Sources & Notes
Where Our Claims Come From
Factual and quantitative claims on this site — including Insights articles — are linked to entries below using inline Source links. Each entry includes the claim or authority, the source, and any relevant caveats. This registry covers two types of sources: survey and research data used for statistics and operational observations, and primary authority (ABA formal opinions) cited in Insights articles on professional responsibility, AI, and confidentiality topics. External URLs are provided for independent verification.
Clio 2024 Legal Trends Report
Clio's annual legal trends report combines survey data from attorneys and legal professionals with original client-experience research — including a secret-shopper study in which researchers contacted law firms as prospective clients and tracked how firms responded. The phone and email findings below are from this secret-shopper methodology. The leads and revenue findings are from the broader survey research.
clio.com/resources/legal-trends →48% of firms were effectively unreachable by phone — not answering calls from prospective clients.
From the 2024 secret-shopper research. The phone and email figures are separate data points; do not combine them into a single "calls and emails" statistic.
Just 33% of firms responded to inquiry emails from prospective clients.
Separate finding from the same secret-shopper study as above. The phone unreachability and email non-response rates are distinct measurements.
73% of people seeking legal help said their first interactions with firms left them unlikely to recommend those firms to others.
From the client experience research in the same report. This is a reputation and referral finding — distinct from the response-rate data above.
Only 30% of prospective clients could easily understand the process for engaging a firm from its website. Only 14% could find pricing information.
From the website usability component of the same 2024 client research. Supports messaging, intake UX, and website clarity work.
Only 36% of firms could clearly explain the process and next steps to a prospective client on the phone.
From the communication quality research in the 2024 secret-shopper study.
Only 18% of firms addressed next steps or expected costs by email.
Only 12% of firms could provide an estimate of total cost by phone.
Firms using client-facing intake capabilities saw 51% more leads and 52% higher revenues, in Clio's 2024 research.
This is an observed association in Clio's research among firms that had deployed client-facing intake features. It is not a guaranteed outcome or a direct measure of any Songbird engagement. It is presented as such on the site.
Clio 2022 Legal Trends Report
A prior edition of Clio's annual legal trends research, covering attorney satisfaction, client outcomes, and practice management technology adoption.
clio.com/resources/legal-trends →Attorneys at firms using cloud-based legal practice management were 43% more likely to have satisfied clients, 34% more likely to be happy at their firm, and 11% more likely to have strong revenue streams.
Observed association from 2022 survey research. The relationship likely reflects the quality of systematic technology adoption rather than any single platform's effect in isolation.
ABA Legal Technology Survey Report
The American Bar Association's annual survey of attorneys and law firms on technology adoption, usage patterns, and professional responsibility related to technology.
americanbar.org/techreport →35% of surveyed attorneys reported capturing 1–10 additional leads per month after integrating online legal intake forms; 13% reported 11–20 additional leads.
The ABA article cited this figure as drawn from MyCase research. Used on the homepage as secondary corroboration alongside Clio intake outcome data — not presented as a primary proof point.
71% of attorneys said they have a responsibility to understand the benefits and risks of technology as part of professional competence.
30.2% of attorneys said their offices were currently using AI-based tools.
Thomson Reuters Institute — 2024 Legal Market Research
Annual legal market research from the Thomson Reuters Institute covering AI adoption, law firm strategy, and professional outlook.
thomsonreuters.com/reports →Note: Thomson Reuters publishes multiple annual research reports. Verify the current landing page at thomsonreuters.com for the most current edition.
79% of law firm respondents said AI will have a high or transformational impact on the legal profession within five years.
30% of respondents worried their firms were moving too slowly on AI adoption.
Reuters Legal — Small Firm Operations Guidance
Reuters Legal reporting on legal operations, technology, and practice management for solo and small law firms.
reuters.com/legal →For small firms, solutions built for larger organizations are often too expensive and time-consuming, while tools designed for non-legal businesses frequently miss practice-specific requirements.
Drawn from Reuters Legal editorial coverage of small-firm technology adoption. Reflects consistent editorial framing rather than a single study or report.
Primary Authority
ABA Formal Opinions & Ethics Guidance
ABA formal opinions are issued by the Standing Committee on Ethics and Professional Responsibility. They are primary authority on attorneys' professional obligations under the Model Rules — not survey data or industry reporting. Insights articles covering AI use, confidentiality, privilege, supervision, and related professional responsibility topics cite formal opinions in this section. State bar ethics opinions may supplement or differ from ABA formal opinions on a jurisdiction-specific basis.
americanbar.org — Formal Ethics Opinions →Note: ABA formal opinions address the ABA Model Rules. Jurisdiction-specific obligations may differ. Insights articles note where state rules or guidance vary materially.
ABA Model Rules of Professional Conduct: Rule 1.1 (competence), Rule 1.4 (communication), Rule 1.5 (fees), Rule 1.6 (confidentiality of information), Rule 5.1 (responsibilities of partners and supervisory lawyers), Rule 5.3 (responsibilities regarding nonlawyer assistance).
The Model Rules are a model. They bind no lawyer anywhere until a jurisdiction adopts them, and jurisdictions adopt them with modifications. Every ABA formal opinion in this group interprets these rules, which is why they are listed separately: an article can reference the duty a rule describes without implying that the ABA's reading of it governs in the reader's state. Where an article names a rule number, it is naming the Model Rule, and the reader's own jurisdiction controls.
ABA Formal Opinion 93-379 (1993): Billing for Professional Fees, Disbursements and Other Expenses
Predates generative AI but is quoted directly in Formal Opinion 512 on the billing question: a lawyer billing on an hourly basis does not fulfill her ethical duty if she bills the client for more time than she has actually expended. Cited here for that proposition as carried forward by FO 512, not as standalone guidance on AI.
ABA Formal Opinion 477R (2017): Securing Communication of Protected Client Information
Addresses lawyers' obligations under Rule 1.6(c) when transmitting client information electronically. Requires a fact-specific analysis of information sensitivity, tools used, and available protections. Holds that competence includes understanding the security implications of communication and storage tools. Directly relevant to attorney use of cloud services, email, messaging platforms, and AI tools that process client data.
ABA Formal Opinion 483 (2018): Lawyers' Obligations After an Electronic Data Breach or Cyberattack
Addresses duties under competence, confidentiality, supervision, and communication when a breach may have compromised client information. Holds that lawyers must take reasonable steps to stop and remedy a breach and must notify affected clients in appropriate circumstances. Relevant to discussions of security posture, vendor due diligence, and the firm's obligation to maintain adequate safeguards over client data.
ABA Formal Opinion 512 (2024): Generative Artificial Intelligence Tools
The most current ABA formal guidance on attorney use of generative AI. Covers five duties: competence (understanding how AI tools work and their limitations); confidentiality (duties when client information is submitted to AI systems); supervision (responsibility over AI-generated work product and non-lawyer staff use of AI); fees (obligations when AI changes the cost of services previously billed to clients); communication (candor and disclosure to clients about AI use). Issued July 2024.
Primary Authority
State Bar and Court Guidance on AI
Guidance issued by state bars, state ethics committees, and court systems. These bind or direct only within their own jurisdictions, and they do not always reach the same test as the ABA Model Rules analysis. They are listed here because the guidance layer has continued to develop since ABA Formal Opinion 512 was issued in July 2024, and a firm policy that cites only the 2024 ABA opinion is citing one layer of a stack that has kept moving.
Note: entries in this group are jurisdiction-specific. Nothing here should be treated as generally applicable. Firms should confirm what their own jurisdictions have adopted.
State Bar of California, Committee on Professional Responsibility and Conduct (COPRAC): Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law (2026)
Replaces the 2023 COPRAC guidance and extends it to agentic AI at the request of the California Supreme Court. Cited for three points: that a lawyer must not input confidential client information into a generative AI solution that may present material risks to confidentiality or security absent informed consent; that reasonable efforts to protect client confidences require more than reliance on generalized marketing assurances and may include reviewing terms of use, privacy policies, and vendor documentation; and that supervision should correspond to the level of system access and autonomy. The confidentiality trigger stated here is not identical to the ABA Formal Opinion 512 analysis, and the article notes the difference rather than reconciling it.
22 NYCRR Part 161, Use of Artificial Intelligence Technology (New York State Unified Court System). Adopted March 25, 2026; effective June 1, 2026.
Applies across the Unified Court System while leaving individual judges discretion. Cited only for the fact of adoption and the effective date. The full rule text was not independently retrieved at the time of writing, and no article makes any claim about its specific provisions. Retrieve the rule text before citing any substantive requirement from it.
New York City Bar Association, Committee on Professional Ethics: generative AI opinion series. Formal Opinion 2024-5 (generative AI generally), 2025-6 (AI recording and summarization of client conversations), and 2026-2 (non-client conversations), issued August 5, 2026.
Cited for the structural point that the guidance layer is a sequence rather than a single pronouncement, with each opinion narrowing to a more specific use. City bar opinions are advisory and do not bind New York courts or the state's disciplinary authorities.
Virginia Legal Ethics Opinion 1901: Reasonable Fees and Generative Artificial Intelligence. Adopted by the Supreme Court of Virginia November 24, 2025.
Addresses fee reasonableness where generative AI changes the time required for a task. Cited here as a dated entry in the guidance timeline. The article does not state its holdings; a firm relying on it should read the opinion.
Vendor Documentation
Anthropic Product and Commercial Documentation
These entries record what the vendor states in its own published terms, privacy documentation, and support articles. That is the right source for a claim about what a vendor commits to, and it is not independent verification that the vendor does it. We cite this material for the contract and configuration facts a firm needs in order to run its own diligence, never as evidence of security outcomes. Vendor marketing pages are not used.
privacy.claude.com →Note: product documentation changes without notice. Every entry below records what the documentation stated as of August 20, 2026. Re-verify before relying on any of it. Anthropic moved its privacy documentation to privacy.claude.com during 2026, so older links may redirect.
Anthropic states that by default it will not use inputs or outputs from its commercial products to train its models. Explicit user feedback (thumbs up or down) is an exception: Anthropic states it may use those chats and coding sessions to train, with that feedback retained up to five years. Organization owners can disable this through the "Rate chats" setting.
Covers Claude for Work, Enterprise, the API, and Claude Gov. The feedback exception is the operationally important part for law firms: it is triggered per conversation by an individual user action, which is why the article treats disabling it as an organization-level governance decision rather than a user preference. Verified August 20, 2026.
Anthropic documents API inputs and outputs as deleted within 30 days subject to stated exceptions, chat products as retaining conversations to provide a consistent experience, and Enterprise custom retention as configurable by a Primary Owner or Owner with a 30-day minimum covering chats and projects. Where no custom period is set, the documented default is that data is retained indefinitely.
Not trained on and not retained are different commitments, and the article separates them for that reason. The documentation addresses chats and projects; it does not state whether uploaded files and audit logs fall inside those controls. The article presents that gap as an open diligence question rather than assuming either answer. Verified August 20, 2026.
Effective June 9, 2026, Anthropic designates certain "covered models" whose prompts and outputs are retained 30 days across all platforms where they are offered. Organizations holding zero-data-retention agreements must enable retention in order to use covered models.
Cited for the governance consequence rather than the product detail: a negotiated zero-data-retention agreement is contingent on which model is used, so approving a new model can change a firm's data posture with no contract change and no visible product change. Model names in this category change; verify the current list rather than relying on any example. Verified August 20, 2026.
Anthropic documents a Compliance API, available to Enterprise plans (excluding Public Sector organizations) and Platform customers, that exposes activity feed events, chat data, and file content.
Content-level, not metadata-level, which is why the article treats it as a second repository of privileged material rather than purely as a supervision benefit. The documentation reviewed does not specify which organizational roles may obtain access, what approval workflow applies, or whether use of an access key is itself logged. The article states those as diligence questions. Verified August 20, 2026.
Anthropic's Commercial Terms of Service address "Third Party Features," providing that customers may elect to use features made available by third parties and that use of third-party content and integrations is subject to that third party's terms.
Cited for what the terms do not resolve. The reviewed terms do not define connectors, integrations, or MCP, and do not allocate data flow or liability when data leaves Claude for a third party. That is the basis for the article's recommendation that firms approve each surface and connector rather than approving a vendor. Terms version effective June 17, 2025; reviewed August 20, 2026.
Research & Reported Incidents
Independent Research and Reported Incidents
Peer-reviewed and preregistered empirical research, plus incidents that are useful as operational illustrations. Where an entry rests on secondary reporting rather than a primary record, the entry says so and the article says so.
Purpose-built AI legal research tools from LexisNexis and Thomson Reuters were each found to hallucinate between 17% and 33% of the time, on queries run May 23 to 27, 2024. The study defines a response as hallucinated if it is either incorrect or misgrounded, meaning key propositions are cited but the source does not support the claim.
Magesh, Surani, Dahl, Suzgun, Manning and Ho, "Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools." Stanford RegLab and HAI preprint (2024), subsequently published in the Journal of Empirical Legal Studies (2025). Described as the first preregistered empirical evaluation of AI legal research tools. Cited in ABA Formal Opinion 512. The figure describes retrieval-grounded legal research products, not general-purpose assistants, which is the reason the article uses it: it sets the reliability expectation for the strongest available category of tool, not the weakest. Methodology, verified against the JELS text: dataset preregistered March 22, 2024; 202 questions; queries run over a five-day window, May 23 to 27, 2024. Reported percentages are shares of all queries; the paper also reports a second, different rate conditional on the system giving a direct response, so the two figures in the paper are not interchangeable. Accuracy was 65% for Lexis+ AI, 41% for Westlaw AI-Assisted Research and 19% for Ask Practical Law AI, with incompleteness of 18%, 25% and 62% respectively. The authors state their own limits: closed systems, model drift over time, a deliberately challenging query set, n=202, and chat-style research only. CoCounsel was not evaluated, and the authors name contract review, deposition summaries and memoranda as open benchmarking problems. Do not quote the range in the present tense as a current product property.
In an employment matter in the Brazilian state of Para (2026), lawyers embedded hidden white-on-white instructions in a filed petition directing an AI reader to contest the petition superficially. The court's AI system flagged the text and the judge imposed a fine of R$84,000, roughly ten percent of the value of the case, with a referral to the bar.
Class caveat: this account rests on consistent secondary reporting across multiple legal-industry and law-firm sources. The court's own document was not retrieved. Foreign jurisdiction with no precedential value in the United States. Used strictly as an operational illustration that prompt injection through adversary-produced documents is a realized technique rather than a hypothetical, and labeled as reported in the article text. Do not cite it as authority or as a settled factual record.
Clio's 2025 Legal Trends research indicates that a majority of firms use flat fees in some form, with meaningful use of subscription pricing alongside them.
Used qualitatively rather than as a precise figure. The underlying percentages were captured during research but the exact population and scope qualifier attached to them were not confirmed, so the article states the direction of the finding and not a number. Do not restate this entry as a specific percentage without re-verifying the scope against the report itself. Separately, the Clio and Thomson Reuters definitions of "mid-sized" differ materially, so figures from the two should not be compared directly.
Vendor Documentation
Legal AI Platform Announcements
Product announcements and repository documentation from legal AI vendors and model providers. These establish what a vendor says it has shipped. They are not evidence that a capability performs as described, and product architecture in this category changes on a scale of weeks. Where a primary repository and press coverage disagree, we cite the repository and say so.
Note: entries here are the shortest-lived material in this registry. Re-verify before relying on any of it.
Anthropic released Claude for Legal on May 12, 2026: twelve practice-area plugins published under an Apache 2.0 license, plus more than twenty MCP connectors linking Claude to systems law firms already use.
Anthropic announcement plus the public repository at github.com/anthropics/claude-for-legal. Plugin and connector counts verified against the repository, not only the announcement. Anthropic names Claude chat, Claude Cowork, Claude for Microsoft 365, Claude Code and the Platform API as surfaces. The announcement makes no quantitative accuracy or hallucination claim, and we make none on its behalf. The model designation given in the May 2026 post may since have been superseded.
In Anthropic's connector documentation, the shipped legal research connectors are CourtListener, Descrybe, Trellis, TopCounsel, Definely and Solve Intelligence. Thomson Reuters (CoCounsel, Practical Law, Westlaw) appears under "Wanted connectors," not "Current connectors." LexisNexis does not appear in the file.
This entry contradicts several trade reports that describe Westlaw research through CoCounsel as a live Claude connector. Read directly from the raw CONNECTORS.md in the anthropics/claude-for-legal repository on August 21, 2026, where the Thomson Reuters line sits under the "Wanted connectors" heading. Possible explanations for the discrepancy include the repository lagging the product, a connector shipping outside the default configuration, or the reporting being wrong; we did not resolve which. The article states what the repository documents and directs readers to verify the current state themselves. This is the most perishable claim on the site and should be re-checked before any republication.
On May 13, 2026, LexisNexis announced that it had integrated Anthropic's Claude legal plugin suite into Lexis+ with Protégé, running within the Lexis environment.
LexisNexis press release. Cited for the fact and direction of the integration, which is the evidence for a legal platform absorbing a frontier model's workflow layer rather than the reverse. The release's corpus figures, including a count of legal documents and a daily addition rate, are vendor marketing claims and are deliberately not restated as fact in any article. Artificial Lawyer reported an earlier LexisNexis engagement with a Claude Cowork legal plugin in February 2026; we did not verify the relationship between the two announcements.
Benchmarks
Legal AI Benchmarks and Evaluations
Benchmark results are only meaningful alongside their method. Each entry below records the unit of measurement, what counted as a failure, who ran the evaluation, and which vendors participated, because those four facts determine whether two published numbers can be compared at all. They usually cannot. The Stanford study, which is the reference point for this group, is listed separately under Independent Research and Reported Incidents.
Note: a benchmark score is not a product rating, and a vendor-created benchmark is evidence without being independent evidence. Both are labeled as such below.
In Harvey's own BigLaw Bench evaluation, Harvey Assistant is reported at roughly a 0.2% hallucination rate, with named foundation models higher. The rate counts sentences containing a hallucinated claim divided by total sentences in a response, defines a hallucination as a factual claim that can be demonstrably disproven against a source of truth, and excludes errors of understanding or reasoning.
Vendor-created and vendor-run, published October 7, 2024, using an automated scoring system with human review of its judgments. Cited on this site only to demonstrate why hallucination rates are not comparable across benchmarks, never as a product ranking and never as a figure to set beside the Stanford range. Four limits matter. The denominator is the sentence rather than the response. Reasoning errors are excluded by design. Misgrounding, as the Stanford authors define it, is not clearly captured by the stated definition. And the comparison is between a product carrying retrieval and workflow scaffolding and foundation models used directly, which is a reasonable thing to measure and is not a like-for-like model comparison. Model versions and the test window are not stated, so the result cannot be dated more precisely than its publication date or reproduced; the model names given refer to generations superseded more than once since. The summary line is also internally inconsistent: of the three published ratio-and-percentage pairs, two reconcile (1 in 500 with 0.2%, 1 in 150 with 0.7%) and one does not, since 1 in 110 implies roughly 0.9% rather than the 1.9% published alongside it. The accompanying data table lists 1.9%, which suggests the ratio is the slip. For that reason no individual foundation-model figure is quoted on this site. The article treats this as evidence that benchmark measurement is genuinely hard rather than as a criticism of the publisher, which documents more of its method than most.
The Vals Legal AI Report (February 27, 2025) evaluated legal AI tools across seven tasks against a lawyer baseline, including data extraction, document question answering, summarization, redlining, transcript analysis, chronology generation and EDGAR research.
Independently operated but vendor-participatory, which is the limitation that matters. Vendors elected which tasks to enter, so the field differs task by task and the results are not a uniform comparison. Lexis+ AI was initially involved and withdrew from the sections covered. Hallucination was not a measured outcome, so a strong task score in this report is not evidence of a low hallucination rate. Read through trade reporting (LawSites, Artificial Lawyer) rather than the primary report; exact scores and rubric should be confirmed against Vals directly before any figure is quoted.
In a Vals legal research benchmark reported October 23, 2025, tested tools scored roughly 79% to 81% on 210 questions across nine research categories, against a lawyer baseline of about 71%. A general-purpose model scored within about one point of the best legal-specific tool. Thomson Reuters, LexisNexis and vLex all declined to participate.
Scored on a weighted rubric combining accuracy, authoritativeness and appropriateness. Cited as the strongest published evidence that general-purpose models have narrowed the gap on legal research, and always together with its central limitation: the three largest legal research vendors were not in the study, so it describes the participants rather than the market. Hallucination was not measured as a distinct outcome. Read through trade reporting of the Vals results; the article states the figures as approximate for that reason.