Which AI is best for legal research?

Legal research quality is mostly a property of the corpus, not the model. All the serious tools now sit on capable frontier models, and none of them can cite a case their corpus does not hold. Three questions separate them.

The longer answer depends on your jurisdiction, your practice area and what you already license.

If your question is what the law says, the default shortlist is the two tools attached to the major legal publishing corpora.

Thomson Reuters CoCounsel Legal is grounded in Westlaw primary law and Practical Law guidance. Its Deep Research capability runs multi-step research against that corpus and produces cited work product, and Thomson Reuters announced a UK expansion in January 2026, which matters if your practice is here rather than in the US.

Lexis+ with Protege, which is what LexisNexis renamed Lexis+ AI to in early 2026, works over the LexisNexis collection of case law, statutes and secondary sources, with citation validation through Shepard's. It takes a multi-model approach and can be pointed at an organisation's own documents where that is enabled.

Between the two, most firms choose the one whose corpus they already license. The switching cost is the corpus subscription, not the AI.

Harvey positions itself as a platform across research, drafting, due diligence and workflow automation rather than a research tool alone. In May 2026 it published an open benchmark of legal agent tasks, which is worth reading whichever tool you buy, because it is one of the few public attempts to measure this category rather than market it.

Legora, a Swedish company, offers a collaborative workspace covering research, review and drafting, with a Word add-in and European data-residency options. Residency is a genuine consideration for UK and EU firms whose information-governance teams will ask where queries and documents are processed.

Vincent AI, from vLex, answers research questions over what vLex describes as a global database of more than a billion legal documents, returning cited answers that link to the underlying sources. Its claim to the job is corpus breadth across jurisdictions.

For any of the three, press the same question you would put to the publishers. Show me corpus coverage for my jurisdiction and practice area, and show me what happens when the answer is not in it.

ChatGPT and its peers answer legal research questions fluently, and that is the problem. They have no authoritative legal corpus, no citation validation, and a documented record of inventing authority. The 2023 Mata v. Avianca sanction was the first well-known case of a lawyer filing fabricated AI citations, and a public database tracking such court decisions has since grown past a thousand documented cases worldwide, with penalties escalating from four-figure fines to five-figure sanctions and, in a few cases, suspensions.

None of this makes general models useless to lawyers. They orient, summarise and draft well. But research means producing authority you can rely on, and for that, citations that resolve to real and current sources are the entry requirement rather than a premium feature.

Everything above concerns law outside your files. A large share of what gets called legal research in practice is a question about material the firm already holds. What did we agree in this engagement, which clause governs, what did the amended agreement end up saying. That is document analysis, and it rewards different evidence: whether the system retrieves the right passage from your own estate, whether its citations point at the exact passage rather than a whole document, and whether access boundaries between matters hold.

Marella sits in that category, not this one. It answers questions over a corpus built from the client's own documents, cites to the exact passage with the source opened alongside and the passage marked, runs as a single-tenant deployment, and fact-checks answers against their sources after generation, flagging inaccurate citations and unsupported claims. It holds no case-law corpus and is not a substitute for the research tools above. If you need both jobs done, you are shopping in two categories.

Vendor benchmarks are run on questions chosen to succeed. The evaluation that predicts anything is the one you run on your own questions.