Bloomberg Law vs Lexis+ AI: how they compare in 2026

The machine learning is deep, long-standing and genuinely load-bearing for several tools, and the product it serves is a legal publishing business that stands without it. The vendor's own framing settles it: its approach page lists differentiated content as one of four pillars and says plainly that the AI does not generate answers on its own but looks through a large collection of vetted information. Strip the models out and 15.5 million court opinions, 200 million dockets, 5 million statutes and regulations, 8,200 practical guidance documents and a legal newsroom remain, which is the product Bloomberg has sold since 2009. What would not survive is a real set of tools rather than a veneer, and that is why this is not lower: BCITE, Docket Key, Points of Law and Litigation Analytics are built by machine learning and do not exist without it. The flat all-inclusive subscription reinforces the reading, since a buyer does not purchase the AI separately or meter it. Graded level with Docket Alarm and one band below UniCourt, where the machine-derived dataset is the whole deliverable rather than an access layer over one.

A generative layer over a research platform that predates it by decades and stands entirely without it. Remove the models and Lexis+ remains a working legal research service with Shepard's, the case law corpus, statutes and Practical Guidance intact; what is lost is conversational search, drafting and summarisation. The vendor states the AI capabilities were built internally with technology partnerships rather than bolted on through acquisition, and the RAG platform is described as proprietary, so the model work is real and owned. Graded on the same basis as Everlaw and Relativity, both B, where a mature platform hosts the model layer rather than depending on it. Distinguished from Reveal, Jhana.ai and Descrybe at A, where removing the models removes the product.

Three of the four things this axis asks for are published and the fourth, the measurement, is not. Grounding is documented with an architecture attached: the vendor names retrieval-augmented generation alongside a proprietary guardrail service, states that generated content is grounded in its own expert-written or primary source material rather than the internet at large, and says Bloomberg Law Answers provides citations and links to the supporting authorities used in producing an answer, with inline citations offering one-click verification. Citation status is checked, which almost nothing else in this pull can say: BCITE is the platform's own citator and reports whether a holding is still valid law. The system states when it has no support, abstaining where no content covers the topic or the query is outside legal scope. What is absent is any number. The page describes an extensive benchmarking process with former attorneys evaluating accuracy from development through beta and into the life of the product, and no result, test set, sample size or error rate from that process was located on 31 Aug 2026. The gap is sharper than a simple omission, because the same page commits the vendor to providing customers with clear documentation of the benchmarking and validation process for its AI, and no such documentation was found. To the vendor's credit it states directly that AI responses are never fully trustworthy and pose a hallucination risk, and advises customers to validate answers against primary sources.

The architecture would support B. What holds it at C is an absolute claim that independent preregistered measurement contradicts, never retracted, answered with data the vendor has not published. Marketed on hallucination free linked legal citations, quoted elsewhere as 100 percent hallucination free. Independently tested in Magesh, Surani, Dahl, Suzgun, Manning and Ho, Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, 22 J. Empirical Legal Stud. 216 (2025), from Stanford RegLab and Stanford HAI, the first preregistered empirical evaluation of commercial legal AI tools, 202 hand built queries across general research, jurisdiction and time specific, false premise and factual recall categories, expert scored, dataset published. Result for this product: hallucinates on more than 17 percent of queries, accurate on 65 percent, incomplete answers on 18 percent. Credit where due and it is substantial: Lexis+ AI was the highest performing system tested, against Westlaw AI-Assisted Research at 33 percent and GPT-4 at 43 percent, and the grounding disclosure is real rather than asserted, covering a proprietary RAG platform, Shepard's editor reviewed sources, inline linked citations, a citation verification tool and in product feedback. Against: LexisNexis disputed the figure and stated its internal data showed lower rates, and has not published it. A preregistered peer reviewed measurement answered with an unpublished internal number is not a rebuttal this index can credit. Caveats belonging to the record: the initial methodology drew objections from both vendors and was revised, access restrictions constrained the design, and the systems were tested in May 2024 and have been updated since.

The design posture is stated in a sentence most vendors would not write: the vendor says it wants its AI to make the job easier, not to do it for the user, and describes a deliberate building-block approach addressing one task at a time rather than producing the end work product. Real control surfaces sit behind that. Answers carry inline citations with one-click verification to the source, the abstention behaviour is documented with two named triggers, all generative tools are beta tested before release, and customers are advised to validate responses against primary sources and use their own judgement. What is not published, checked across the approach-to-AI page, the AI product page and both agreements on 31 Aug 2026: nothing describes what happens after an answer proves wrong, no correction or error-reporting route is documented, and no confidence threshold or configuration is exposed. The gap matters most on Deep Thinking, described by the vendor as automatically planning and executing multi-step research across its content before synthesising results, which is the least supervised thing on the platform and has no described checkpoint between the request and the finished answer.

Oversight is stated at governance level and instrumented at product level, which is a stronger pairing than most of this roster manages. The RELX Responsible AI Principles name human oversight as a governing commitment, and LexisNexis states that human experts review model output for legal accuracy and that continuous review and monitoring operate across products. In the product itself the oversight is practical: citations are linked inline so a user can open the authority behind any proposition, a user can submit a citation for verification and be told when it may be wrong, and in product feedback is collected and stated to feed accuracy improvement. Held at B because nothing is bounded or quantified. No autonomy threshold, no statement of what the system does without review, no description of what the human expert review actually covers or how often, and no escalation behaviour. The independent evaluation is also relevant here: incomplete answers on 18 percent of queries means the system does decline or under answer at a measurable rate, and the vendor documents no refusal behaviour at all.

Testimonials stand in for measurement. The pricing page carries three quotes, and one is better than most in this pull because it names an organisation and describes an outcome: a corporate counsel at Synovos, Inc. says the platform let the team pull state summaries of laws and regulations without calling outside counsel, reducing tens of calls on minor issues and so reducing annual legal fees. That is a real deployment account with a real effect, and it has no figure, no timeframe and no date attached. The other two are attributed only by role and segment, a practice area leader in small law and a partner in large law, and both speak to the flat-fee pricing rather than to outcomes. Searched the home page, the AI product page, the approach-to-AI page, the pricing page and the site navigation on 31 Aug 2026 and located no case study with a stated method, no customer count, no named law firm, and no quantified result of any kind. The solutions pages for law firms, in-house, government and law schools were not opened and are the most likely home for anything stronger.

Deployment breadth is documented by segment and the product is the only one on this index whose real world performance has been independently measured and published. The commercial preview from May 2023 is described as covering global law firms, corporate legal departments, small law firms and United States courts, with the vendor stating that preview feedback drove specific product refinements before general availability in October 2023. Independent evidence: the Stanford RegLab and HAI evaluation establishes 65 percent accuracy across 202 expert scored queries, which is an outcome measure produced by a third party rather than a vendor claim, and is the only such figure on this index. Held at B rather than A because no customer, firm or court is named anywhere in located material, no usage figure is published, no vendor outcome measure exists, and the segment level description of the preview cohort is not something a buyer can verify or contact.

The documents are detailed, readable before signing, and what they disclose about research activity is the concern rather than the reassurance. The privacy policy states that the information collected includes user content, usage information, browsing history and search history, along with reading habits and inferences drawn to create a profile reflecting a user's preferences and behaviour. It then describes an advertising regime over that material: creation of lookalike or matched audiences for targeted advertising on social platforms, sharing of a hashed email address with third-party advertising companies, and Sharing for cross-contextual behavioural advertising under California and Virginia law, with opt-outs offered. The policy applies to the marketing site and to the products together and does not say whether in-product search history is carved out of any of it, which is the ambiguity a litigator would need resolved, because the cases and statutes a lawyer searches disclose the matter. Alongside that, the subscription terms state that the site is not designed or intended as a repository of personally identifiable or health-related information and give no warranty as to how such information would be held, and reserve a right to audit and monitor use of the services. Real commitments do exist and are graded here: a subpoena notice undertaking, Standard Contractual Clauses and an intra-group data protection agreement, data centres stated to be dedicated to the vendor's own operations. What is missing is the whole confidentiality core: no statement either way on training with customer input, no retention period, no segregation model between users or matters, and no privilege or work product treatment anywhere.

Confidentiality is genuinely evidenced at platform level and privilege is not addressed at all. The annual SOC 2 Type 2 examination of Lexis and Lexis+ covers confidentiality and privacy as named Trust Services Principles, not merely security, which is the broadest attested scope located in this pull and is a real confidentiality position rather than a claim. Supporting statements cover encryption, data minimisation and pseudonymisation, and privacy embedded through product development. What is absent: any treatment of legal professional privilege or attorney work product, any statement about the confidentiality of documents uploaded to the product, and any matter level segregation model. The gap matters because document upload is a headline capability, so client material enters the system by design. Marketing language around state of the art encryption and industry leading data security is noted and not credited, being unfalsifiable on its face. Checked the product pages, the trust centre material, the UK trust explainer and the launch material on 29 Aug 2026.

A clear advice-line position, published twice and reinforced outside the contract. Both the website terms and the subscription terms of service carry a capitalised clause stating that none of the services constitutes actual legal advice, opinion or recommendation, that a competent professional should be sought where legal assistance is required, that the user assumes all responsibility for decisions, conclusions and opinions reached as a result of using the services, and that no attorney-client relationship is formed with the vendor or any content supplier. Who may use the product is defined rather than assumed: the terms require a user to be eighteen, a member of a professional, business or academic community whose primary purpose is professional research, working in a court, government agency, corporation or similar entity, or enrolled as a law or business student, and exclude bots outright. The competence and supervision dimension appears where a buyer will actually read it, on the public approach-to-AI page, which advises customers to validate responses against primary sources and use their best judgement and says the AI is meant to make the job easier rather than do it. Two things hold this short of the top: no bar or ethics guidance is named anywhere, including ABA Formal Opinion 512, and no jurisdictional limit on the product's answers is stated.

Not located. No statement that output is not legal advice, no positioning on the reviewing lawyer's role, and no engagement with professional conduct rules was found in the material read. The question is live for this product because the independent evaluation established a measurable hallucination rate on a tool marketed to the whole profession, which makes the verification duty concrete rather than theoretical. Checked the launch and pressroom material, the product pages, the UK trust explainer and the trust centre material on 29 Aug 2026. Research limitation recorded: LexisNexis publishes a very large body of practitioner and ethics content and this pass did not survey it, so this grade is a correction candidate rather than a settled absence.

A published framework with a described mechanism behind it, and no results and no bias work at all. The approach-to-AI page, last modified 18 June 2026, sets out four pillars and puts substance under the accuracy one: retrieval-augmented generation plus a proprietary guardrail service, generative output grounded in vetted first-party or primary material rather than the open internet, beta testing of every generative tool before release, and an accuracy evaluation carried out with former attorneys running from development through internal testing, beta and customer testing and onward through the life of the product, with continuous tuning against model drift. It commits to documenting where and when the products use generative AI, the sources the AI uses, and the benchmarking and validation process. That is a mechanism a buyer can ask about, not a set of principles. What is not there is the other half of this axis. No accountable owner inside the vendor is named. No testing result is published, so the benchmarking commitment is unredeemed by anything located on 31 Aug 2026. And nothing addresses bias or uneven output, across matter types, jurisdictions or populations, on a platform whose analytics profile named judges, attorneys and firms. The trust centre carries a document titled AI Frequently Asked Questions whose contents were not read and which is the most likely place for any of this.

A named parent level framework with stated commitments, published and attached to the product. The RELX Responsible AI Principles are named and enumerated as transparency, fairness, human oversight and respect for privacy, and LexisNexis states the product is built with them, that human experts review AI models for legal accuracy and ethical performance, and that continuous review and monitoring align products with accountability. That is materially more than the several records in this pull with no governance disclosure of any kind, and it is published on a trust surface rather than in a press release. Held at B because nothing behind the principles is published: no model card, no bias or fairness testing methodology or result, no evaluation output, no accuracy monitoring figures, no named governance body or review cadence, and no ISO 42001. A fairness principle with no measurement is a commitment rather than evidence, and the one substantive accuracy measurement in the public record was produced by researchers rather than by this framework.

Most of the ground is covered in public, with the retention period the notable hole. The privacy policy, updated 9 July 2026, states that personal information is kept on secure servers under administrative, technical, personnel and physical measures, and sets out how long it is held by naming the criteria rather than a number: the time needed for the collection purpose, whether the user has stopped using the products, the existence of relevant legal proceedings, and legal or regulatory requirements. Its intra-group statement is more specific than most, describing a personal data governance programme, a maintained data map inventory documenting flows within the vendor's systems and disclosures to third-party systems, ongoing monitoring and testing of safeguards, company-wide privacy training, and data centres stated to be dedicated solely to the vendor's own products and operations with monitored access controls. Standard Contractual Clauses and an intra-group data protection agreement cover cross-border transfer, and additional processing terms including SCCs are made available where the vendor acts as processor. The trust centre adds data erasure, backups, access monitoring, least privilege and logging as named control areas, and publishes a full responsible disclosure policy with a reporting address and stated expectations on both sides. Missing: no retention period for anything, no subprocessor named on any page read, and no breach notification commitment located.

Handling controls are stated and the model training question is unanswered. Published: encryption, data minimisation, pseudonymisation, privacy embedded at every stage of product development, secure cloud infrastructure, a robust set of information security policies, and incident response plans that are updated and tested periodically with technical, administrative, business and executive escalation paths and external firms on retainer. Incident response detail at that level is uncommon and is credited. What was not located: any statement on whether customer prompts, uploaded documents or research queries are used to train or improve any model, any retention position for that content, and any tenant separation description. For a product whose headline capabilities include document upload, the training and retention position is the disclosure a buyer needs most and it is the one not made. Checked the trust centre material, the UK trust explainer, the launch material and the product pages on 29 Aug 2026.

The allocation is published in full, is precise, and runs almost entirely one way. The subscription terms of service, updated November 2025, cap aggregate liability at 500 dollars regardless of the form of action, disclaim all express and implied warranties including merchantability, fitness for a particular purpose and non-infringement, provide the services as is and as available, and state that the vendor does not guarantee the accuracy, correctness, completeness or timeliness of the services. One exclusion is written for this product class specifically and is worth a buyer's attention: liability is excluded for any inability or failure to perform legal or other research or related work, or to perform it properly or completely, even where assisted by the vendor. The terms also place the burden squarely, saying the user is solely responsible for the accuracy and adequacy of the data used and the resulting output, and impose a one-year limitation period on any action. Governing law is Virginia with exclusive jurisdiction in Arlington County. The only carve-out from the cap is death or personal injury from negligence where the law forbids limiting it. Nothing runs toward the customer: no indemnity, no warranty on output, no AI-specific liability term. The trust centre lists a certificate of liability document, which was not read and is the one route to an insurance position.

No published position located. Nothing was found on liability for AI output, warranty, indemnity, service levels, or remedy where a generated answer or citation is wrong. The absence is more consequential here than on most records because the vendor marketed an accuracy absolute that independent measurement contradicts: a buyer who relied on hallucination free as a representation has no published recourse framework to look to. Checked the product pages, the trust centre material, the UK trust explainer and the launch and pressroom material on 29 Aug 2026. Research limitation: subscription terms for this product are behind a customer agreement rather than published, and no public terms page was located in this pass, so this is a correction candidate.

Integrations are named in passing and nowhere documented. Two are identifiable: Bloomberg Law Contract Solutions is described in the vendor's own newsroom as integrating with Microsoft Word alongside a centralised contract repository, and the subscription terms carry a special provision for users who reach Bloomberg Law through the Bloomberg Professional service, meaning the Terminal, whose own agreement then governs. Beyond that, nothing. Checked the full primary navigation covering products, solutions, insights, events and about, the AI product page, the approach-to-AI page and the site footer on 31 Aug 2026: there is no integrations page, no developer or API documentation, and no named connection to a document management system, matter management, e-billing or court filing. Dashboard Legal is offered as a project management and collaboration product rather than as a bridge into systems a firm already runs. This is the least well evidenced row on this record and the limit is worth naming: the Contract Solutions and workflow tools product pages were not opened, and either could carry integration detail that would move this grade.

Nothing located in the material read, and this grade carries the weakest research basis on the record. No integration, connector, API or document management system relationship was found across the product pages, launch material, trust centre material and UK trust explainer checked on 29 Aug 2026. Stated plainly rather than dressed up: this pass researched the product's capability claims and its trust estate and did not survey integration documentation, and a vendor of this scale very likely publishes some. Recorded as a documented absence across the surfaces actually checked, on the specific date, and flagged as the strongest correction candidate on this record. The honest reading is that this is a gap in the research rather than an established gap in the vendor's disclosure.

One infrastructure statement, no product deployment position. The privacy policy's intra-group data protection statement says the vendor's data centres are dedicated solely to its own products, services and operations and have secure and monitored access controls, which is more than the cloud-by-implication most of this lane offers and is the reason this is not lower. Cross-border movement is acknowledged rather than hidden: the policy states that uses and disclosures may involve transferring and processing personal information in various countries with differing levels of protection, including the United States, and that Standard Contractual Clauses and an intra-group agreement identifying exporters and importers are used as safeguards. What is not published anywhere located on 31 Aug 2026: no region or residency option for a customer to select, no tenancy model for the platform, no distinction between where processing happens and where data is stored, and no statement of where the generative features run. The trust centre lists a data centre item whose contents were not read. The statements that do exist sit in an annex about personal data protection rather than in any product or security material a buyer would be pointed to.

Nothing specific located. Published material refers to secure cloud infrastructure hosted on enterprise grade platforms, which names no provider, no region and no residency commitment, and to state of the art encryption, which is not a deployment disclosure. No single tenant or dedicated instance option is described and nothing states where customer queries or uploaded documents are processed and stored. Residency is a live question for a product sold across the US, UK, Canada, Ireland and Australia under different data protection regimes, and the Canadian trust material raises PIPEDA and Quebec Law 25 as customer obligations without stating the vendor's own residency position. Checked the trust centre material, the Canadian and UK trust pages, the launch material and the product pages on 29 Aug 2026. Correction candidate: the SafeBase trust centre may hold residency detail behind its registration gate, which was not entered.

A trust centre that is easy to find and thin on the one detail that makes an attestation an attestation. Every Bloomberg Law page footer links to it directly, under the label Trust and Security, with a product identifier in the URL, which is a materially better position than the unlinked parent portals seen elsewhere in this lane. The portal is a SafeBase instance reachable without a sales conversation and its document set is the richest located in this pull: a penetration test report, a security whitepaper, SOC 1 and SOC 2, a SIG Lite self-assessment, a certificate of liability, a business continuity management system document and an AI frequently asked questions document, alongside named control areas across product security, access control, application security, network security, business continuity and training, and a responsible disclosure policy published in full. What is missing is the substance of the attestation. SOC 1 and SOC 2 are named without a Type, no auditor is identified, no coverage period or report date is given, and no scope statement was located. Documents are split into public and private with an access request whose tier the portal does not state. One retrieval note: the overview section returned on fetch on 31 Aug 2026 described the group's tax and accounting software rather than Bloomberg Law, most likely because the product selector resolves in the browser, so it is recorded as a limit rather than graded.

The strongest security disclosure located in this pull, and the first A on this axis. A live trust centre operates at trust.lexisnexis.com on SafeBase, with a self serve registration route through which authorised users access compliance reporting, which under the three tier test is a request flow rather than a sales gate. The attestation is named with its scope stated in full: an independent third party auditor performs an annual SOC 2 Type 2 examination of Lexis and Lexis+ against all five Trust Services Principles, security, availability, processing integrity, confidentiality and privacy. Naming all five is materially broader than the norm, since most vendors scope to security alone or to three criteria, and the difference is exactly what a reviewer needs to know. A SOC 1 Type 2 report for digital content management and information technology services is separately published and its update announced. Annual cadence addresses currency. Held short of a perfect record by two things stated here rather than hidden: the auditing firm is not named, and the Lexis+ AI specific examination is described on one vendor page as scheduled for Q1 2024 with no subsequent confirmation located, so the AI assets' current attestation status is unclear even though the platform's is not. Calibration ladder for later records: Regology and Onspring sit at B with two of four elements each, being auditor and scope respectively without a portal; this record carries scope, currency and a self serve portal, which is strictly more. Note also located and deliberately not characterised: the trust centre carries a customer notification concerning a security matter, and the notice text was not read in full in this pass, so nothing is asserted here about its scope, date or which LexisNexis entity it concerns. It must be read before any adverse statement is made, and conflating LexisNexis Legal and Professional with LexisNexis Risk Solutions would be exactly the domain hazard this index has hit three times already.

The architecture is described and the providers are not, which is the shape this band exists for. What is published: retrieval-augmented generation named as the framework, a proprietary guardrail service named as the layer over it, an explicit statement that generation is grounded in the vendor's own expert-written and primary source content rather than the internet at large, and a statement that its large language models train on that specialised content. The vendor also describes the wider technology base, naming machine learning, natural language processing, information retrieval, recommendation systems and large language models, and says Bloomberg Industry Group employs hundreds of AI researchers and engineers who publish research annually. What a buyer cannot learn from anything located on 31 Aug 2026: whether any third-party foundation model sits underneath and whose, since the phrase used throughout is our large language models without stating they are first-party; no model name or version; no processing location for inference; no subprocessor named; and no commitment to notify customers when the model or its training data changes. The trust centre carries a subprocessors item and an AI FAQ, neither of which was read, and either could name what is missing here.

The existence of external model relationships is disclosed and no party is named. The vendor states that the AI capabilities were built internally at LexisNexis with world leading technology partnerships, which confirms that third parties are involved in the model layer while identifying none of them, and describes the RAG platform as proprietary. No foundation model provider, model family or version is named, no subprocessor list was located, and nothing states which models process customer uploaded documents as opposed to the public corpus. For a RELX company selling to regulated buyers whose own supervisors ask subprocessor questions, the naming gap is notable. Compare Onspring at B, the only record on this index that names its model provider outright. Checked the launch and pressroom material, the product pages, the UK trust explainer and the trust centre material on 29 Aug 2026.

The shape of the deal is published and the number is not. The charging model is stated clearly and repeatedly, and it is real information in a market where competitors meter: one platform at one price, flat and all-inclusive, with unlimited access rather than per-search or per-document charges. The pricing page makes the consequence explicit through a customer quote, that flat all-inclusive pricing means occasional use of a tool carries no additional expense, and the request form splits the purchase into two named lines, Bloomberg Law Research and Bloomberg Law Dashboard Legal. That is where disclosure stops. Searched the pricing page, the getting-started entry, the AI product page and the full navigation on 31 Aug 2026 and located no figure at any level, no unit of charge, no tier names or feature splits, no term length and no statement of what implementation adds. Every route to a number is a form: the primary navigation item is Request Pricing and the page it opens is a lead capture form. The firm-size bands in that form, from solo through fifty or more practitioners, imply price varies with headcount, but they are form fields rather than a published band structure.

No pricing located at any level. No price, no range, no tier structure, no unit of charge, and no statement of whether Lexis+ AI is licensed separately or bundled into a Lexis+ subscription, which is the first question a buyer with an existing contract would ask. Every route is a contact or demo request. Independent commentary describes tools in this segment as marketed at substantial monthly cost to professionals without publishing a figure, which is characterisation rather than disclosure and is not credited. Checked the product pages, the launch and pressroom material and the trust centre material on 29 Aug 2026. Standard practice for enterprise legal research incumbents and still an absence a buyer cannot work around.

Segments are documented and the boundary is not. Four buyer segments each have a dedicated solution page, covering law firms, in-house counsel, government and law schools, and the government page states unlimited access to state and federal coverage. The organisation types the vendor sells to are enumerated in its own forms as law firm, tax and accounting firm, corporation, nonprofit or association, consulting firm, education institution and government, with firm size bands from solo to fifty or more practitioners. Content coverage is quantified with unusual precision for this lane: more than 15.5 million court opinions growing by around 50,000 a month, roughly 200 million dockets with more than 21 million associated pleadings, more than 5 million codified statutes and regulations, 75 million EDGAR filings and more than 8,200 practical guidance documents, with Docket Key classifying over 200 motion and brief types across all federal district courts and Chart Builders supporting state-by-state rule comparison. What is absent is any statement of where it stops. No practice area list with supported and unsupported areas, no jurisdiction-by-jurisdiction coverage table, no statement of state court depth, and no acknowledgement of a gap anywhere. Checked the home page, the four solution page descriptions in the navigation, the AI product page and the approach-to-AI page on 31 Aug 2026.

Corpus depth is the incumbent advantage and it is described in category terms rather than enumerated. Grounding content is named by type: case law reviewed by Shepard's editors, statutes, and Practical Guidance secondary material, described collectively as the largest repository of legal content, with the product available across the United States and further LexisNexis markets including the UK, Canada, Ireland and Australia under regional trust and product material. Shepard's coverage is itself a coverage claim of substance, since a citator only works where the corpus is complete. Held at B rather than A because nothing is enumerated at the level a researcher checks: no jurisdiction or court list, no historical date range, no update lag or refresh frequency for any content type, and no statement of which jurisdictions the AI features are actually available in as opposed to where the underlying platform is sold. The largest repository is a comparative claim rather than a measurable one.

No located term or policy states whether customer input is used to train or refine any model, either way. The privacy policy lists improving, upgrading or enhancing the products and developing new ones among its purposes for collected information, which covers the ground without naming training. The approach-to-AI page says the vendor's large language models train on its own specialised content, meaning its court opinions, guidance and news, and says nothing about customer material. The asymmetry is explicit in the agreement: the subscription terms prohibit a customer from using product content in prompts or for training, tuning or incorporating into any language model without the vendor's written consent, and define language model broadly enough to catch any generative system. The prohibition runs one way. Read 31 Aug 2026.

Silent. The quoted phrase is the full scope of the annual SOC 2 Type 2 examination of Lexis and Lexis+, and it is the strongest confidentiality and privacy evidence on this record, but an attestation covers controls rather than commitments: it says the company is audited against privacy criteria, not that customer content is excluded from model training. No statement in either direction was located on whether prompts, research queries or uploaded documents are used to train or improve models. Document upload is a headline capability, so client material enters the system by design and the question is not theoretical. Checked the product pages, the launch and pressroom material, the UK trust explainer and the trust centre material on 29 Aug 2026. Recorded as silent, not as a negative commitment. Correction candidate: the SafeBase trust centre registration gate was not entered and may hold an AI data use statement.

Retention is addressed with criteria instead of a period. The privacy policy states that personal information is kept for as long as necessary for the described purposes and legitimate business reasons, and names four factors used to set the window: the time needed to achieve the collection purpose, whether the user has stopped using the products, the existence of relevant legal proceedings, and legal or regulatory requirements. Search history and user content are expressly within the categories collected, so questions put to the AI features fall inside this framework, but no period attaches to them and no customer-configurable or zero-retention setting is offered. The subscription terms point the other way on uploads, stating the vendor has no obligation to retain or store user generated content. Checked the privacy policy, both agreements and the approach-to-AI page on 31 Aug 2026.

Not addressed. No retention period is published for prompts, generated answers, conversational search history or uploaded documents, and nothing indicates whether retention is configurable or can be set to zero. Data minimisation is named as a principle, which speaks to collection rather than to duration. Checked the product pages, the UK trust explainer, the trust centre material and the launch material on 29 Aug 2026. Same registration gate limitation as above applies.

Nothing addresses walls or matter-level separation, and one published provision runs against the idea. The privacy policy states that personal information may be disclosed to the party who authorised the user's access, naming the employer or educational institution, which means a subscribing firm is a potential recipient of information about what its own users did rather than a party walled off from it. The subscription terms treat each registration as single-user and require credentials to be kept confidential, which is account hygiene rather than a permission model. Searched the privacy policy, both agreements, the trust centre landing page, the AI product page and the approach-to-AI page on 31 Aug 2026 and located no statement on separation between customer organisations, no access model within a subscribing firm, and no matter-level concept anywhere in the product.

Not addressed. No permission model, matter level restriction or tenant segregation description was located, and no document management system integration was found that would let retrieval inherit a firm's own permissions at query time. The product accepts uploaded documents and is sold into large firms where walls are a routine requirement, so the absence is material rather than incidental. Checked the product pages, the UK trust explainer, the trust centre material and the launch material on 29 Aug 2026.

A real notice commitment, published and qualified. The privacy policy addresses disclosure to unaffiliated third parties in response to a subpoena, court order or other legal process, and states that in such an event the vendor will use reasonable efforts to give the user reasonably prompt written notice so the user has an opportunity to comment or object and to preserve the confidential nature of the information, and to cooperate at the user's expense with the user's efforts to do so. That is the substance this signal asks for. Three qualifications a buyer should carry: the undertaking is reasonable efforts rather than an unconditional commitment, cooperation is at the user's cost, and the subscription terms contain a separate provision reserving the right to use and disclose information obtained from or input by the user as part of any legal process or as required by law, with no notice attached, which nothing reconciles. No transparency report was located on 31 Aug 2026.

Not addressed. No government or law enforcement request clause, no commitment to notify a customer before producing their data, and no transparency report were located in the material read. Checked the trust centre material, the UK and Canadian trust pages, the product pages and the launch material on 29 Aug 2026. Correction candidate on the same registration gate footing as the rows above.

The corpus is identified by type and quantified in detail, with the rights basis unstated. Published figures cover more than 15.5 million court opinions growing by around 50,000 a month, roughly 200 million dockets with more than 21 million associated pleadings, more than 5 million codified statutes and regulations, 75 million EDGAR filings and more than 8,200 practitioner-written guidance documents, so both the primary law and the vendor's own editorial layer are described and a cadence is given for the opinions. No upstream supplier is named. The privacy policy refers only to third party data suppliers as a category, and discloses something a buyer should note in passing: the databases available to registered users contain public records, publicly available information and non-public information from those suppliers. The subscription terms acknowledge third-party content providers whose own rules attach to their material without identifying any. No licence, public domain basis or rights statement for any part of the corpus was located on 31 Aug 2026.

Sources named by type with no licence basis stated. The grounding corpus is described as case law reviewed by Shepard's editors, statutes and Practical Guidance, which names the content classes and, unusually, names the editorial process applied to the case law rather than presenting it as raw text. Editorial review of the underlying authority is a provenance statement of real substance and few records in this pull can make it. Not stated: the licence or public domain basis for any content class, a jurisdiction or court enumeration, a historical date range, or an update lag for either the corpus or the citator. The vendor's description of the largest repository of accurate and exclusive legal content implies proprietary licensing without setting out its terms, and exclusive is a commercial characterisation rather than a provenance disclosure.

The platform operates its own citator and says how it is built. BCITE is described as Bloomberg Law's court opinion citator, using machine learning and natural language processing to let a user determine quickly whether the holding of an opinion is still valid law. That is a first-party treatment signal with a stated method rather than a licensed commercial citator or a prompt telling the reader to go and check. The description is where the disclosure ends: no coverage statement for which courts or date ranges BCITE treats, no accuracy or error rate, no explanation of how a treatment determination is reached beyond naming the techniques, and nothing published on whether the generative features surface BCITE status alongside authority they cite. Read 31 Aug 2026.

Licensed citator, and the strongest value recorded on this signal in the pull. Shepard's Citations is a long established named commercial citator owned by this vendor, and the product is stated to check all citations in a generated response against Shepard's for validation, so treatment checking is applied automatically to AI output rather than left as a separate manual step the user must remember. A user can additionally submit a specific citation for verification and be told when it may be wrong. This is the value the axis was written for: a named citator with an established treatment methodology, integrated into the generation path. Worth pairing with the Citation Accuracy note: automatic Shepard's validation did not prevent an independently measured hallucination rate above 17 percent, which tells a reader that citator validation addresses whether an authority exists and how it has been treated, and not whether the proposition drawn from it is correct.

An abstention path is documented with its triggers named. The vendor states that the Bloomberg Law Answers and AI Assistant features will abstain from answering when there is no available content covering the topic in question, or when the query falls outside legal scope. It sets that alongside an unusually direct acknowledgement that AI-generated responses are never fully trustworthy and carry a hallucination risk, and advice that customers validate responses against primary sources. What is not published is any demonstration: no worked example of a refusal, no evaluation or measurement of how often the behaviour fires or fails, and no confidence or grounding score surfaced to the user. Documented rather than demonstrable on that basis, checked 31 Aug 2026.

Not addressed by the vendor, and measured by someone else. No published statement describes an explicit no answer path, an abstention behaviour, a confidence score or what the system does when the corpus does not support an answer. The independent Stanford RegLab and HAI evaluation found the product returned incomplete answers, meaning refusals or ungrounded responses, on 18 percent of 202 queries, which establishes that some abstention behaviour exists in the system without the vendor describing it. Recorded as not addressed because the signal asks what the vendor discloses, and an externally observed rate is not a disclosure. The gap is significant on a product marketed as hallucination free: the honest counterpart to that claim would be a documented account of what happens when the system does not know, and it is not published.

No court order, opinion or disciplinary record naming this product has been located as of 31 Aug 2026. Instrument searched: the AI Hallucination Cases database maintained by Damien Charlotin, which tracks decisions worldwide where a court addressed hallucinated AI content and records the tool implicated where known, searched on the product name and the publisher's name, alongside 2026 sanctions trackers, law library commentary and trade press. This is a statement about the public record on the date shown rather than a clearance, and it is bounded by what that database covers. Two appearances should be distinguished from a hit, since the name recurs constantly in this literature: practitioner guidance repeatedly names Bloomberg Law as one of the databases against which a citation should be verified, and the vendor's own newsroom covers the sanctions phenomenon as a reporting subject.

None located, with the instrument named and a distinction preserved that matters on this record. General web searches combining the vendor and product names with court, order, opinion, sanction, disciplinary and fabricated citation terms returned nothing on 29 Aug 2026. No named docket database or court record tracker was searched, so the instrument is weaker than this product's prominence warrants. The distinction: the Stanford RegLab and HAI study is an academic evaluation measuring a hallucination rate, not a court record, and it is recorded on the Citation Accuracy axis where it belongs. It is not a product named adverse value here, and treating a peer reviewed study as though it were a judicial finding would collapse exactly the distinction this signal exists to preserve. Recorded as a statement about what this search found, not as a clearance.

Searched the approach-to-AI page, the AI product page, the about page, the pricing page, both agreements, the privacy policy and the trust centre landing page on 31 Aug 2026. No bar or ethics authority is engaged with anywhere in the vendor's product or policy material, including ABA Formal Opinion 512 and any state bar guidance, and nothing maps the product to a professional conduct obligation a firm could hand to its risk committee. The advice-line disclaimers in both agreements and the validation guidance on the approach-to-AI page are the vendor stating its own position rather than addressing the rules its buyers are bound by, and neither cites the guidance those rules come from. One distinction worth drawing, because it cuts the other way superficially: Bloomberg Law's newsroom covers legal ethics and AI sanctions extensively as an editorial subject, which is journalism about the guidance rather than the product being measured against it.

Not addressed in the material read. No named ethics opinion, no ABA Formal Opinion 512 and no state bar guidance engagement was located across the product pages, launch and pressroom material, the UK trust explainer and the trust centre material checked on 29 Aug 2026. Research limitation recorded rather than glossed: LexisNexis publishes a very large practitioner content estate including practice guidance and professional responsibility material, and this pass surveyed the product and its trust surfaces rather than that estate, so this value is a correction candidate. What can be said on the material checked is that no bar guidance engagement appears on the surfaces where a buyer evaluating the product would look.

Efficiency claims without a word on the client's side of the bill. The vendor's material is built around time and cost saved: AI applications that help customers maximise efficiency and complete research faster, summarisation to speed consumption of news and research, and a flat all-inclusive subscription pitched on the basis that occasional use of a tool costs nothing extra. A customer quote on the pricing page goes further, describing reduced annual legal fees from fewer calls to outside counsel. Searched the pricing page, the AI product page, the approach-to-AI page and both agreements on 31 Aug 2026 and located no per matter record of AI-assisted work intended for fee purposes, and no guidance on billing, fee or client disclosure treatment where an AI answer informs advice. The flat pricing model does remove the disbursement question that metered research platforms create, which is adjacent to this signal without meeting it.

Not addressed. Nothing published addresses billing for AI assisted time, and no exportable record was located that a firm could use to show a client what portion of work was machine generated. The product's own commercial terms are not published either, so neither side of the fee question has a public answer. Checked the product pages, the launch and pressroom material and the trust centre material on 29 Aug 2026.

The artifacts a firm would need are catalogued on the trust centre and their contents could not be established from outside. The portal, linked from every product page footer and reachable without a sales conversation, lists a subprocessors entry, a data processing agreement, a data subject requests entry, a security whitepaper, a SIG Lite self-assessment and a document titled AI Frequently Asked Questions. None of those rendered its contents on fetch on 31 Aug 2026, no subprocessor or model provider is named anywhere on the portal landing page or on any Bloomberg Law page, and the document library is split into public and private with a get-access request whose tier the portal does not state. Recorded at the lower tier on that basis: the material demonstrably exists and is catalogued, and whether a firm can obtain it without an executed agreement is not stated. That is a better position than nothing published and short of a list a firm could forward to its client today.

On request, through a real and functioning route, which is better than most of this roster and short of a forwardable pack. The Trust Center at trust.lexisnexis.com runs on SafeBase and invites users to register to access compliance reporting including the SOC 1 Type 2 report, and the annual SOC 2 Type 2 examination of Lexis and Lexis+ is described with its full five principle scope. A firm responding to an outside counsel guideline questionnaire has a defined place to go and named attestations to point to, without needing a sales conversation. Held at on request rather than higher because nothing is open: no downloadable summary, no published subprocessor list, no named model provider and no DPA were located outside the gate, and the registration gate itself was not entered in this pass so the contents of the portal are unverified.

Sources are traceable per answer and the AI provenance a standing order asks for is not recorded. The vendor commits that when generative AI provides an answer or suggests text it includes links to source material or an explanation of why it returned that result, that Bloomberg Law Answers provides citations and links to the supporting authorities used in producing an answer, and that inline citations offer one-click verification to source, so a lawyer can produce the authorities behind an answer and check each one. It also commits to documenting where and when its products employ generative AI, which is the disclosure question at the platform level. What cannot be established: no model is identified or versioned anywhere, so which system produced a given passage is unknowable; nothing records who verified an answer; and no export designed for an AI-use disclosure or certification was located on 31 Aug 2026.

Partial record, strong on sources and silent on the rest. Citations are linked inline in every response and validated against Shepard's, and a user can submit a citation for verification, so the authorities relied on are identifiable and their treatment status is checkable, which is the sources retrieved limb answered better than almost anywhere on this index. The other two limbs a judicial standing order asks for are absent: nothing records which model produced a given output, and no human verification record is captured or exportable. No export artifact of any kind was located. The independently measured hallucination rate makes the verification limb the consequential one, since a court asking whether a human checked the output would find the product captures no evidence either way.