originalism. I believe it could possibly be useful for many who need to take that path and discover a new means of taking a look at one thing.

We’ve all the information. They will search it, they will use the software to search out issues it wasn’t doable to search out previously. So, I may see them utilizing our software in that means. I suppose it’s as much as the attorneys to find out how they’re going to make use of the product. We’re not constructing it as a result of we’re attempting to vary the regulation. We’re constructing it as a result of we’re attempting to assist attorneys do the duties that they need to do.

However I take a look at the sweep of the tech business — not the authorized business, however the tech business — over the previous 15, 20 years, and boy, have I heard that reply many, many occasions. The social media firms all stated, “Properly, you need to use it for good or evil. We’re impartial platforms.” It seems perhaps they need to have considered a few of these harms earlier.

Have a look at the AI firms at present. Who is aware of if coaching is copyrighted. We all know the reply. You’ll be able to’t truly simply decide out of copyright regulation. Now, we’re going to do the lawsuits and we’ll see what occurs. Who is aware of if OpenAI doing Sora, which is TikTok for deepfakes — truly, we all know. We all know the reply is, you need to have some guardrails.

So, I’m posing you a similar query. We see a very originalist judiciary hell-bent on utilizing originalism to vary precedent at alarming charges. I might say it’s alarming for me as a result of I paid for a regulation diploma that I now suppose is ineffective, however that’s why it’s alarming to me. It’s alarming as a result of lots of people have had their rights taken away as effectively. Every single day that is occurring. And one of many methods they’re going to do this is to defer to an AI determination engine. They’re going to say, “We requested the AI, ‘What did “all folks” imply when the 14th Modification was drafted?’” and this shall be how we get to a birthright citizenship case. I’m simply connecting this to the dialog we had originally. We’re going to provide our reasoning to a pc in a means that it’s not essentially accountable for, and we’re going to belief the pc. The strategies of considering and that rigor may go away.

So I’ve heard the reply that the software is impartial from tech firms for years, and I’ve seen the outcomes. I’m asking you. You’re constructing a tech product for attorneys, and so they’re already utilizing it on this particular means. I’m questioning for those who’ve thought concerning the guardrails.

We function beneath accountable AI rules, and that features plenty of issues. One, we at all times attempt to contemplate the real-world implications of any product we develop. We need to guarantee that there’s transparency when it comes to how our product works. We open up the black field so folks can see the logic that we’re utilizing, and so they can truly go in and alter it if they need. So, we need to guarantee that there’s transparency and there’s management.

We at all times incorporate human oversight into product improvement. Privateness and safety is one other one in every of our core tenets in accountable AI creation. One other factor we’ve included is the prevention of bias introduction. So, these are the RELX rules for AI improvement, and we adhere to these. We need to create merchandise that do good issues for the world.

When you requested Lexis AI if the 14th Modification ensures birthright citizenship to all folks born in america, will it make the argument that it doesn’t?

I’ve by no means requested it that query. I can’t let you know.

Do you’ve got your cellphone on you? There’s a cell app.

I may pop up right here and ask it, I suppose. Let me pop into Protégé right here. “Does the 14th Modification assure birthright citizenship or are there exceptions?” Let’s see.

It’s producing a response, so we will come again to it in a minute.

I’m very curious to see what it says, as a result of up till just lately there’s solely been one reply to that query. Now, the Trump administration is saying, “Nope, truly, that’s not what ‘topic to the jurisdiction thereof’ means.” With a purpose to win on the Supreme Court docket, they’ll have to assemble an originalist argument to that query, and I’m assured that the way in which they’re going to do this is by feeding a bunch of knowledge into an AI mannequin and saying, “That is what was truly meant on the time of the 14th Modification’s drafting.” That’s a factor that AI shall be used for that’s very damaging.

I’m not an lawyer, so I’m simply going to learn the reply right here:

“The 14th Modification of america Structure ensures birthright citizenship to all individuals born or naturalized in america, and topic to its jurisdiction. The phrase ‘topic to its jurisdiction’ has been interpreted to incorporate almost all people born on US soil with just a few slim exceptions. These exceptions embrace overseas diplomats, youngsters of overseas diplomats, youngsters of enemy forces in hostile occupation, youngsters born on overseas public ships, and, traditionally, youngsters of members of Native American tribes who owed allegiance to their tribe somewhat than america.”

It’s best to ship that to [Chief Justice] John Roberts proper now. Can Protégé try this? As a result of that’s the reply.

The query is, are a bunch of conservative influencers going to say Protégé is woke now? That is the cultural conflict that you simply’re in.

It does acknowledge that “latest circumstances have affirmed this interpretation rejecting makes an attempt to broaden the exceptions of birthright citizenship,” so it does additionally acknowledge that there have been efforts to interpret it otherwise. The reply goes on fairly a bit.

The rationale I ask that query very particularly is as a result of Reconstruction is up for grabs in a really possible way on this case. Do you suppose you’ve got a duty because the software maker? That’s actually the query for therefore many AI firms. You’re the software maker. Do you’ve got a duty to not deepfake actual folks? Do you’ve got a duty to not present folks pretend concepts? I believe you had been very clear on that, you’ve got a duty to not hallucinate, however right here you’ve got —

We don’t need to introduce or perpetuate any bias that may exist both. And to do this, we depend on the regulation versus a consumer-grade mannequin that in all probability simply makes use of information articles, which could have a really completely different interpretation of issues relying on the information articles. There are more likely to be biases from launched information articles than black-letter regulation, for instance.

The rationale I’m interested by that’s as a result of there’s a spectrum. I don’t suppose there’s anyplace for telling folks what they will do with Microsoft Phrase working regionally on their laptop computer. Do what you’ve acquired to do. Telling folks what they will do with a consumer-grade AI software constructed into Fb? I believe Fb has plenty of duty there, particularly as a result of the chance to distribute that content material far and large is at their fingertips.

That’s a giant alternative spectrum, and right here within the center there’s these AI firms. Do you’ve got the duty to say, “Properly, if you wish to go make the argument that birthright citizenship doesn’t defend everybody in america, you’ve acquired to do this by yourself. Our robotic’s not going that will help you.” Do you’re feeling any of that stress?

We strive to not get into politics or any of that debate.

I don’t suppose that’s politics.

We’re attempting to develop a system that doesn’t have bias launched into it, that will provide you with the details, and attorneys can do the work that attorneys do to make these vital selections. Our job is to provide them the knowledge that they want: the precedents, the details, all the knowledge that they should then develop their argument, no matter that could be. However we actually don’t get into any of the politics of birthright citizenship being assured or not.

Properly, in some unspecified time in the future you do. That is — once more, to deliver us again the place we began, I first encountered LexisNexis as a database of circumstances and a few case notes. There have been some regulation professors who had been very proud that their case notes had been in LexisNexis once I was in regulation faculty. Now we’re drafting a little bit bit, going to go do the analysis. Now we have now a agentic AI that’s making the arguments. Possibly in the future we’ll automate all the way in which to submitting. You’re taking over extra of the burden. You make the arguments. The corporate is making the arguments. The place is the road? As a result of there are many attorneys who wouldn’t take that case, who wouldn’t make that argument. Is there a line for you?

I might say our strategy is to arm the attorneys with the very best info, and assist them with the drafting of these paperwork. We’re actually simply being led by our clients and what they’re asking us to do. We definitely are usually not attempting to interpret the regulation. We’re not attempting to form the authorized system. We’re not attorneys. We’re not attempting to do the work of attorneys. We’re attempting to assist attorneys do the work they do in a extra environment friendly means and, hopefully, assist them drive higher outcomes.

Nevertheless it’s at all times their prerogative to interpret the knowledge that we offer, which is what attorneys do. That’s what they’re nice at. The rationale we have now circumstances is as a result of there are folks on each side. The 2 people are going to make reverse arguments, we need to help each of these attorneys as finest that we will.

I get it if you’re the database of circumstances. I get it if you’re the phrase processor. I get it if you’re the specialised phrase processor or the case administration platform. The factor that I’m pushing on repeatedly right here is that if the AI system is definitely doing the work, do you’re feeling like you’ve got completely different guardrails?

I believe our duty is to develop AI in a accountable means.

Give me an instance of one thing you wouldn’t let your AI do — an argument that you simply wouldn’t let your AI make, or a movement that you simply wouldn’t let your AI draft.

I don’t know that we’d need to essentially prohibit the AI in that means. We’re referring again to the knowledge that we have now, which is our authoritative assortment of paperwork and supplies that helps attorneys perceive what the details are, what the precedent is, and what the background is, to allow them to do the true, deep authorized work and make these trade-off selections, judgment selections, the vital issues that, once more, attorneys went to regulation faculty to do.

I believe these questions are going to come back up again and again. We should always have you ever again to reply them as you study extra. As you look out over the horizon — the subsequent two or three years — what’s the subsequent set of capabilities you see for LexisNexis, and what do you suppose the pressures that may change the way you make a few of these selections shall be?

It’s arduous to say precisely what the principle factor that’s going to vary the trail going ahead goes to seem like as a result of if I look again two years in the past, I might’ve by no means guessed we’d be doing what we’re doing at present as a result of the know-how didn’t exist or it was too costly to implement. That’s completely modified over the past two years, and I believe over the subsequent two years, it’s going to vary once more. So, it’s actually arduous to say the place we’re going to go. Our imaginative and prescient stays the identical, which is that we need to assist attorneys. We need to present them with a personalised, AI-powered product that understands their apply space, their jurisdiction. It has entry to our authoritative set of supplies and their prior work product. It understands their preferences, it understands their fashion, it understands what they’re attempting to do, and it may well automate duties that they do at present manually.

We are going to proceed to take that newest accessible know-how, present it to our clients, and have them assist us perceive how we will use that know-how to serve them in additional trendy and related methods. That’s actually what’s going to information our roadmap sooner or later.

Sean, this was nice. Let me know if you develop a system that may truly navigate an digital case submitting web site as a result of among the smartest folks I do know can’t try this. However this was nice. We’ve acquired to have you ever again quickly. Thanks a lot.

Thanks a lot. I actually loved our time at present. Take care.

_{Questions or feedback about this episode? Hit us up at decoder@theverge.com. We actually do learn each e mail!}