Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

AI legal research tools generally combine two steps: they retrieve material from a legal content collection, then use a generative model to draft an answer from it. A citation link can help you inspect a source, but it does not prove that the answer accurately describes the case, that the case applies to your issue, or that it remains good law. Treat the output as a starting point and verify every authority you may rely on.

How AI finds case law

The exact technology and source coverage differ by product, and vendors do not publicly describe every query-processing or ranking detail. At a high level, many legal research assistants follow a retrieval-and-generation workflow: they search a defined collection for potentially relevant materials and use a generative model to compose a response based on what was retrieved.

  1. Frame the question. The researcher asks a question in natural language. In a peer-reviewed 2025 evaluation, Lexis+ AI and Ask Practical Law AI accepted chatbot-style queries, while Westlaw AI-Assisted Research retrieved material from Westlaw legal databases. A clear question should identify the jurisdiction and legal issue; include important facts or procedural context when they affect the answer.
  2. Retrieve candidate sources. Depending on the service and subscription, the collection may include cases, statutes and regulations, secondary sources, editorial material, or practice guidance. Thomson Reuters says its Deep Research uses Westlaw and Practical Law tools and content, including primary law, administrative materials, secondary sources, and current-awareness material. That describes the vendor’s offering; it is not independent proof that the collection is complete for a particular question.
  3. Generate an answer and source links. The model synthesizes material into a narrative and may attach links to source documents or citations. For example, Thomson Reuters says CoCounsel Legal draws on Westlaw, Practical Law, and firm knowledge and includes linked citations. Lexis describes linked citations and Shepard’s verification features. The link is a route to the document—not proof that the generated sentence is a faithful account of it.

In other words, retrieval asks which materials might be relevant; legal verification asks whether an authority actually supports the proposition and applies to the question at hand. The second task still requires reading the authority in context.

Where errors can enter

A tool can fail before it drafts a sentence, or it can retrieve a useful source and then describe it incorrectly. Magesh and co-authors’ peer-reviewed 2025 evaluation documents several kinds of failure:

What’s actually slowing this PC down?

Pick the symptom - the matching free tool is one click away.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.
Stage Possible failure What to check
Understanding the question The system may misunderstand the legal issue or accept a mistaken premise, leading it toward adjacent rather than controlling material. Confirm that the answer addresses the actual doctrine, issue, and facts—not just similar wording.
Retrieving authorities A result may concern a different jurisdiction, legal context, or procedural situation. In one example from the evaluation, a system retrieved a source about “moral turpitude” in response to a question about the distinct “moral wrong doctrine.” Check the court, jurisdiction, issue, and procedural posture of every material case.
Synthesizing the source Generated prose may misstate a passage or make a source sound more supportive than it is. Read the cited decision itself and compare its holding and relevant passage with the answer’s specific claim.

A citation can be genuine and still be irrelevant, distinguishable, or misrepresented. Conversely, a plausible-sounding answer may rely on a source that does not exist. Source links reduce the effort of locating documents; they do not eliminate the need to check them.

How to verify an AI-cited case

Use the assistant to identify leads, then verify each authority independently before relying on it in legal advice or a filing. A practical review sequence is:

  1. Open the cited decision. Do not rely only on the AI’s summary, quotation, or link text. Confirm that the cited case exists and that the document is the decision the answer describes.
  2. Check identity and context. Confirm the court, date, jurisdiction, procedural posture, and relevant facts. A decision from another jurisdiction or a different procedural setting may not answer the question.
  3. Test the proposition against the opinion. Locate the passage the answer relies on. Read enough surrounding text to understand what the court decided, whether the statement is a holding or dicta, and whether the case actually supports the proposition as phrased.
  4. Check later history and treatment. Use a citator such as Shepard’s or KeyCite, where available, to review subsequent history and citing references. Read relevant later decisions and treatment; do not treat a positive status signal as proof that the case controls your issue.
  5. Assess fit and completeness. Consider whether the authority governs the jurisdiction and issue, whether there are important distinctions, and whether the answer may have omitted a more central or contrary authority. Apply legal judgment to the facts rather than treating the generated answer as a conclusion.

The Law Society of England and Wales has warned that members encountered cases that, on checking, “have turned out to be a fake citation, a misrepresentation of a document, or even a piece of legislation from another jurisdiction incorrectly described as English and Welsh law.” Its warning illustrates why checking both source authenticity and legal fit matters.

What published evaluations show—and what they do not

Published error figures are evidence of risk in specific tests, not a score you can apply to every query or current product release. The New York State Unified Court System Advisory Committee’s 2025 report gives figures from a Stanford evaluation: the tested Lexis product had a reported 17% hallucination rate and 65% accuracy rate; the tested Westlaw product had a reported 33% hallucination rate and 42% accuracy rate. Magesh and co-authors’ 2025 paper reports hallucination rates from 17% to 33% across the three tools in its benchmark.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

Those numbers describe the particular systems and query sets tested. They do not predict the result of an individual search, establish how a current version performs, or provide a universal comparison of all legal research tools. Product versions and features can change.

The same New York court-system report describes summer 2024 trials involving nearly 100 judges, court attorneys, law clerks, and law librarians. Participants saw potential time savings for preliminary tasks such as finding on-point sources and preparing first drafts, while also reporting that outputs were imperfect and needed review and correction. The committee’s 2025 report states: “Even when using the AI-enhanced features that have been incorporated into established legal research platforms, any content generated by AI should be independently verified for accuracy.”

Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Support on Ko-Fi

How to compare legal research assistants

Product names and features alone do not tell you whether a tool is suitable for a specific matter. When comparing services, assess the workflow and the authority you can actually inspect:

  • Corpus and jurisdiction: Which legal materials and jurisdictions are available under your subscription? Are the relevant primary authorities included?
  • Source trail: Does the answer link to primary authority and, where offered, pinpoint passages? Can you see what material informed the answer?
  • Treatment checking: Is a citator available, and can you inspect subsequent history and treatment rather than relying on a single status indicator?
  • Issue handling: Does the tool recognize jurisdiction limits, procedural posture, false premises, and distinctions in the facts?
  • Review and security: What human-review process and data controls does your practice require when using the service?

Vendor materials can explain what a product says it offers; independent evaluations and direct inspection of authorities are needed to assess performance. Thomson Reuters’ Deep Research help page puts the intended role plainly: “Use Deep Research to accelerate your research, not to replace it.”

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.