Legal AI With No Hallucination? What the Evidence Says and How to Test a Tool

The database nobody wanted to need

Damien Charlotin, whom the Fifth Circuit calls "a French lawyer and data scientist," keeps a public database of court decisions in which a party relied on AI-hallucinated content. On 12 October 2026 it said it had been last updated on 10 October 2026 and listed 2,184 cases identified, 1,492 of them in US courts (damiencharlotin.com/hallucinations). The line has not bent toward zero. We read it in more detail in Charlotin's tracker, read like a risk manager.

So buyers ask a fair question: which legal AI avoids hallucinating cases, and how would you know before you bet a filing on it?

The honest answer is not the one most vendors give. No tool is shown to avoid hallucination entirely. What differs between tools is how much they reduce it, and how quickly you can check what they tell you.

Reviewed and corrected on 12 October 2026. This is general information, not legal advice. Product versions change, so the figures below describe the 2024 versions Stanford tested.

Short answer: Stanford's preregistered study found that research tools from LexisNexis and Thomson Reuters, built on retrieval, hallucinated less than GPT-4 but "each hallucinate between 17% and 33% of the time." The best of them answered 65% of the study's queries accurately. So the useful question is not "which AI never hallucinates" but "which tool lets me open every cite and read the source."

TL;DR

• Charlotin's database listed 2,184 decisions worldwide, 1,492 in the US, as of 10 October 2026.
• Stanford (Journal of Empirical Legal Studies, 2025) tested 2024 versions: Lexis+ AI hallucinated on 17% of queries, Westlaw AI-Assisted Research on 33% and Ask Practical Law AI on 17%, against 43% for GPT-4.
• Retrieval reduces errors without removing them. Courts have said so too: in Farris (6th Cir., 3 April 2026) counsel used Westlaw's CoCounsel and the court said the risk persists "even when new tools are sponsored by trusted legal technology providers."
• The harder failure is a real case attached to a holding it never made. It passes a quick "does it exist" check.
• "RAG" on a label proves nothing if the model can override what retrieval found.

Part of our legal AI verification and hallucination guide series.

For related coverage, see AI hallucinations in legal research and how to avoid sanctions and how to verify AI legal citations before filing.

Two ways to build a lawyer's tool

Strip away the branding and there are two designs under most legal AI products.

The first is generative recall. A large language model answers from what it absorbed in training. It "remembers" that a Fourth Amendment cell-site case exists and reconstructs the citation from patterns. Sometimes it lands on Carpenter v. United States, 585 U.S. 296 (2018). Sometimes it produces a citation that looks just like it, for a case that was never decided. The model is not lying. It is generating likely text.

The second is grounded retrieval. The system first searches a corpus of opinions, pulls the documents, and asks the model to write from them. The model's job shrinks from recalling the law to summarizing documents it was just handed. Fabrication becomes harder, because the cases are in front of the model.

A product that says "RAG" on the label can still run on recall with a retrieval garnish, if the model is free to ignore what was retrieved. Retrieval is only grounding if the answer is constrained to it. Our explainer on how AI legal research works covers the mechanics.

What Stanford found

In 2025, Stanford RegLab and HAI researchers published a preregistered evaluation in the Journal of Empirical Legal Studies (22:216, Magesh et al.; the preprint appeared in May 2024). They tested three research products with queries run between March and May 2024: Lexis+ AI, Ask Practical Law AI and Westlaw AI-Assisted Research. They compared them with GPT-4.

These are the figures in the paper's Figure 4. (The paper's running text rounds Westlaw's and Ask Practical Law AI's accuracy to 41% and 19% and Ask Practical Law AI's incomplete share to 62%.) Two points stand out. First, the legal tools hallucinated less than GPT-4, which is what retrieval should do. Second, the picture is not simple: Ask Practical Law AI gave incomplete answers (refusals or ungrounded responses) for 63% of queries, and the paper reports 20% of its answers as accurate.

The paper's abstract says the providers' claims that these tools eliminate or avoid hallucinations are "overstated"; its introduction cites 2023 language from Casetext, Thomson Reuters and LexisNexis. Its authors also make a point about definitions. They count a response as a hallucination if it is incorrect, or if it describes the law correctly but cites a source that does not support the claim. A tool that "links only to Brown v. Board of Education on every query," they write, would on a narrow definition be "hallucination-free," which they call "a plainly irrational result."

The lesson is not that one product is good and another bad. Versions change and a study is a snapshot of 2024 products. The lesson is that retrieval reduces hallucination without eliminating it, and that building a tool for lawyers does not by itself make its output safe to file.

The failure that actually hurts

When people picture an AI hallucination they picture Mata v. Avianca: opinions that do not exist at all, which a database search catches. The quieter failure is a real case, correctly cited, attached to a proposition it does not support.

Imagine a tool tells you Terry v. Ohio, 392 U.S. 1 (1968), supports a warrantless search of a locked container in a parked car. Terry is real and the citation is right, but it concerns a stop and frisk of a person on a sidewalk. A check that confirms only that the case exists would pass it.

Real courts have seen versions of this. In Farris (6th Cir., 3 April 2026) none of the cases was fake, but the briefs carried invented quotations and misdescribed the holdings of real cases. The court denied counsel his fee for the appeal and referred the matter for possible discipline. In Noland v. Land of the Free (Cal. Ct. App., 12 September 2025), 21 of 23 case quotations in the opening brief were fabricated.

How a buyer can test which machine they are using

You will rarely get the architecture diagram. You can learn a lot in fifteen minutes by probing where an answer comes from.

Ask something obscure. Pick a narrow question in a small jurisdiction. A grounded system with little to retrieve should say so. A system answering from memory will often produce a citation anyway.

Open every citation, not just the first. Check whether the case, at the pinpoint, says what the summary claims. If the tool cannot take you to the passage it relied on, it is not showing its grounding.

Ask for authority against the answer. A grounded tool searches again and returns real contrary authority, or reports there is little. A recall-based tool may generate symmetrical citations for both sides.

Test a known change in the law. Ask about Chevron deference. A current tool should mention that Loper Bright Enterprises v. Raimondo, 603 U.S. 369 (2024), overruled it. If it does not, you have learned something about its currency.

Make it quote a statute. Ask for the text of 17 U.S.C. § 107 or 42 U.S.C. § 1983 and compare it to the US Code. Statutory text is a clean test because one wording is right.

None of these require trusting a vendor. They require the product to behave correctly in front of you, which is the only evidence that matters.

What courts have said about relying on a tool

The orders do not distinguish by tool. In Farris, counsel admitted using Westlaw's CoCounsel for a first draft, and the Sixth Circuit wrote that the risk remains "even when new tools are sponsored by trusted legal technology providers." In Fletcher v. Experian (5th Cir., 18 February 2026) counsel named Casetext and vLex among her sources, and the court did not find which tool produced the errors. In Lacey v. State Farm (C.D. Cal., 5 May 2025) a special master sanctioned two firms $31,100 after an outline generated with AI tools was built into a brief without anyone cite-checking it.

ABA Formal Opinion 512 (29 July 2024) says the amount of verification "will necessarily depend on the GAI tool and the specific task that it performs," and that a tool designed for law "may require less independent verification or review, particularly where a lawyer's prior experience with the GAI tool provides a reasonable basis for relying on its results." It does not remove the duty. See our ABA 512 guide.

What to ask a vendor

• Can I open the source behind every citation, and does it take me to the passage?
• What happens when you have no good source? Does the product say so?
• Does the answer come only from retrieved documents, or can the model add from memory?
• Which versions of the product has independent testing covered, and when?
• Where do my inputs go, and are they used for training? See where your legal AI data actually goes.

FAQ

Which legal AI does not hallucinate cases?

None has been shown to avoid hallucination entirely. Stanford's study of 2024 products found every tested tool hallucinated on some queries, at rates from 17% to 33% for the legal research tools. Treat "does not hallucinate" as "lets me verify each citation quickly."

What is the most accurate legal AI?

In Stanford's 2024 testing, Lexis+ AI was the most accurate of the three research products, with 65% of answers accurate. Versions change, so a benchmark from 2024 says little about a product today. Ask the vendor for current independent testing.

Is there a legal AI with no hallucination?

No tool in the evidence we reviewed meets that bar, and the Stanford authors found provider claims of "hallucination-free" legal citations overstated. A defensible claim is narrower: every cited case traces to a document you can open.

Does retrieval-augmented generation stop hallucinations?

It reduces them. In Stanford's testing, the retrieval tools hallucinated less than GPT-4 but still did on 17% to 33% of queries. It helps only if the answer is constrained to what retrieval found.

What is the worst kind of legal AI hallucination?

A real case, cited correctly, that does not support the proposition. A fake case is caught by a database search. A misdescribed holding or an invented quotation from a real case needs someone to read the opinion.

How do I test whether a legal AI hallucinates?

Ask an obscure question in a small jurisdiction, open every citation and read the passage, ask for authority against the answer, test a known change such as Loper Bright, and have it quote a statute you can check against the official text.

Why are lawyers still being sanctioned for AI citations?

Because the failures are quiet and some filings are not checked. Courts have sanctioned unverified filings in cases such as Mata, Wadsworth, Whiting and Lnu. See our sanctions guide.

For the US statutes, regulations and court rules these laws sit in, see the US primary law API.

Vaquill AI

Product & Content

The US primary law data API: statutes, regulations, constitutions, court rules, executive orders and agency guidance across 53 US jurisdictions, each section cited and linked to its official source, via REST, MCP or SQL.