Lexisnexis Ceo Says The Ai Law Era Is Already Here

Today, I’m talking pinch Sean Fitzpatrick, nan CEO of LexisNexis, 1 of nan astir important companies successful nan full ineligible system. For years — including erstwhile I was successful rule schoolhouse — LexisNexis was fundamentally nan library. It’s wherever you went to look up lawsuit law, do ineligible research, and find nan laws and precedents you would request to beryllium an effective lawyer for your clients. There isn’t a lawyer coming who hasn’t utilized it — it’s basal infrastructure for nan ineligible profession, conscionable for illustration email aliases a connection processor.

But endeavor companies pinch immense databases of proprietary accusation successful 2025 can’t defy nan siren telephone of AI, and LexisNexis is nary different. You’ll perceive it: erstwhile I asked Sean to picture LexisNexis to me, nan first connection he said wasn’t “law” aliases “data,” it was “AI.” The extremity is for nan LexisNexis AI tool, called Protégé, to spell beyond elemental research, and thief lawyers draught nan existent ineligible penning they taxable to nan tribunal successful support of their arguments.

That’s a large deal, because truthful acold AI has created conscionable arsenic overmuch chaos and slop successful nan courts arsenic anyplace else. There is simply a accordant drumbeat of stories astir lawyers getting caught and sanctioned for relying connected AI devices that mention hallucinated lawsuit rule that doesn’t exist, and location person moreover been 2 tribunal rulings retracted because nan judges appeared to usage AI devices that hallucinated nan names of nan plaintiffs and cited facts and and quoted cases that didn’t exist. Sean thinks it’s only a matter of clip earlier an lawyer location loses their licence because of sloppy usage of AI.

So nan large committedness LexisNexis is making astir Protégé is simply accuracy — that everything it produces will beryllium based connected nan existent law, and overmuch much trustworthy than a wide intent AI tool. You’ll perceive Sean explicate really LexisNexis built their AI devices and teams truthful that they tin make that committedness — LexisNexis has hired galore much lawyers to reappraisal AI activity than he expected, for example.

But I besides wanted to cognize what Sean thinks devices for illustration Protégé will do to nan occupation of rule itself, to nan occupation of being a lawyer. If AI is doing each nan ineligible investigation and penning you’d usually person inferior associates doing, really will those inferior associates study nan craft? How will we create caller elder group without a pipeline of inferior group successful nan weeds of nan work? And if I’m submitting AI ineligible penning to a judge utilizing AI to publication it, aren’t we getting adjacent to automating a small excessively overmuch of nan judicial system? These are large questions, and they’re coming existent accelerated for nan ineligible industry.

I besides pressed Sean beautiful difficult connected really judges, peculiarly blimpish judges, are utilizing computers and exertion successful work of a judicial mentation called originalism, which states that laws tin only mean what they meant astatine nan clip they were enacted. We’ve run stories astatine The Verge astir judges letting automated linguistics systems effort and understand nan originalist intent of various statutes to scope their preferred outcomes, and AI is only accelerating that inclination — particularly successful an era wherever virtually each portion of nan Constitution appears to beryllium up for grabs earlier an incredibly partisan Supreme Court. 

So I asked Sean to demo Protégé doing immoderate ineligible investigation for me, connected questions that look to beryllium settled but are recently up for grabs successful nan Trump administration, for illustration birthright citizenship. To his credit, he was crippled — but you tin spot really taking nan institution from 1 that provides elemental investigation devices to 1 that provides existent ineligible reasoning pinch AI will person large implications crossed nan board.

This 1 is weedsy, but it’s important. 

Okay: LexisNexis CEO Sean Fitzpatrick. Here we go.

This question and reply has been lightly edited for magnitude and clarity.

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

Thank you. Great to beryllium here.

Thank you for joining me. This is my first question and reply backmost from parental leave. Apologies to nan assemblage if I’m rusty, but apologies to you if I’m conscionable wholly loopy.

Congratulations!

I’m very excited to talk to you. I’m very overmuch a grounded lawyer, my woman is simply a lawyer, there’s a batch of lawyers connected The Verge team. The ineligible occupation successful America is astatine a infinitesimal of absolute change, a batch of chaos, and an tremendous magnitude of uncertainty. And LexisNexis, if nan assemblage knows, tends to beryllium astatine nan bosom of what lawyers do each day. Most lawyers are utilizing LexisNexis each infinitesimal of each day. What that merchandise is, what it tin do, and really it helps lawyers do their occupation connects to a batch of themes that we’re seeing some successful nan ineligible occupation and successful exertion and AI generally. 

So, commencement astatine nan start. What is LexisNexis? How would you explicate it to nan layperson?

LexisNexis is an AI-powered supplier of information, analytics, and drafting solutions for lawyers that activity successful rule firms, corporations, and authorities entities.

That’s a caller conception of LexisNexis. When I was successful rule schoolhouse successful nan early 2000s, it was conscionable nan point I searched to find lawsuit law.

Yes, we’ve transformed complete time. We were conscionable that investigation provider. Over time, we’ve acquired and integrated much businesses. In 2020, erstwhile we launched our Lexis+ product, we integrated each those things together, becoming an integrated ecosystem of solutions. Then, successful 2023, we launched Lexis+ AI, and that’s erstwhile we really became an AI-powered supplier of accusation analytics, determination tools, and drafting solutions. AI capabilities person really allowed america to do much things than what we’ve traditionally done successful nan past

That jump from being nan golden modular database of ineligible opinions, reasonings, lawsuit notes, and each that to “we’re going to do nan activity for you aliases thief you do nan work” is simply a large one. That’s a taste jump. Obviously, location were immoderate acquisitions on nan measurement that helped you to make that jump, which you tin talk about. What drove you to make that jump, to say, “Actually, nan lawyers request thief drafting nan motions, nan projected opinions they mightiness springiness to a judge?” What made you say, “Okay, we’ve sewage to measurement into really doing nan work?”

I deliberation it’s been a earthy evolution. As exertion has evolved, it’s opened up caller avenues of things we tin do. We thin to return nan latest technology, present it to our customers, and walk clip talking to them astir really they deliberation that exertion tin beryllium champion applied successful nan ineligible environment. Then, we construe nan ideas they came up pinch into products that resoluteness aliases reside those opportunities.

Let maine inquire you a beautiful philosophical question. It’s 1 that I struggle pinch each nan time. It’s 1 that I talk to our assemblage astir each nan time. I deliberation astir group who brushwood nan ineligible strategy deliberation it’s beautiful deterministic. Our assemblage is beautiful technically focused. They’re utilized to computers. Computers are, until recently, beautiful deterministic. You put successful immoderate inputs, you get immoderate outputs. Most group who brushwood nan ineligible strategy deliberation it’s beautiful deterministic. You put successful immoderate inputs and you get immoderate predictable outputs.

And what I’m ever saying is, “That’s not really it useful astatine all.” You show up to court, nan judge is successful a bad mood, you person nary thought what’s going to happen. You’re a large institution pinch an antitrust appeal, you show up to nan three-judge appellate reappraisal board, and you person nary thought what’s going to happen. Literally thing could hap astatine immoderate time. The judicial strategy is fundamentally not deterministic. Even though it’s system for illustration a computer, trying to deliberation astir it for illustration 1 tin get you successful each kinds of trouble. Maybe nan champion illustration of this is group connected Facebook putting nan words “no copyright intended” connected nan bottommost of movies. They deliberation they tin rumor these magic words and nan ineligible strategy is solved, and they conscionable can’t.

AI is that problem successful a nutshell. We’re going to return a computer, make it amended astatine earthy language. We’re going to make nan machine fundamentally not deterministic — you can’t really foretell what an AI is going to do — and past we’re going to use that to nan fundamentally non-deterministic, quality quality of nan tribunal system. Somewhere successful location is simply a large philosophical problem astir applying computers to nan justness system. How do you deliberation astir that?

First of all, you person these monolithic investments happening pinch nan foundational models. Each of these hyperscalers — Microsoft, Amazon, Google — are putting successful adjacent to $100 billion. So these models conscionable proceed to get amended and amended complete time. That’s astatine nan foundational exemplary level. We don’t really run astatine that level. We build applications that utilize these foundational models. And astatine that level, we spot prices are dropping. We utilized to salary $20 for 1 cardinal tokens 2 years ago, and coming we mightiness salary 10 cents for 1 cardinal tokens. That allows america to do things astatine velocity and astatine standard that we’ve ne'er been capable to do before. 

And location are a batch of things astir nan rule that make these models attractive. Most of nan rule is language-based, and these models are really awesome pinch connection problems. The rule is precedent-based, and truthful — 

Well, that’s up for grabs. We’ll travel backmost to it.

I’ll assistance you that. You look astatine nan activities that lawyers do: they draught documents, they do research, they summarize things. The models are each really bully astatine these types of things. So, you person this cleanable storm, pinch this exertion and nan things lawyers do coming together. 

Yet, erstwhile group effort to usage these consumer-grade models, location are each kinds of problems pinch them. Like you said, it’s not deterministic. You can’t conscionable put accusation into a machine and get an reply out. If that were nan case, we wouldn’t request a tribunal system. These models are conscionable not built for nan ineligible system. You can’t spell into tribunal and say, “I recovered this connected nan internet.” You person to person charismatic content.

The cut-off day for GPT-4o was 2023, I believe. You request to person accusation that’s perpetually updated. Your assemblage astir apt doesn’t cognize this,, but there’s nan citator, which traditionally has said, “This is bully law” aliases “It’s not bully law, it’s been overturned.” Now, it’ll show you if it’s nan rule astatine each aliases if immoderate strategy conscionable made it up. These systems are probabilistic. They want to put together an reply that’s astir apt right. Well, that’s not nan modular we person successful legal. You can’t spell successful pinch thing that’s astir apt right. So, you person this full database of issues that these models don’t address. 

What we’ve tried to do is reside those pinch a courtroom-grade solution. Our strategy is backed by 160 cardinal documents and records. Our curated postulation is our grounding data. So can’t spell into tribunal and say, “I recovered this connected nan internet,” but you tin mention to a circumstantial case. We besides person what we telephone a citator supplier that’ll cheque that lawsuit to make judge that it wasn’t fabricated by nan strategy and is really still bully law. You tin besides look astatine nan lawsuit rule summary truthful you cognize what nan lawsuit is about. You tin look astatine nan headnotes truthful you tin spot nan peculiar points of rule that were addressed successful that lawsuit and spot if it’s still a valid case. 

Privacy is different issue. There’s a typical narration that exists betwixt attorneys and their clients successful that attorney-client privilege, truthful there’s immoderate privateness requirements that you request successful bid to support that. If you’re utilizing 1 of these conscionable consumer-grade models, you don’t person nan level of privateness and information that you need. Transparency is different issue. You put a mobility in, you get an reply back. Well, based connected what? What was nan logic that nan strategy used? We unfastened up nan achromatic container truthful you tin spot nan logic that’s being applied. We springiness nan attorneys nan expertise to spell successful and really alteration that. If this exemplary is getting thing wrong, nan lawyer has nan opportunity to alteration it truthful that they get nan result that they’re driving for. But, arsenic you said, nan rule is not deterministic. There are tons of different factors that spell into this, but you request to person a strategy that’s legally driven, that’s intent built for ineligible situations successful bid to really run successful a courtroom-type environment.

There’s 2 things I really want to push on. Again, I was not a bully lawyer. I don’t want to ever dress connected nan show, to you, aliases to anyone other that I was immoderate bully astatine this. But you study a peculiar measurement of reasoning successful rule school, which is simply a beautiful rigorous, system measurement of approaching a problem, going to find nan applicable cases and precedents, and past trying to manner immoderate solution based connected that. That feels for illustration we’re conscionable moving words around, but it’s really a measurement of thinking. Before AI showed up, we would mash together utilizing a connection processor and reasoning a definite way. Now, we’re pulling them apart. We’re saying that nan machine tin move nan words astir and make immoderate thinking. So, that’s 1 point I want to push on. I’m very funny astir that because it feels for illustration nan lawyering portion of being a lawyer is being subsumed into a system, and that mightiness alteration really we lawyer. 

The different portion is if anyone is going to look astatine nan activity being done. We’re already seeing lawyers get sanctioned for filing briefs pinch hallucinated lawsuit citations successful them. There was conscionable a lawsuit where, I believe, a tribunal had to rescind an sentiment because it had a hallucinated lawsuit citation successful it. This is bad. This is conscionable straightforwardly a threat to nan strategy and really we mightiness deliberation astir lawyers, judges, and courts. It’s not clear to maine that anyone’s going to usage nan devices arsenic rigorously arsenic you want.

So connected nan 1 manus there’s “We’ve made nan reasoning easier.” On nan different hand, it’s, “Oh, boy, everyone’s going to get really lazy.” They’re some successful your answer. They’re some saying, “We’re making it easier to look astatine this stuff. We’re making it faster to do nan research.” I’m conscionable wondering wherever you deliberation nan reasoning comes in.

I don’t deliberation that these models switch nan lawyers. I deliberation they thief nan lawyer and augment what nan lawyer does. So, if you deliberation astir an activity that a lawyer mightiness do — let’s opportunity they were preparing for a deposition. They request to travel up pinch a database of questions that they’re going to inquire nan individual that’s being deposed. You tin return nan facts astir that peculiar case, load them into a vault, and constituent nan strategy to that vault and say, “Based connected nan facts of this peculiar matter, create a database of deposition questions.” That’s thing that a lawyer would’ve done connected their own. In nan past, they whitethorn person referred to a database of questions that they had antecedently aliases thing —

Actually, tin I conscionable drawback that example? Maybe a lawyer would’ve done that, but much often a lawyer would’ve told a bunch of inferior associates to beryllium successful nan basement and do that. That was really those inferior associates learned really to do their job. That’s what I mean. We’re farming retired nan thinking, and immoderate group mightiness ne'er really do that thinking. That mightiness alteration nan occupation down nan statement successful really substantive ways.

Right. It is an apprentice system. So, if you commencement to return immoderate of nan layers retired of nan bottom, really does everyone skip nan bottommost furniture and still make it to nan 2nd pinch nan aforesaid capabilities and skills? That’s a existent challenge. I deliberation nan systems are allowing lawyers to not person nan subordinate do that work. Now they tin say, “Generate maine 300 aliases 700 questions.” It doesn’t return that agelong to spell done 700 questions, and nan models ne'er get tired. From our experience, they’ll spell done that database of questions and say, “First question? Yep, that’s a bully question. I would’ve thought of that.” The strategy made it a small spot faster, but it didn’t really thief them. Second question, aforesaid thing. Third question, aforesaid thing. Fourth mobility doesn’t moreover make immoderate sense, scratch it disconnected nan list. With nan 5th mobility they’ll say, “Oh, that’s interesting, I wouldn’t person thought to inquire that but that’s astir apt important, truthful I’m going to adhd that to my list.” So, there’s an ratio constituent to it, but I deliberation there’s besides a amended result component. 

In position of nan apprenticeship piece, I deliberation group are struggling correct now to fig retired really that’s going to effect nan apprenticeship model. Someone was describing to maine that they had worked connected a business wherever they were looking astatine securitized assets. When they were an associate, they did this task for a institution that had 50 states worthy of coverage, and truthful they became nan master successful nan patient connected plus securitization successful each 50 states. For 4 aliases 5 years, anytime personification had a question, they came to that individual. It was a awesome measurement to make a career. Now, nan strategy tin do each that accusation for you. So his mobility was: “How is that ever going to hap now successful this caller world?” 

I deliberation firms are going to struggle pinch that, but I besides deliberation they’re going to fig it out. We thin to get immoderate of nan smartest and brightest group going into nan ineligible profession, and truthful far, they look to person figured retired each situation that’s faced nan industry. I deliberation they’ll fig this 1 retired arsenic well.

What are immoderate solutions you’ve seen arsenic group effort to fig this out?

I don’t cognize that folks person travel up pinch a batch of solutions astir nan apprenticeship model. What we’re for judge seeing is that group are embracing AI. It’s here, it’s successful nan courtroom, it’s successful nan rule firm. Two-thirds of attorneys are utilizing AI successful their work, according to our surveys, and our survey’s astir apt a small outdated. I’d opportunity nan number’s astir apt higher. I don’t cognize astir you, but I usage AI each day.  It’s now successful my individual and activity lives. I deliberation nan ineligible occupation is perfectly suited for it, truthful it’s only going to expand.

When you spot nan lawyers getting sanctioned and nan courts having to rescind opinions, is location a solution that involves utilizing LexisNexis truthful it won’t hap to you? Or do you deliberation that’s a denotation of thing else? Everyone’s conscionable utilizing AI, I get it. Probably nan biggest divided for our assemblage correct now is betwixt nan information that says everyone’s utilizing this worldly each nan clip and nan hostility our assemblage expresses astir nan tools, their quality, and nan truth that a batch of that usage is driven by large companies conscionable putting it successful beforehand of them. There’s thing happening where, to warrant these tremendous investments, nan devices are showing up whether nan consumers are asking for them aliases not, and past we’re pointing retired that everyone’s utilizing nan tools.

What I perceive from our assemblage is, “Well, I can’t move disconnected nan AI overview. Of people I’m utilizing nan instrumentality because it’s conscionable successful beforehand of maine each nan time. I can’t make Microsoft Office extremity telling maine that it’s going to thief me. It’s conscionable successful beforehand of maine each nan time.” So, erstwhile you spot nan errors being made successful nan ineligible strategy coming — nan lawyers getting sanctioned, nan lazy AI use, nan deficiency of apprenticeship that’s going to effect nan full adjacent procreation of lawyers and really rigorous they are —  really do you make your merchandise reside that? Or are you conscionable not reasoning astir that correct now?

No, we’re decidedly reasoning astir it, and we’ve incorporated things into our product. These things ever make nan headlines erstwhile they happen, but I deliberation it’s a mini percent of attorneys that are causing these problems. Just taking thing and bringing it into tribunal has ne'er been nan standard. You’ve ever had nan work arsenic a lawyer to cheque nan worldly and make judge that it’s valid earlier going into court. And immoderate individuals aren’t doing that. We surely saw that successful nan [Tim] Burke case wherever immoderate attorneys submitted a archive to nan tribunal that I deliberation had 8 citations successful it and 7 of them were conscionable wholly — 

But that was inevitable. The time ChatGPT showed up, half of nan ineligible pundits I cognize were like, “This is inevitable. This result will happen,” and past it happened. There wasn’t moreover a stutter step. It conscionable happened immediately. That’s what I’m trying to push on. Is nan solution conscionable that LexisNexis has a instrumentality that’s amended and you should salary for it, aliases is nan occupation going to person to build immoderate caller guardrails arsenic we return nan rigor distant from nan younger associates?

Well, you tin ne'er extremity an lawyer from taking it into tribunal and not doing nan due work. That’s going to proceed to happen. I deliberation somebody’s going to suffer their licence complete this astatine immoderate point. We’re seeing nan sanctions commencement to ratchet up. So, a couple attorneys sewage fined $5,000 a piece, and past immoderate attorneys successful national tribunal down successful Alabama got referred to nan authorities barroom relation for disciplinary action. I deliberation nan stakes are expanding and increasing. What we do pinch our strategy is supply a nexus to a citation if we person it, truthful you tin click connected it and spot it successful our system.

And there’s nary fabricated cases wrong our system. We person a postulation system that ensures that each lawsuit successful location is valid. It’s shepardized and has headnotes and different devices that lawyers tin use. So, we make it really easy for you to usage our strategy to cheque and make judge that nan citations you’re bringing into tribunal are not only valid and still bully rule but are besides successful nan correct format. Format is important. We cheque for each these things and make it really easy for nan lawyer to do nan activity they request to do. They request to make judge that lawsuit is connected point, that nan lawsuit is still valid.

One of nan galore reasons I was a horrible lawyer was because of that infinitesimal erstwhile you get your first rule patient occupation and you recognize your leader conscionable has a room of their favourite motions connected file. They’re conscionable going to propulsion from nan paper catalog, alteration immoderate names and dates, and record nan motion. The judge will admit nan mobility and nan attorney, and this is each conscionable a weird formality to get done nan adjacent shape of nan process. Maybe we’ll ne'er get to nan substantive portion of nan lawsuit because we’re conscionable going to settee it, but we request to record this mobility we had and elaborate. This genuinely was demoralizing. I was like, “I’m conscionable doing paperwork. There’s thing astir this that is real.”

I’m astir apt describing what each first twelvemonth subordinate goes done until nan cheque hits, and it conscionable didn’t activity for me. How adjacent are you to having Lexus AI conscionable do that thing, person it admit nan infinitesimal and say, “We person nan banked mobility and we’re conscionable going to record it to a system?”

Well, we tin link into a archive guidance strategy (DMS) that has an attorney’s anterior motions. We person our vault capabilities, truthful they tin load their motions up. They tin still usage nan motions they’ve already developed. And that’s a perfectly good measurement to do things because —

Well, I’m saying, from scratch.

Right. We person nan expertise to do it from scratch too, but a batch of attorneys don’t want to do it from scratch because they’ve reviewed each azygous connection successful that mobility and they cognize that it’s good. If they do it from scratch, past they person to reappraisal each azygous word. But if they want to do it from scratch, we tin do that for them today, and if they want, we tin usage their anterior activity merchandise arsenic nan grounding contented to create a caller motion, aliases we tin usage our charismatic material. They tin take nan root and nan grounding content.

I conjecture I’m asking what level of automation is there. So, you’re an attorney. You’ve sewage a archive guidance system, you’ve sewage a caller client, and you request to record immoderate standardized mobility that you ever record for immoderate point you request to do, for illustration a continuance. At what constituent does Lexis [AI] say, “I’m watching this case. I’m going to record this for you. I’m conscionable going to deed nan buttons for you. Don’t interest astir it,” successful nan measurement that a awesome ineligible adjunct mightiness do?

We’re ever going to springiness nan lawyer nan opportunity. We don’t want to conscionable beryllium doing things connected their behalf unsupervised, truthful we’re going to springiness them nan opportunity. We could get to nan constituent wherever we say, “It looks for illustration you request a continuance. Here’s a draught of a continuance push,” and it will automatically record it. We’re not astatine that constituent today, but if you request a continuance, we could draught it for you. Our imagination is that each lawyer is going to person their ain personalized AI assistant, and it’s going to understand their believe area and their jurisdiction, on pinch having entree to their anterior activity product. The systems are only arsenic bully arsenic nan contented down them, truthful it’s going to person entree to our 160 cardinal documents and records, and it’s going to beryllium capable to automate tasks that they do today. If you deliberation astir each nan different types of attorneys and each nan different tasks that they perform, there’s astir apt 10,000 tasks that could beryllium automated. 

So, we’re moving pinch our customers to understand what nan astir important tasks are, and we’re moving pinch them to automate those tasks today. We person nan largest and astir robust backlog of projects that we’ve ever had successful our company’s history because location are truthful galore of things that tin still beryllium automated, and we’re moving pinch our customers to do that. If they show us, “What we really want is for you to automatically record this” aliases for america to supply them pinch an alert that says, “Hey, this deadline is coming up and you request to record this. Here’s a draft. Do you want to record it?” I’m judge we tin create it.

We’re not astatine that constituent today, but we are successful nan drafting phase. That imagination is not a five-year imagination aliases a three-year vision, that’s disposable today. That’s Protégé. That’s what Protégé does today. There are tasks that it tin do, but we haven’t vanished that monolithic backlog yet.

If you look astatine nan expanse of different CEOs who’ve been connected Decoder, they’re going to show you, “You conscionable merge our machine imagination strategy and we’ll usage [electronic lawsuit files] for you to record this motion.” They’ll each beryllium very happy to waste you that product, I’m sure. 

The logic I’m asking it this measurement is because erstwhile I get nan user AI CEOs connected nan show, they emotion to show maine that they’re going to constitute my emails for maine pinch AI, and past nan adjacent condemnation they opportunity is, “Then, we’ll benignant your inbox pinch AI.” At immoderate point, nan robots are conscionable penning emails to each different and I’m reference summaries. Something very important has been mislaid successful that chain. One of nan funniest outcomes of AI is my iPhone abruptly conscionable summarizing emails and generating emails for different iPhones to summarize, and I person nary thought what’s going on. 

That’s bad successful nan ineligible context. We’re automating archive procreation to make nan lawsuit for our clients. On nan different side, nan judges and clerks mightiness beryllium utilizing these aforesaid devices to ingest nan cases, summarize them, understand nan arguments, and constitute nan opinions that are nan outcomes. Culturally, I deliberation it’s important for you to person a constituent of position connected wherever that should extremity because different we are conscionable going to person a afloat automated justness strategy of LLMs talking to each other. Maybe there’ll beryllium immoderate guardrails that different group don’t have, but we’ve taken an tremendous magnitude of humans retired of nan loop.

I deliberation you person to person nan quality successful nan loop.  It’s an important portion of nan process. I could spot nan bots going backmost and distant connected things for illustration if personification says, “Hey, tin you meet astatine 9 o’clock?” and your strategy opens up nan calendar, says you’re disposable to meet this personification connected your precocious privilege list, and sets up nan meeting. When you’re talking astir substantive ineligible matters, nan stakes are excessively high. You’re talking astir a abnormal seasoned getting aliases not getting their benefits. You’re talking astir a unfortunate of a earthy disaster getting aliases not getting security proceeds. You’re talking astir a azygous mother getting aliases not getting use benefits. These are each ineligible matters, and they really person a immense effect connected people’s lives. The stakes are measurement excessively precocious for bots to beryllium going backmost and distant and sharing information.

Do you deliberation that clerks and judges should beryllium utilizing AI nan aforesaid measurement lawyers should be? That’s wherever I would tie nan line. I deliberation nan clerk should beryllium made to publication and construe everything arsenic humans, and nan judges should beryllium made to constitute everything arsenic humans, but it doesn’t look for illustration that statement has been formalized anywhere.

I don’t deliberation a judge should constitute each line. I deliberation that they could usage AI. It’s awesome erstwhile you put concepts in, put nan words astir that conception and building them successful an orderly way. I deliberation that location is simply a constituent of nan activity that could beryllium done by AI, but it shouldn’t beryllium a bot talking to a bot. I don’t deliberation it should beryllium afloat outsourced to AI. You’ve sewage a work arsenic a judge, arsenic a rule clerk, arsenic a lawyer to reappraisal that archive and make judge it’s really saying what you intend it to say.I deliberation astir attorneys are utilizing it that way. It will create a awesome draft, possibly astatine 80 percent, which allows you to do 20 percent of nan work. But that 20 percent is nan deep, analytical thought work, nan things you really went to rule schoolhouse to do arsenic opposed to what you were describing earlier. It’s going to let lawyers to do much of that type of work.

I’m funny to spot really different jurisdictions and circuits attack nan mobility of what nan judges and clerks should beryllium doing.  I consciousness that that unit is going to definitive itself successful different ways crossed nan field. 

Judges are becoming forensic auditors. They’re reviewing this accusation looking for clone cases. We don’t want them doing that. That should not beryllium their job. I deliberation things do request to alteration successful immoderate of these areas.

Using AI to drawback AI is different taxable that comes up connected Decoder all nan time. 

I person utterly forgotten to inquire you nan Decoder questions. So fto maine do that, and past I want to zoom retired a small spot farther. These are my ain questions. You tin tell, I’m a small rusty.

I’m looking astatine nan LexisNexis activity structure, it’s very complicated. There’s a CEO who’s not you, Mike Walsh, but past you’re nan CEO of nan US and nan UK. There’s a bunch of different VPs everywhere. You’ve sewage a genitor institution called RELX. Explain really LexisNexis is system and really your portion fits into it.

RELX is nan genitor company, and it’s publically traded. It has 4 divisions. Legal and Professional is 1 of those divisions, and its CEO is Mike Walsh. I study to Mike. I’m nan CEO of our North America, UK, and Ireland businesses. The measurement that we’re organized, it’s a matrix. We spell to marketplace based connected customer segments. So, we person a ample rule business, a mini rule business, a firm ineligible business, a national authorities business, a authorities and section authorities business, a Canadian business, and a UK business. 

Then, we person functional groups that support that. So, we person merchandise management, and they’re responsible for our merchandise improvement roadmap and nan merchandise strategy. We person an engineering team, and they return nan guidance from merchandise guidance but really build nan products. We person functional groups that support that, finance, HR, legal, and world operations that does things for illustration cod contented for us. Once you get utilized to it, it’s not that analyzable of a structure. It’s really good integrated and seamlessly integrated together, which allows america to run really quickly. We tin get things done quickly and efficiently. And I would opportunity that nan full process is customer driven.

I’m really willing successful nan structure, peculiarly nan truth that you person nan UK, Ireland, and North America. I’m fascinated by firm structures, and 1 of nan things that strikes maine is that you are not successful power of nan taxonomy of your product, right? These countries’ governments are successful power of nan taxonomy of their ineligible systems. The English ineligible strategy and nan American ineligible strategy person commonalities but wildly different structures. The Canadian ineligible strategy and nan US ineligible strategy person wildly different structures. Canada really has much successful communal pinch nan UK fixed their shared history. How do you deliberation astir that? Are those different teams? Do they person different database structures? How does each that work?

We do person different teams and different database structures, but we’re really trying to consolidate to nan grade that we tin because erstwhile we person akin things, we shouldn’t person them marked up otherwise successful different databases. Getting them marked up successful a accordant measurement will let america to do what we telephone “extreme reuse,” which is to fundamentally usage that aforesaid exertion we create successful aggregate jurisdictions pinch constricted changes to that system. What that allows america to do is really attraction connected that halfway strategy and rotation it retired quickly, truthful that everyone crossed nan world gets nan benefits of each those changes. But you person civilian rule successful immoderate jurisdictions and communal rule successful others, and nan laws are system successful different ways. So, you do person things that make that much challenging, but that’s nan wide thought down what we’re trying to do.

Can you use nan aforesaid AI systems to these different ineligible systems successful nan aforesaid way, aliases are you actively localizing them differently?

I would opportunity that we actively localize them, but we effort to minimize nan magnitude of activity that we do because a batch of it tin beryllium done successful a akin way.

Generally, there’s a batch of interest astir American ineligible precedents walking crossed nan ocean, peculiarly successful nan UK. You tin spot nan American civilization warfare gets exported and shows up successful a batch of different ways. Do you deliberation your instrumentality will make that amended aliases worse? If you’re not pulling them isolated and are really trying to minimize nan differences, you mightiness spot repetition arguments aliases repetition structures conscionable based connected nan measurement nan AI works.

Each 1 is based connected nan contented of nan individual jurisdiction. So, we don’t operation nan content, but we do effort to utilize nan aforesaid technology. For example, there’s hunt relevance exertion to find nan lawsuit that’s astir intimately associated pinch nan matter that personification is moving on. We tin return that and build it for nan US marketplace aliases nan UK market, and past we tin move it to different marketplace and it will activity beautiful well. Then, we request to do immoderate modifications to make it activity really good for that peculiar jurisdiction. We get 80 percent of nan DNA transferred complete successful that halfway model.

I was recently talking to Mike Krieger, who is nan main merchandise serviceman of Anthropic — conscionable a wholly different speech connected a different point — but he said this point to me, which is stuck successful my mind. He said, “I admit Claude, I tin spot Claude’s writing.” He said, “That’s my boy,” which is cute. Does your AI person a personality? Can I admit its penning successful each these different jurisdictions?

We usage a multi-model approach, and truthful it’s astir apt a small little clear which peculiar exemplary drove something. Of people pinch agentic AI, things person really changed. I deliberation that was astir apt existent a twelvemonth and a half ago, but now pinch agentic AI, erstwhile personification puts successful a query… let’s opportunity they wanted to draught a document. Maybe a customer is sending successful a petition and she’s willing successful a premises liability rumor astir nan work to pass a trespasser astir a vulnerable information connected a portion of land. The query will spell into a readying agent, who will past allocate that query retired to different agents. 

It needs to do immoderate heavy research, truthful possibly it uses OpenAI o3  because it’s really bully astatine heavy research. At nan end, it needs to draught a document, truthful possibly it uses Claude 3 Opus, which is really bully astatine drafting. We’re exemplary agnostic, and we’ll usage immoderate exemplary is champion successful a peculiar task. So, nan consequence you get backmost was really perchance done by aggregate different models, which astir apt makes it a small spot harder conscionable to spot if it was drafted by OpenAI.

Is that reflected successful your structure? You picture engineering, product, and your localization, but you’ve sewage to build that agentic orchestration furniture and determine which models are champion for each purpose. You could creation an engineering statement astir that problem specifically. Is that really you’ve done it aliases is that done differently?

We person an engineering squad that focuses connected nan planner supplier and nan duty of nan tasks to different agents.

Is that wherever nan bulk of your finance is aliases is it paying nan token fees?

I haven’t really surgery it retired that way, truthful I couldn’t show you. The token fees are surely an important portion of nan investment. Engineering is simply a immense information of nan investment. The attorneys that we prosecute to reappraisal nan output and show america if it’s bully aliases not bully are a monolithic portion of nan investment. So, it’s dispersed retired complete galore different things, but we’ve surely spent a batch of money connected that peculiar issue.

Tell maine astir those attorneys. You prosecute attorneys to fundamentally do archive reappraisal of nan AI? Are they very elder attorneys? Are they moonlighting from large firms? Are they a bunch of inferior associates successful a basement?

It’s based connected nan task. What we effort to do is get attorneys that person acquisition successful a peculiar matter. So, if we’re looking astatine documents related to a mergers and acquisitions transaction, we want those to beryllium looked astatine by personification who has immoderate acquisition successful mergers and acquisitions. They tin show america that nan archive looks great, aliases show america if it’s missing peculiar things. Then, we tin spell backmost and say, “Why did we miss those peculiar things and what changes do we request to make to really we’re training and directing these models to correct that business going forward?”

What’s nan biggest point you’ve learned from that process?

The biggest point I’ve learned is really important it is to person attorneys doing that work. I expected to prosecute a batch of method group and information scientists to do this work. I didn’t really expect to prosecute an service of attorneys. But I deliberation 1 of nan concealed condiment components of our solution is that our outputs are lawyer reviewed. That’s really we support getting much applicable results.

Where were you champion astatine to commencement pinch and wherever were you worst astatine to commencement with? 

We weren’t really bully astatine thing to statesman with, and I deliberation we’re building things out. Sometimes it’s a believe area, sometimes it’s a task. If you look astatine each nan different tasks attorneys do that we were talking astir earlier, successful galore cases nan task’s output is immoderate benignant of a document. So, we’re really focused correct now connected really to amended our archive drafting.

Is each this gross affirmative yet? Are you making money connected each this finance aliases do you spot that connected nan horizon?

Our maturation complaint is decidedly accelerated arsenic a consequence of this. The main point that we’re focused connected is nan customer outcome. What we’re seeing is that nan customers are getting happier and happier pinch nan solution, truthful I would opportunity that it’s been very successful successful that regard. It’s nan fastest increasing merchandise that we’ve ever had.

Growing accelerated but losing money pinch each query is bad, right?

We’re not there. We’re not losing money pinch each query.

Are you breaking moreover aliases are you making money?

Out profit is growing.

Specifically connected AI tools, aliases overall?

Most of our finance is successful AI tools.

Let maine return nan past spot present and zoom retired moreover much broadly. I mentioned that I would bring up precedent again successful this conversation. 

I deliberation if you’re paying attraction to nan ineligible strategy of America correct now, you cognize that it’s beautiful overmuch successful a authorities of axenic upheaval. You’ve sewage territory tribunal judges calling retired nan Supreme Court, which is not a point that usually happens. You person a Supreme Court that is overturning precedents successful a measurement that makes maine consciousness for illustration I learned thing successful rule school. Chevron deference is retired nan door. Humphrey’s Executor, nan rule that keeps nan president from firing FTC commissioners is, I’m guessing, retired nan door. Roe v. Wade was out nan door. Just these foundational precedents of American rule retired nan door.

A batch of that is based connected what blimpish judges would telephone originalism. I person a batch of feelings astir originalism, but a large inclination wrong of originalism is utilizing AI, aliases what they telephone “corpus linguistics,” to find what group meant successful nan past. Then, you return nan AI and you say, “Well, it did nan occupation for me. This is nan answer.” Are you worried that your devices will beryllium utilized for that benignant of effort? Because it really puts a batch of unit connected nan AI instrumentality to understand a batch of things.

I’m not that worried. I don’t deliberation nan Supreme Court is asking LexisNexis what we deliberation it should do.

But surely courts up and down nan concatenation are.

They’re asking ineligible questions, they’re getting answers back, and past they’re interpreting those answers. We are providing them pinch nan earthy contented that they request to make nan determinations, but we’re not practicing law. We’re not making those decisions for them.

I’m going to outpouring this connected you, but present it is: John Bush is simply a Trump-appointed judge. He cited nan emergence of corpus linguistics successful nan ineligible field, and he said, “To do originalism, I must undertake nan highly laborious and time-consuming process of sifting done each this. But what if AI were employed to do each nan reappraisal of nan hits and compile statistic connected connection gathering and usage? If nan AI could beryllium trusted, that would make nan occupation overmuch easier.”

That is him saying, “I tin outsource originalist reasoning to an AI.” This is simply a trend. I spot this peculiarly pinch nan originalist judges, that nan occupation they deliberation they’re meant to do is find what a connection meant successful nan past. And AI is awesome astatine being like, “Statistically, this is what that connection meant successful nan past, and we’re going to outsource immoderate ineligible reasoning.”

This is, I think, very odd. My thoughts astir originalism and stare decisis successful America successful 2025 aside, saying, “I will usage an AI to scope into nan past and find this meaning” seems very odd. I’m wondering really you consciousness astir your instrumentality being utilized successful that way.

I decidedly understand your constituent there. I deliberation astir nan affinity of a brick. You tin usage a ceramic to build a infirmary and return attraction of sick children, aliases you could return a ceramic and propulsion it done a window. One usage is really awesome and different is beautiful negative, but successful either case, it’s a brick. I deliberation astir our instrumentality arsenic being neither bully nor bad. I deliberation it could beryllium utilized for good. I deliberation it could beryllium utilized for immoderate type of activity that attorneys [need]. I wouldn’t want to opportunity originalism is simply a bad thing. I deliberation it could beryllium utilized for galore different things. I deliberation it could beryllium utilized for originalism. I deliberation it could beryllium adjuvant for those who want to return that way and find a caller measurement of looking astatine something.

We person each nan data. They tin hunt it, they tin usage nan instrumentality to find things it wasn’t imaginable to find successful nan past. So, I could spot them utilizing our instrumentality successful that way. I conjecture it’s up to nan attorneys to find really they’re going to usage nan product. We’re not building it because we’re trying to alteration nan law. We’re building it because we’re trying to thief attorneys do nan tasks that they want to do.

But I look astatine nan expanse of nan tech manufacture — not nan ineligible industry, but nan tech manufacture — complete nan past 15, 20 years, and boy, person I heard that reply many, galore times. The societal media companies each said, “Well, you tin usage it for bully aliases evil. We’re neutral platforms.” It turns retired possibly they should person thought of immoderate of those harms earlier.

Look astatine nan AI companies today. Who knows if training is copyrighted. We cognize nan answer. You can’t really conscionable opt retired of copyright law. Now, we’re going to do nan lawsuits and we’ll spot what happens. Who knows if OpenAI doing Sora, which is TikTok for deepfakes — actually, we know. We cognize nan reply is, you should person immoderate guardrails. 

So, I’m posing you nan aforesaid question. We spot a peculiarly originalist judiciary hell-bent connected utilizing originalism to alteration precedent astatine alarming rates. I would opportunity it’s alarming for maine because I paid for a rule grade that I now deliberation is useless, but that’s why it’s alarming to me. It’s alarming because a batch of group person had their authorities taken distant arsenic well. Every time this is happening. And 1 of nan ways they’re going to do that is to defer to an AI determination engine. They’re going to say, “We asked nan AI, ‘What did “all people” mean erstwhile nan 14th Amendment was drafted?’” and this will beryllium really we get to a birthright citizenship case. I’m conscionable connecting this to nan speech we had astatine nan beginning. We’re going to springiness our reasoning to a machine successful a measurement that it’s not needfully accountable for, and we’re going to spot nan computer. The methods of reasoning and that rigor mightiness spell away. 

So I’ve heard nan reply that nan instrumentality is neutral from tech companies for years, and I’ve seen nan outcomes. I’m asking you. You’re building a tech merchandise for lawyers, and they’re already utilizing it successful this circumstantial way. I’m wondering if you’ve thought astir nan guardrails.

We run nether responsible AI principles, and that includes a number of things. One, we ever effort to see nan real-world implications of immoderate merchandise we develop. We want to make judge that there’s transparency successful position of really our merchandise works. We unfastened up nan achromatic container truthful group tin spot nan logic that we’re using, and they tin really spell successful and alteration it if they want. So, we want to make judge that there’s transparency and there’s control. 

We ever incorporated quality oversight into merchandise development. Privacy and information is different 1 of our halfway tenets successful responsible AI creation. Another point we’ve incorporated is nan prevention of bias introduction. So, those are nan RELX principles for AI development, and we adhere to those. We want to create products that do bully things for nan world.

If you asked Lexis AI if nan 14th Amendment guarantees birthright citizenship to each group calved successful nan United States, will it make nan statement that it doesn’t?

I’ve ne'er asked it that question. I can’t show you.

Do you person your telephone connected you? There’s a mobile app.

I could popular up present and inquire it, I suppose. Let maine popular into Protégé here. “Does nan 14th Amendment guarantee birthright citizenship aliases are location exceptions?” Let’s see.

It’s generating a response, truthful we tin travel backmost to it successful a minute.

I’m very funny to spot what it says, because up until precocious there’s only been 1 reply to that question. Now, nan Trump management is saying, “Nope, actually, that’s not what ‘subject to nan jurisdiction thereof’ means.” In bid to triumph astatine nan Supreme Court, they will have to conception an originalist statement to that question, and I americium assured that nan measurement they’re going to do that is by feeding a bunch of information into an AI exemplary and saying, “This is what was really meant astatine nan clip of nan 14th Amendment’s drafting.” That’s a point that AI will beryllium utilized for that is very destructive.

I’m not an attorney, truthful I’m conscionable going to publication nan reply here:

“The 14th Amendment of nan United States Constitution guarantees birthright citizenship to each persons calved aliases naturalized successful nan United States, and taxable to its jurisdiction. The building ‘subject to its jurisdiction’ has been interpreted to see astir each individuals calved connected US ungraded pinch a fewer constrictive exceptions. These exceptions see overseas diplomats, children of overseas diplomats, children of force forces successful dispute occupation, children calved connected overseas nationalist ships, and, historically, children of members of Native American tribes who owed allegiance to their people alternatively than nan United States.”

It goes on.

You should nonstop that to [Chief Justice] John Roberts correct now. Can Protégé do that? Because that’s nan answer. 

The mobility is, are a bunch of blimpish influencers going to opportunity Protégé is woke now? This is nan taste warfare that you’re in.

It does admit that “recent cases person affirmed this mentation rejecting attempts to grow nan exceptions of birthright citizenship,” truthful it does besides admit that location person been efforts to construe it differently. The reply goes connected rather a bit.

The logic I inquire that mobility very specifically is because Reconstruction is up for grabs successful a very existent measurement successful this case. Do you deliberation you person a work arsenic nan instrumentality maker? That’s really nan mobility for truthful galore AI companies. You’re nan instrumentality maker. Do you person a work to not deepfake existent people? Do you person a work to not show group clone ideas? I deliberation you were very clear connected that, you person a work to not hallucinate, but present you person —

We don’t want to present aliases perpetuate immoderate bias that mightiness beryllium either. And to do that, we trust connected nan rule arsenic opposed to a consumer-grade exemplary that astir apt conscionable uses news articles, which mightiness person a very different mentation of things depending connected nan news articles. There are overmuch much apt to beryllium biases from introduced news articles than black-letter law, for example.

The logic I’m funny astir that is because there’s a spectrum. I don’t deliberation there’s immoderate spot for telling group what they tin do pinch Microsoft Word moving locally connected their laptop. Do what you’ve sewage to do. Telling group what they tin do pinch a consumer-grade AI instrumentality built into Facebook? I deliberation Facebook has a batch of work there, particularly because nan opportunity to administer that contented acold and wide is astatine their fingertips.

That’s a large opportunity spectrum, and present successful nan mediate there’s these AI companies. Do you person nan responsibility to say, “Well, if you want to spell make nan statement that birthright citizenship doesn’t protect everyone successful nan United States, you’ve sewage to do that connected your own. Our robot’s not going to thief you.” Do you consciousness immoderate of that pressure?

We effort not to get into authorities aliases immoderate of that debate.

I do not deliberation that’s politics.

We’re trying to create a strategy that does not person bias introduced into it, that will springiness you nan facts, and attorneys tin do nan activity that attorneys do to make those important decisions. Our occupation is to springiness them nan accusation that they need: nan precedents, nan facts, each nan accusation that they request to past create their argument, immoderate that mightiness be. But we really don’t get into immoderate of nan authorities of birthright citizenship being guaranteed aliases not.

Well, astatine immoderate constituent you do. This is — again, to bring america backmost wherever we started, I first encountered LexisNexis arsenic a database of cases and immoderate lawsuit notes. There were immoderate rule professors who were very proud that their lawsuit notes were successful LexisNexis erstwhile I was successful rule school. Now we’re drafting a small bit, going to spell do nan research. Now we person a agentic AI that’s making nan arguments. Maybe 1 time we will automate each nan measurement to filing. You’re taking connected much of nan burden. You are making nan arguments. The institution is making nan arguments. Where is nan line? Because location are tons of lawyers who wouldn’t return that case, who wouldn’t make that argument. Is location a statement for you?

I would opportunity our attack is to limb nan attorneys pinch nan champion imaginable information, and thief them pinch nan drafting of those documents. We’re really conscionable being led by our customers and what they’re asking america to do. We surely are not trying to construe nan law. We’re not trying to style nan ineligible system. We’re not lawyers. We’re not trying to do nan activity of lawyers. We’re trying to thief lawyers do nan activity they do successful a much businesslike measurement and, hopefully, thief them thrust amended outcomes. 

But it’s ever their prerogative to construe nan accusation that we provide, which is what lawyers do. That’s what they’re awesome at. The logic we person cases is because location are group connected some sides. The 2 individuals are going to make other arguments, we want to support some of those attorneys arsenic champion that we can.

I get it erstwhile you’re nan database of cases. I get it erstwhile you’re nan connection processor. I get it erstwhile you’re nan specialized connection processor aliases nan lawsuit guidance platform. The point that I’m pushing connected many times present is if nan AI strategy is really doing nan work, do you consciousness for illustration you person different guardrails?

I deliberation our work is to create AI successful a responsible way.

Give maine an illustration of thing you wouldn’t fto your AI do — an statement that you wouldn’t fto your AI make, aliases a mobility that you wouldn’t fto your AI draft.

I don’t cognize that we would want to needfully restrict nan AI successful that way. We’re referring backmost to nan accusation that we have, which is our charismatic postulation of documents and materials that helps lawyers understand what nan facts are, what nan precedent is, and what nan inheritance is, truthful they tin do nan real, heavy ineligible activity and make those trade-off decisions, judgement decisions, nan important things that, again, attorneys went to rule schoolhouse to do.

I deliberation these questions are going to travel up complete and complete again. We should person you backmost to reply them arsenic you study more. As you look retired complete nan sky — nan adjacent 2 aliases 3 years — what’s nan adjacent group of capabilities you spot for LexisNexis, and what do you deliberation nan pressures that mightiness alteration really you make immoderate of those decisions will be?

It’s difficult to opportunity precisely what nan main point that’s going to alteration nan way going guardant is going to look for illustration because if I look backmost 2 years ago, I would’ve ne'er guessed we’d beryllium doing what we’re doing coming because nan exertion didn’t beryllium aliases it was excessively costly to implement. That’s wholly changed complete nan past 2 years, and I deliberation complete nan adjacent 2 years, it’s going to alteration again. So, it’s really difficult to opportunity wherever we’re going to go. Our imagination remains nan same, which is that we want to thief attorneys. We want to supply them pinch a personalized, AI-powered merchandise that understands their believe area, their jurisdiction. It has entree to our charismatic group of materials and their anterior activity product. It understands their preferences, it understands their style, it understands what they’re trying to do, and it tin automate tasks that they do coming manually.

We will proceed to return that latest disposable technology, show it to our customers, and person them thief america understand really we tin usage that exertion to service them successful much modern and applicable ways. That’s really what’s going to guideline our roadmap successful nan future.

Sean, this was great. Let maine cognize erstwhile you create a strategy that tin really navigate an physics lawsuit filing website because immoderate of nan smartest group I cognize can’t do that. But this was great. We’ve sewage to person you backmost soon. Thank you truthful much.

Thank you truthful much. I really enjoyed our clip today. Take care.

Questions aliases comments astir this episode? Hit america up astatine decoder@theverge.com. We really do publication each email!