AI
LexisNexis Bets Its Library Will Outlast the Legal Models
LexisNexis is renting OpenAI, Anthropic, and Google, then gating them with Shepard’s and a 250-person bench, a bet that the library, not the model, is the product.
LexisNexis is wrapping models from OpenAI, Anthropic, and Google in an agent harness that checks their output against Shepard’s, because its chief technology officer says agents will always invent. The company set that wager out on October 2, 2026, in a feature on the Legal Intelligence Engine and agent harness behind Lexis+ with Protégé.
Lawyers are not asked to pick a model. They describe a matter, and the harness chooses sources, agents, and models, then carries context forward as the work moves into Word, Excel, or PowerPoint.
LexisNexis Is Betting the Library Outlasts the Model
The models in that stack are someone else’s. OpenAI, Anthropic, and Google supply general systems; LexisNexis adds purpose-built legal models and then refuses to let any of them speak unchecked. Greg Dickason, chief technology officer of LexisNexis Legal & Professional, has been blunt about why. In August 2026 he said an agent can always make something up, that no one will reach 100% non-hallucination, and that the only answer is a harness around the output.
That is a bet against legal AI products that sell the model as the product. LexisNexis, part of RELX, is selling the library, the citator, and the routing layer, and it is collecting on that bet even when a lawyer opens a rival assistant. In July 2026 the company said AI-related products accounted for 90% of new business. It employs 11,900 people and serves customers in more than 150 countries.
THE CONTENT BEHIND THE HARNESS
- The collection: More than 200 billion legal and news documents and public records, drawn from more than 50,000 sources.
- The feed: Near-real-time updates from more than 3,000 courts, with about 4 million new records each day.
- The check: The LexisNexis legal knowledge graph plus Shepard’s Citations, used to test whether a cited case is still good law.
- The human layer: A specialized team of 250 professionals scores AI output on accuracy, fluency, and authoritativeness and feeds those scores back into testing.
Foundation models get cheaper and more interchangeable every quarter. A citator that attorney-editors still read, and a corpus that updates from thousands of courts, does not. That is the asset LexisNexis is putting in front of the lawyer, and it is also the asset it has started licensing into other platforms.
How Lexis+ With Protégé Routes a Task
Lexis+ with Protégé reached general availability in the United States on February 24, 2026, replacing the Lexis+ AI name. On May 7 the company added Protégé Work, agentic drafting for contracts, motions, briefs, and deal documents, plus Workrooms and a Shepard’s Verify pass. The August 24 release is the one that changed the routing. Earlier agentic flows were scripted in advance. The new harness picks tools on the fly from a plain-language request and keeps matter context as research turns into a draft.
Dickason’s example is a single instruction: find the latest law on a point and draft a client document, done together rather than as two steps. Alexandra Smyth, executive vice president and general counsel of LexisNexis Legal, described the same change from the user side. A question that once meant opening Intelligize for SEC material, then Lex Machina for a judge’s history, now goes to the harness, which calls the source and drops the answer into Office.
WHAT THE AGENT HARNESS SELECTS
- The models: Purpose-built legal systems plus general models from OpenAI, Anthropic, and Google, chosen per task rather than locked to one vendor.
- The sources: LexisNexis primary law and news, Shepard’s, and a firm’s own policies, precedents, and prior work, with context walled off by matter so one client’s file cannot leak into another’s.
- The agents: Specialized skills for research, drafting, and citation checks, including a pass against Shepard’s rather than a trust-the-model step.
- The output: Review-ready material aimed at Word, PowerPoint, or Excel, with links back to the underlying authority so a lawyer can open the case, not just the summary.
Smyth’s test for the product is whether she can check what it gave her. Dickason’s test is harsher, and it is the line the architecture is built on.
It doesn’t matter how good your agent is, it can always make up something. You’re never going to get to 100% non-hallucination. So you’ve got to put the harness around it.
Greg Dickason, Chief Technology Officer, LexisNexis Legal & Professional, August 2026
That sentence is the product brief. The harness is not a flourish on top of GPT-class models. It is the admission that those models, including the legal ones LexisNexis trains itself, will keep inventing, and that the company intends to sell the invention-catcher.
Fake Citations Keep Showing Up in Court
The reason a citator is the product again is not subtle. In Mata v. Avianca, Inc., 678 F. Supp. 3d 443, Judge P. Kevin Castel of the Southern District of New York on June 22, 2023 sanctioned two lawyers and their firm $5,000 after a brief cited six decisions that did not exist. Steven A. Schwartz had used ChatGPT for the research. When he asked the chatbot whether the cases were real, it said they were, and that they could be found in databases such as LexisNexis and Westlaw. They could not.
Castel wrote that lawyers may use a reliable AI tool, and that existing rules still require a human gatekeeper on filings. The $5,000 figure became a teaching hypo. It did not empty the dockets. Damien Charlotin, who runs a public tracker of court decisions that address AI-invented material, lists 2,145 tracked hallucination cases, with new entries still landing in early October 2026, including fabricated case law in U.S. and foreign courts.
FROM THE FIRST FINE TO THE HARNESS
- June 22, 2023: Castel sanctions the Mata lawyers $5,000 for six ChatGPT-invented authorities that the chatbot had placed, falsely, on LexisNexis and Westlaw.
- June 18, 2025: LexisNexis and Harvey announce an alliance that puts Protégé answers, U.S. primary law, and Shepard’s inside Harvey.
- February 24, 2026: Lexis+ with Protégé goes generally available in the United States.
- May 7, 2026: Protégé Work, agentic drafting, Workrooms, and Shepard’s Verify land as the next layer.
- August 21, 2026: The Customer Innovation Lab opens in New York City.
- August 24, 2026: The context-carrying agent harness is added to the Legal Intelligence Engine.
- October 2, 2026: LexisNexis publishes the feature tying the engine, the lab, and the 250-person evaluation bench into one pitch.
Shepard’s is older than any of those dates. The Shepard’s Citations Service is editorial analysis by attorney-editors of how later courts treated a case, including positive “followed” treatment that a negative-only citator never shows. Signals on the report, the red stop among them, tell a researcher whether an authority is still good law. Protégé now runs that check as a service against documents the models draft, which is a different job from hoping the model declined to invent a caption.
Stanford’s 2024 Test Still Hangs Over the Pitch
LexisNexis has, at times, sold linked citations as hallucination-free. Stanford’s RegLab and Institute for Human-Centered AI put that class of claim on a pre-registered set of more than 200 open-ended legal queries. The paper, led by Varun Magesh and Daniel E. Ho and later published in the Journal of Empirical Legal Studies, found that retrieval-augmented legal tools beat raw GPT-4 and still invented at rates no brief can absorb.
On that 2024 snapshot, Lexis+ AI, the product name then, hallucinated on more than 17% of queries and answered 65% accurately, the best of the legal tools in the test. Westlaw’s AI-Assisted Research hallucinated more than 34% of the time and was accurate on 42%. General-purpose chatbots hallucinated on 58% to 82% of legal queries in the related Stanford work. Those figures describe the 2024 systems. They do not score the 2026 harness, the lab, or the 250-person bench, and treating them as a live Protégé report card is a category error that still circulates.
STANFORD’S 2024 LEGAL AI TEST
| Tool in the 2024 study | Queries with hallucinations | Queries answered accurately |
|---|---|---|
| Lexis+ AI | More than 17% | 65% |
| Westlaw AI-Assisted Research | More than 34% | 42% |
A 17% miss rate on research questions is not a rounding error when the miss is a case that does not exist. It is also not a verdict on Protégé’s current routing layer. The honest read is narrower. RAG cut the problem versus a naked chatbot and did not finish it, which is exactly the gap Dickason is building the harness to cover. Whether that cover dropped the rate is a study nobody has published on the August 2026 engine.
Harvey Pays for the Same Citator
The cleanest proof that LexisNexis thinks the library is the scarce piece is that it sold access to the rival most often held up as the AI-native alternative. On June 18, 2025, LexisNexis and Harvey announced a deal to put LexisNexis generative AI, U.S. case law and statutes, and Shepard’s inside the Harvey workspace, and to co-build motion-to-dismiss and summary-judgment workflows. Harvey users ask Protégé in natural language and get answers anchored in LexisNexis fine-tuned models, Shepard’s Knowledge Graph, and Point of Law Graph data. Ask LexisNexis later opened as a knowledge source inside Harvey, with Shepard’s Signals in the thread.
Our customers trust LexisNexis for authoritative legal content, and we’re excited that they will benefit from LexisNexis capabilities within the Harvey experience.
Winston Weinberg, Chief Executive Officer, Harvey, June 18, 2025
Sean Fitzpatrick, then CEO of LexisNexis North America, UK, and Ireland, called the alliance a way to put LexisNexis AI and content directly in front of Harvey customers. In May 2026, as CEO of LexisNexis Global Legal, he drew the quality line the same way: legal AI has to produce work a lawyer can verify, defend, and trust. If Harvey is the interface a Big Law associate actually lives in, LexisNexis still gets paid when that associate needs a case to survive Shepardizing.
That double collect is the part of the wager most product write-ups skip. LexisNexis competes with Harvey for the assistant seat, and it supplies the grounding Harvey cannot mint from model weights. Thomson Reuters is running the same play from the other library, with CoCounsel on Westlaw and KeyCite. The model vendors remain upstream suppliers. The fight that will still be here in a year is over who owns the file of what the law was yesterday.
Lawyers Now Sit in the New York Build Room
The Customer Innovation Lab opened in New York City on August 21, 2026, as a physical change in who is present when a prototype is written. Legal professionals sit with LexisNexis engineers and product teams, test working builds in the same session, and watch some of those builds move into Protégé. Dickason’s line for the lab is that customers help originate what gets built from day one, and that LexisNexis still has to validate it, ground it in citable authority, and decide what is ready to scale.
Work already credited to that room includes a customer-managed Protégé Vault for documents, video, images, and audio, with a two-hour deposition as the stock example, findings tied to timestamps, and Workrooms with permissioned sharing among counsel, clients, and experts. Around October 1, 2026, LexisNexis also began putting Lexis+ with Protégé in front of lawyers in the UAE as it marked 16 years in the MENA region, the same week it published the engine feature in New York.
A lab does not make a citator honest. It does change the failure mode. If a drafting skill is wrong, the first people to see it are supposed to be the lawyers who will be on the hook for the filing, not a product manager two releases later. That only holds if those sessions are a gate, not a showroom.
A 250-Person Bench Still Grades Every Model
The 250-person evaluation team is the least glamorous number in the October 2 feature, and it is the one that matches Dickason’s 100% line. Accuracy, fluency, and authoritativeness are scored by people, then pushed back into development. Shepard’s itself still depends on attorney-editors reading new opinions. The harness can call that layer. It cannot replace the person who will sign the brief.
Associates at well-staffed firms already live that split: the model drafts, a human cite-checks, a partner files. Software did not remove the second set of eyes. It moved the first draft earlier in the night. For a solo, there is no spare associate. A linked citation and a Shepard’s flag are the entire safety net, which is why LexisNexis keeps putting the citator in the screenshot.
The wager is now operational. Models are rented. Context is kept inside the matter. Output is run at Shepard’s. Customers sit in the lab. Two hundred fifty people still grade the machine. Charlotin’s list, at 2,145 decisions and still growing in October 2026, is the docket that bet is aimed at, and it is also the reminder that a harness is a check, not a promise that the next invented caption never reaches a judge.
-
AI5 months agoFable 5 Came Back Under a Commerce On-Off Switch
-
AI5 months agoGoogle’s SpaceX GPU Lease Has a Sept. 30 Deadline
-
CRYPTO5 months agoPlasma One’s XPL Locks Face a 1.81 Billion Cliff
-
APPS5 months agoDGO’s Rs 549 World Cup Pass Cost Fans Sleep and Data
-
AI5 months agoMoonshot AI’s $30 Billion Ask Became a $35 Billion Close
-
NEWS6 months agoColorOS 17 Device List Spans Oppo, OnePlus and Realme
-
GAMING5 months agoThe RTX 4050 Under Rs 70,000 Hides a Wattage Gap
-
AI5 months agoBoris Cherny’s Loops Still Demand a Higher Production Bar
