Can AI do legal research?
Verify the source.

A clickable source can still mislead.

When a legal answer cites an authority, check more than whether the link opens. Does the case or statute exist? Is it the authority the answer names? Does the cited passage actually support the stated rule? Is the authority relevant to the jurisdiction, issue, and date in question?

A citation to a real but unrelated case is not reliable support. A summary can also confuse a party’s argument with the court’s holding, ignore a later decision, or leave out the limits that shape a rule. These are different failures, and a simple “citation found” indicator catches only some of them.

Legal research depends on authority and context. A persuasive answer must identify the applicable law, distinguish binding from nonbinding materials, and follow how a rule changed. Fluent prose and neat formatting establish none of those things.

Retrieval can bring documents into view.

Retrieval-augmented generation (RAG) searches a collection of documents, then gives selected material to a language model as context for an answer. In legal tools, the collection may include cases, statutes, regulations, or practitioner guidance.

This can make a system more useful than asking a general-purpose model to answer from learned patterns alone. But retrieval does not decide whether a document governs the question, whether it is still good law, or whether a passage entails the sentence written around its citation. Those steps require legal interpretation and careful checking.

A 2025 preregistered study compared three commercial legal-research tools with GPT-4 on 202 open-ended U.S. legal questions. The queries covered general doctrine, jurisdiction- and time-specific issues, false premises, and factual recall. The researchers classified 17% to 33% of responses from the retrieval-based tools as hallucinated in that evaluation; those systems hallucinated less often than the closed-book GPT-4 comparison.

The study’s versions were queried in 2024, before publication, and its deliberately challenging question set is not a representative sample of every lawyer’s everyday use. Treat those numbers as evidence that retrieval did not eliminate the problem in that evaluation—not as a current score for today’s products.

Asking another model to verify is not a final audit.

A 2026 research preprint examined automated checking of legal citations. The authors built a benchmark of 1,300 legal-brief excerpts containing citation errors drawn from observed failure types, and tested models with and without an agent that could search legal materials. Tool-assisted checking improved detection, but errors involving exact page references, quotations, and whether case content truly supports a claim remained difficult.

This benchmark is a controlled test with inserted errors, not a trial of AI review in courts or law firms. It shows that automated checking can help find some problems while also leaving important ones for a qualified reader. Better retrieval makes review more informed; it does not make review unnecessary.

Turn the answer back into research.

1. Fix the scope: Name the jurisdiction, court or agency, relevant dates, and legal issue before judging the answer.
2. Open the primary source: Go to the actual opinion, statute, rule, or official document—not only a chatbot summary or a secondary commentary.
3. Match claim to passage: Read the cited page or paragraph and confirm it supports the precise proposition attributed to it.
4. Check legal status: Look for later history, amendments, reversals, or controlling authority that changes how the source applies.
5. Separate voices: Distinguish a party’s argument, a dissent, dicta, and the holding that controls the result.
6. Check completeness: Ask what the answer left out, what counterauthority applies, and whether the question contains a false premise.
7. Protect confidential material: Understand what a tool stores and who can access it before entering client or case information.

That workflow is an evaluation method, not legal advice. A useful system can help surface a research trail, but the person responsible for a decision still needs access to the authority, time to inspect it, and the expertise to judge what it means.

Evidence follows the whole path.

For the retrieval method itself, read what RAG can and cannot do ↗. For a framework for judging AI claims, explore how to read AI benchmarks ↗ and how to put accountability behind a system ↗.