L

Lex Machina

Machine learning builds the asset, but what the buyer pays for is the resulting database. Lex Machina says proprietary technology and AI assisted attorney review turn raw legal documents into data sets and fill gaps in court records, and that its AI analyzes documents, reads signature blocks, updates case data and groups related entities. Without those models the analytics would not exist in usable form. But the product sold is a queryable dataset, valued for coverage and accuracy, with the generative layer arriving as Protégé on top rather than as the product itself.

Accuracy is claimed in superlative terms and measured nowhere. The API section describes the offering as the industry's most accurate litigation data, and the product is positioned on filling gaps in court records that others miss, but no figure supports either claim. Searched the product page in full and located no accuracy rate, no error rate for extraction or entity resolution, no benchmark, no test set and no published evaluation.

Nothing addresses what Protégé does when the underlying data does not support an answer to a prompt. What is published instead is provenance and scale: coverage figures carrying an explicit as-of date of April 2025, which tells a reader how current the data is without telling them how right it is.

A real review point is published for the data pipeline, and nothing for the generative layer. AI assisted attorney review is how raw documents become data, putting lawyers between model output and what customers see, and the RELX Responsible AI Framework names human oversight as a principle. Checked 31 Aug 2026, nothing says what the extraction models decide alone, what share of output attorneys review or against what threshold, what happens when an error is found in published analytics, or what oversight applies to Protégé when a user queries the data by prompt.

Adoption evidence has a cited basis, and the testimonials thin out on inspection. The vendor says over 90 percent of the largest firms use it, citing the 2025 Law360 Pulse Leaderboard Report and the 2025 AmLaw100 Report, with data dated April 2025. Of three testimonials, one is a named practitioner at a named firm, John Johnson, a partner at Fish & Richardson, with no figure. The second is from an unnamed Chief IP Litigation Counsel at a Fortune 500 company.

The third, from Miriam Rivera, former Deputy General Counsel at Google, says what she would do if she were at Google today, which is opinion rather than deployment evidence. No case study with figures was found.

A clear and unqualified training commitment, short of the detail around it. The product page states that LexisNexis never uses customer data to train AI models, which is stated flatly with no carve-out for internal models, external models or aggregates, and adds that identifiable information is removed from AI interactions so performance can be improved without compromising privacy. Robust data retention and deletion policies are asserted, privacy by design is claimed at every stage of systems, products and business processes, and advanced encryption is referenced.

Three gaps checked 31 Aug 2026: no retention period is stated, no segregation or access model is described, and privilege and work product are not addressed. The commitments also sit on a product marketing page rather than in any agreement, since the LexisNexis terms were not opened. Note the exposure here differs from a document platform, because what a customer puts in is largely queries and research rather than client files.

Nothing published on the advice line was found. The product page, read in full including the data security, privacy and governance section and the footer, has no statement that the analytics are not legal advice, no professional responsibility or ethics page, no bar or ethics guidance such as ABA Formal Opinion 512, and no jurisdiction limits. The product predicts outcomes and forecasts a motion's chances of success, exactly the output a lawyer must judge independently, and the marketing invites that use without addressing it. The risk is lower than for a consumer tool because buyers are lawyers, but nothing is published.

A named governance framework addresses bias explicitly, without a named owner or published results. LexisNexis follows the RELX Responsible AI Framework and lists its commitments: AI that protects privacy, is transparent and explainable, avoids unfair bias, includes human oversight and is designed for real world impact. No person or role is named as accountable for model behavior, no testing before release is described, no evaluation method is published, and no finding on uneven output across case types, courts, parties or populations is disclosed. On a product that profiles named judges and counsel, published bias testing would matter a great deal.

Principles are published where controls should be. The product page promises privacy by design, protection of customer data wherever it is stored, sent or accessed, advanced encryption, enterprise cloud providers, continuous monitoring, regular third party audits and robust retention and deletion policies, all assurances rather than specifics. Read in full, it gives no retention period, named subprocessor, incident or breach notification practice, encryption standard or access control model.

The one concrete statement is that identifiable information is removed from AI interactions. A LexisNexis Trust Center is openly linked at trust.lexisnexis.com; it was not opened and may say more.

Read: lexmachina.com/terms/, which leads to the LexisNexis product page rather than a Lex Machina agreement, and the LexisNexis General Terms and Conditions effective 5 June 2026, which say they govern any LexisNexis product linking to them; no separate Lex Machina agreement exists. Section 7.1 indemnifies against third party patent, trademark, service mark, copyright or trade secret claims, with five conditions and three remedies, the last being termination with a pro rata refund, and 7.2 makes this the sole remedy.

Section 6.3 caps liability at the lesser of actual direct damages or fees paid in the prior twelve months, in place of all other remedies. Section 6.4 excludes consequential damages with four named exceptions. Nothing covers output: section 6.1 disclaims liability for losses from errors or omissions in the services or materials, section 5.2 provides them as is, and section 4.3(a) says AI systems may not be accurate or error free and users must verify. No insurance position is named. For a product sold for case assessment, the errors clause is the one to read first.

Integrations are named and all of them are inside the parent's own portfolio. The page states that Lexis+ with Protégé, CourtLink, Law360 and CaseMap+ AI have integrations with Lex Machina and complement it, and Protégé itself is embedded in the product. An API is offered and described as a route for customers to build their own solutions using the underlying litigation data, linked to a community article. What is not published is any integration into the systems litigation work otherwise lives in: no document management, no matter management, no billing, no filing system and nothing outside the LexisNexis family.

Nor is there documentation an implementer could use, since neither the API article nor any developer index was opened, and nothing on the page describes what any integration moves or in which direction.

Nothing is published on where the software runs or where customer data sits. The product page, read in full including its data security, privacy and governance section, says only that LexisNexis partners with trusted enterprise cloud providers, naming no provider, region, jurisdiction or tenancy model. Nothing addresses residency options or processing location, or separates storage from processing. The LexisNexis Trust Center was not opened and may say more.

A trust center exists, and the product page names no standard. It says LexisNexis systems meet the highest standards for information security and that the company has regular third party audits to keep industry leading certifications, and links to trust.lexisnexis.com. No certification is named, so a reader cannot tell whether that means SOC 2, ISO 27001 or something else, and no auditor, period, scope or report route appears.

The trust center is openly linked and is a SafeBase style portal; it was not opened, so whether reports are released on registration or only through sales is unknown. Assurance language without a named standard is weaker than a badge without scope.

Nothing is published about the model supply chain a customer inherits. The product page, read in full, names Protégé as the AI assistant and calls the extraction technology proprietary, but identifies no model provider, model or version, architecture, processing location for the model layer or subprocessor list, and commits to no notice of changes. Naming a branded assistant and proprietary technology does not say what sits underneath.

No pricing information is published at any level, including the unit of charge. Searched the product page in full on 31 Aug 2026, including its on-page navigation covering features, audience, resources, testimonials and related products. There is no pricing page, no tier structure, no per-seat or per-practice-area unit, no volume banding and no statement of what implementation adds. Every call to action is a demo request, and the demo form is the only route offered.

Nothing published would let a prospective buyer form any view of cost before entering a sales process, which is notable on a product sold to segments as different in scale as AmLaw firms, solo practitioners, law schools and state government.

Coverage is stated in detail on both who it serves and what it covers. Eight audiences are named: large law firms, small law firms, corporate legal departments, insurers, courts and judges, state and local government, federal government, and law schools, including government and court use. Coverage is equally precise and states its limits: all 94 federal district courts, the 13 courts of appeal, the PTAB and specialty venues, with state coverage described as enhanced state courts rather than complete and a separate pool of 18 million state cases supporting party analytics only; appeals analytics limited to federal civil cases filed since 2012; motion metrics limited to more than 40 motion types; and the whole claim dated as of April 2025. Published litigation reports on patent, class action and administrative law show practice depth.