AI Agents for Lawyers

Use Cases

AI Legal Research in 2026: What Works, What Hallucinates

AI legal research now beats a lawyer baseline on accuracy yet misapplies law in 24% of answers. Which tools are grounded, where errors hide, a routine that holds.

By Dr. Niklas SchmidtUpdated 13 November 202611 min read
On this page
  1. Two numbers that appear to contradict each other
  2. Grounded and ungrounded: the tools, as of September 2026
  3. Where the errors hide: five failure types
  4. A research routine that survives a show-cause order
  5. Austria, Germany and the EU: what the agents can and cannot reach
  6. What to do this week

As of September 2026, AI legal research works when the tool retrieves from a citator-backed database and the lawyer opens every authority before relying on it. In Vals AI's October 2025 test the best systems beat a lawyer baseline on US questions, yet HAQQ's June 2026 benchmark found that 24% of answers cited or misapplied law that did not support the claim. Both findings are correct, and the gap between them is where the risk lives.

Two numbers that appear to contradict each other

The first number comes from Vals AI's legal research study of 14 October 2025: on 199 US research questions, ChatGPT Team scored 80% weighted accuracy, Counsel Stack 78%, Alexi and Midpage 74–77%, against a lawyer baseline of 69%. The lawyers averaged about 23 minutes per question. Every system lost roughly 11–14 points on multi-jurisdiction questions, and Thomson Reuters, LexisNexis and vLex declined to take part (Vals AI).

The second comes from the HAQQ benchmark of 5 June 2026: 3,000 graded answers to 300 tasks across 51 practice areas and 20+ jurisdictions. Claude Opus 4.8 came first overall, GPT-5.5 had the lowest citation hallucination rate at 3% but was the slowest (134 seconds) and dearest ($0.082 per task), o3 hallucinated in 32% of answers and Mistral Large produced citation errors in 64%. Across all models, 24% of answers "cited or misapplied law that didn't support claims" (HAQQ).

These measure different things. Vals scored whether the answer was right. HAQQ scored whether each authority said what the model claimed. A model can reach the correct conclusion by citing the wrong case, or the right case for the wrong proposition, and a partner reading the memo will never know unless someone opened the source.

StudyDateWhat it scoredHeadlineWhat it means at your desk
Stanford RegLab, "Hallucination-Free?"Tools tested May 2024 (older products)Hallucination on 202 expert-scored queriesLexis+ AI 17%, Westlaw AI-Assisted Research 33%, GPT-4 43%Grounding cut hallucination but did not remove it, even in the incumbents' products (RegLab)
Vals AI Legal Research14 Oct 2025Weighted accuracy of the answerBest tools 74–80% vs lawyers 69%AI is competitive on outcome; the multi-jurisdiction drop is the warning
HAQQ5 Jun 2026Whether each citation supports the claim24% of answers misapplied law; 3–64% citation hallucination by modelRead the pincite, not just the case name
Vals Legal Research Bench15 Sep 2026Leaderboard, 61 models1. Muse Spark 1.3 Max, 2. Claude Opus 5, 3. Claude Fable 5.1Rankings change monthly; test on your own questions
Astra for Law (OpenAI-reported)17 Sep 2026Vals Legal Research Bench, 200 questions54% correct vs 38.7% for GPT-6 with plain web searchA legal index helps a general model; it does not make it a citator (LawSites)

The benchmarks guide explains the scoring rules behind each of these, and why a vendor's headline figure and a strict figure can differ by 80 points.

Grounded and ungrounded: the tools, as of September 2026

"Grounded" here has a narrow meaning: the tool retrieves authorities from a licensed or curated legal database and checks them against a citator before it answers. Everything else, however good the model, is answering from memory plus a web search.

ToolGroundingVerification built inReachPrice signal
Lexis+ with ProtégéLexis case law, statutes, Practical GuidanceShepard's-grounded verification; Legal Intelligence Engine (24 Aug 2026) presents an executable plan for approval before runningUS GA 24 Feb 2026; 300+ workflowsSubscription, not published
CoCounsel Legal (next generation)Westlaw, Practical LawKeyCite; Deep Research Verify; built on Anthropic's Claude Agent SDK plus TR's own legal LLMUS GA 20 Aug 2026; Canada, UK, Australia later in 2026Published range $104–$639 per user per month
Clio Work (vLex Vincent)1bn+ documents, 110+ jurisdictionsSource-linked answers; Vincent "refused to answer, rather than hallucinate" in Vals' February 2025 testStandalone since 21 Apr 2026$199 per user per month
Astra for LawLegal Search Index of 230M+ URLs of US law built with CourtListener/Free Law ProjectNo citator; 26 partner plugins including Thomson Reuters and HarveyAnnounced 17 Sep 2026; US lawChatGPT Enterprise; zero-data-retention API
Claude for Legal (open source)Whatever connector you attach: CourtListener, Trellis, CoCounsel, LexisNexis, MidpageAny citation not checked against a connected research tool is tagged [verify]Runs in Cowork, Claude Code, Microsoft 365No extra charge on paid Claude plans
ChatGPT, Claude, Gemini with web searchTraining data plus live webNoneAny jurisdiction with public sources$20–$200 per month
Gemini Notebook (ex-NotebookLM)Only the sources you uploadPassage-level citations to your own filesClosed universe; cannot search the open webWorkspace or enterprise account

Sources: LawSites on Protégé (Feb 2026), Thomson Reuters (20 Aug 2026), LawSites on Clio Work (Apr 2026), Claude for Legal repo. The site's reviews of Lexis+ Protégé, CoCounsel and Astra for Law go into each product.

Two cautions on the ungrounded row. Adding web search does not fix citations: a University of Pennsylvania study of 3 April 2026 found search-augmented models and deep-research agents fabricated 3–13% of URLs, and the agents hallucinated more because they cited more (arXiv). And the sanctioned lawyers were not only using consumer chatbots: Lacey v. State Farm (C.D. Cal., 6 May 2025) involved CoCounsel, Westlaw Precision and Gemini, and Fletcher v. Experian (5th Cir., Feb 2026) traced 16 fabricated quotations to vLex and CoCounsel. Grounding lowers the rate. It does not delegate the duty.

Where the errors hide: five failure types

Damien Charlotin's database of court decisions involving hallucinated AI material stood at 2,041 on 14 September 2026, with 811 involving lawyers. It separates the errors into types, and the split is instructive: 1,690 instances of fabricated case law, 854 misrepresented citations and 548 false quotations (Charlotin).

  1. The invented case. The Mata v. Avianca (S.D.N.Y. 2023) pattern. Easiest to catch: the citation returns nothing in any database. Still the most common.
  2. The misgrounded citation. A real case, cited for something it does not say. Charlotin logs 854 of these, and they are the ones a name-check misses. Clearbrief's founder has said many "are not easy to spot with the human eye". Only reading the pincite catches them.
  3. The false quotation. Real case, invented words in quotation marks. Oregon courts now price this at $500 per fabricated citation and $1,000 per fabricated quotation, the formula behind a $10,000 fine in March 2026 (David Kemp's Q1 2026 review).
  4. The invented rule or paragraph number. In Cork v Smith [2026] EWHC 1199 (Ch) a Pinsent Masons associate's research produced "Insolvency Rule 12.37(5)", which does not exist; the tool had flagged its own answer as possibly inaccurate, and nobody checked (UK Human Rights Blog). German courts see the same with commentary Randziffern.
  5. Agreement with your premise. Stanford HAI's 2026 AI Index reports that when a false premise is framed as the user's own belief, model accuracy collapses: GPT-4o from 98.2% to 64.4%, DeepSeek R1 from over 90% to 14.4% (Stanford HAI). A partner who asks "confirm that X applies" will usually be confirmed.

A sixth is jurisdiction drift. The "Place Matters" study (2025) found GPT-4o, Claude 3.5 Sonnet and Gemini 1.5 Pro hallucinated on 45% of Los Angeles questions, 55% for London, 61% for Sydney, and 100% on Australian residential tenancy law (AI Law Librarians). The further the question sits from US federal law, the less the model knows and the more confidently it fills the gap.

A research routine that survives a show-cause order

Courts describe citation verification as a "nondelegable duty" that AI "in no way abrogates". The routine below keeps the duty where it belongs and still captures most of the speed.

  1. Frame with a general model, and forbid citations. Use Claude, ChatGPT or Gemini to sharpen the issue, list the elements and generate search strings. Tell it explicitly not to supply authorities.
  2. Search in a grounded tool. Run the strings in Protégé, CoCounsel, Clio Work or your national database. Export the authorities you intend to use, with pincites.
  3. Feed the verified set back. Ask the model to analyse only the authorities you have provided and to mark any proposition it cannot tie to a provided source.
  4. Run three checks per citation. Existence (the case is real and the citation resolves), support (the pincite says what you claim; read the paragraph), treatment (KeyCite or Shepard's shows it is still good law).
  5. Keep a research log. Tool, model, date, query, who verified which citation and when. If a show-cause order arrives, this is the exhibit.

The framing prompt for step 1:

You are assisting a [jurisdiction] lawyer with legal research. Do NOT provide case citations, statute section numbers, or quotations from authorities; I will retrieve those from a citator-backed database myself.

Question: [state the question and the facts in three sentences]

Produce:
1. The legal issues raised, as a numbered list, each in one sentence.
2. For each issue, the elements or tests a court would apply, stated generically.
3. Five to eight search strings I can run in a legal database, using terms of art for [jurisdiction].
4. The strongest counter-argument to the client's position, in two sentences.
Flag any point where the answer likely differs between [jurisdiction] and US federal law.

The step 3 prompt is one line: "Work only with the authorities I have provided in the attached list. Where a point is not supported by a provided authority, write [no source] instead of citing anything." Anthropic's own litigation plugin ships with the equivalent convention, tagging unchecked citations [verify] (Claude for Legal).

One rule never to break: do not ask the model whether its citations are real. The lawyers in Mata did that. In April 2026 a Cherry Hill, New Jersey lawyer was sanctioned a second time, $5,000, for "using AI to verify AI" (Compliance Hub). The full citation-checking protocol on this site expands steps 4 and 5 into a checklist, and the brief-drafting page takes the verified set through to filing.

Austria, Germany and the EU: what the agents can and cannot reach

The US tools above are built on US databases. For Austrian, German and EU questions the picture as of September 2026 is uneven.

  • Free primary sources are reachable. RIS (the Austrian federal legal information system), EUR-Lex, the CJEU's CURIA site and gesetze-im-internet.de are public, so an agent with web search can read the consolidated text of § 1295 ABGB or Article 28 GDPR. The risk is currency (an outdated consolidation) and invented paragraph or margin numbers, not access.
  • Commentaries are not. beck-online, juris, RDB and Linde sit behind licences a general agent cannot see, so it cites Palandt or Koziol from memory. That is how a Cologne family court came to find commentary citations "frei erfunden" in AG Köln 312 F 130/25 (2 July 2025) and warned of § 43a Abs. 3 BRAO (beck-aktuell). The Kammergericht Berlin (17 WF 144/25, 20 November 2025) rebuked counsel for citing a "fantasierende KI"'s non-existent BGH decision (heise).
  • Austria's highest criminal court has already dismissed a filing unread. In OGH 14 Os 95/25i (7 October 2025) a Graz defence lawyer's AI-drafted Nichtigkeitsbeschwerde contained invented OGH decisions; the court rejected it outright under § 285d Abs 1 StPO and pointed to § 9 Abs 1 RAO on the lawyer's personal responsibility (dataprotect.at).
  • Local grounded products exist. Noxtua, built with C.H. Beck (€80.7m raised, April 2025), JUPUS in Germany, Lexroom in Italy ($50m, May 2026), Doctrine in France and AI:ssociate in Austria (ÖRAK partner, from €39 per user per month) each ground on national content; the big platforms (Harvey, Legora) are used by Schoenherr, Hengeler Mueller, CMS and others but do not replace a citator for German-language commentary.
  • Bar rules are explicit. ÖRAK's four practice rules keep final responsibility with the lawyer and require EU servers; BRAK's December 2024 Hinweise treat tools that train on client data as a breach of § 43a Abs. 2 BRAO. The Austrian and German guidance page has the detail.

Charlotin counts 11 German, 15 Italian and 14 French decisions, small next to 1,395 in the USA, which says more about how recently continental courts started noticing than about how often it happens. The deep research tools guide compares how Claude, ChatGPT, Gemini and Perplexity handle EU sources in long-running research runs, and the Gemini Notebook review covers the closed-universe alternative for a bundle of authorities you have already collected.

What to do this week

  • Take five questions you answered last month, with the authorities you actually relied on. Run them through one grounded tool and one general model. Score each citation against the five failure types above, not just "right or wrong".
  • Write the research log template and put it in the matter file structure. A table with six columns is enough.
  • Send one line to the team: no model confirms its own citations, ever.
  • Decide which national sources your agents can and cannot reach, and write that down for associates who assume a search box sees everything.

Module 2 of the AI Lab for Lawyers course runs this routine live on real research questions, including the EU and Austrian sources a US-trained model handles worst. Read the hallucinations guide for why the errors happen, and the direct answer if a colleague asks whether the problem is solved yet. It is not, and the routine above is what closes the gap.

Frequently asked questions

Is AI legal research accurate enough to rely on in 2026?

For finding and summarising authorities, grounded tools now score above a lawyer baseline on US questions. For relying on a citation without opening it, no tool is accurate enough. HAQQ's June 2026 benchmark found 24% of answers across leading models cited or misapplied law that did not support the claim, so the reading step stays with the lawyer.

What is the difference between Westlaw AI, Lexis+ Protégé and ChatGPT for legal research?

Westlaw's CoCounsel and Lexis+ with Protégé retrieve from licensed case-law databases and check citations against KeyCite or Shepard's before answering. ChatGPT, Claude and Gemini answer from training data plus web search, with no citator, which is why they fabricate or misattribute authorities more often. Vals AI's 2025 test still found ChatGPT competitive on raw accuracy.

Which AI tool has the lowest hallucination rate for legal research?

On HAQQ's June 2026 benchmark GPT-5.5 had the lowest citation hallucination rate at 3%, but it was also the slowest and most expensive model tested, while Claude Opus 4.8 scored highest overall. Rates change with each release, so treat any published figure as a snapshot and run your own five-question test before choosing.

Can AI research Austrian, German or EU law reliably?

Only partly. Agents with web search can read free sources such as RIS, EUR-Lex and gesetze-im-internet, but cannot see paywalled commentaries on beck-online or juris, so they cite them from memory. German and Austrian courts have already dismissed filings with invented BGH and OGH decisions. Use local products or a firm knowledge base for commentary, and verify every paragraph number.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion

Playbooks

How to Verify AI Legal Research: A Citation-Checking Protocol

Verify AI legal research with a five-check protocol: open every authority, match the quote, confirm holding and status, log it. Template, time budgets, tools.

11 min read

Guides

AI Hallucinations in Legal Practice: Why They Happen and How to Catch Them

AI hallucinations in legal practice: why models invent cases, what 2,041 court decisions and the HAQQ benchmark show, and a verification routine that holds.

9 min read

Tools

Lexis+ with Protégé Review: Workflows, Vault and the Legal Intelligence Engine

Lexis+ with Protégé replaced Lexis+ AI in February 2026 with 300+ workflows, Shepard's-grounded verification and an agentic engine. What it does, costs and still gets wrong.

8 min read

Tools

CoCounsel Legal (Thomson Reuters) Review: Westlaw-Grounded Agents

CoCounsel Legal reviewed as of September 2026. The Claude-built rebuild, Brief Builder, Tabular Analysis, Deep Research Verify, $104 to $639 pricing and grounding.

8 min read

Questions

Can AI Do Legal Research Without Hallucinating? The Honest Answer

No AI legal research tool is hallucination-free as of September 2026. Measured rates, why grounding does not cure misreading, and the checks that protect you.

6 min read

Regulation & Ethics

AI Rules for Austrian and German Lawyers: ÖRAK, BRAK and CCBE Guidance

ÖRAK's four AI rules and § 40 Abs 3 RL-BA bind Austrian lawyers; BRAK's § 43a BRAO line binds German ones; CCBE guides cover both. With a bilingual checklist.

11 min read