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.
On this page
To verify AI legal research, open every authority the model cites in a primary database, confirm it exists, that the quoted words appear at the pinpoint, that the holding supports your proposition, that it is still good law, and that the jurisdiction and date fit, then log each check. No AI tool performs this for you, and courts treat the duty as nondelegable.
Why "it looked right" has stopped working
The sanctions of the last twelve months share one feature: the lawyer believed a check had happened. In Couvrette v. Wisnovsky (D. Or., order 12 December 2025) three briefs contained 15 non-existent cases and 8 fabricated quotations; Magistrate Judge Mark D. Clarke dismissed the $12 million elder-abuse claim with prejudice and ordered $110,204.38 in fines and fees against the plaintiff's two lawyers, calling it the "appropriate case" for terminating sanctions (ABA Journal). In April 2026 the Alabama Supreme Court dismissed an entire appeal after a Mobile lawyer's apology for AI-invented cases cited two more non-existent cases in the same paragraph (Bloomberg Law, 24 April 2026). In Cork v Smith [2026] EWHC 1199 (Ch) a Pinsent Masons associate filed an invented "Insolvency Rule 12.37(5)" although the AI had flagged its own answer as possibly inaccurate; the firm was referred to the SRA (UK Human Rights Blog, 4 June 2026).
Two design points follow. The tool is not the problem: Fletcher v. Experian (5th Cir., 18 February 2026) involved 16 fabricated quotations from vLex and CoCounsel, and Lacey v. State Farm (C.D. Cal., 6 May 2025) cost K&L Gates and Ellis George $31,100 for citations produced with CoCounsel, Westlaw Precision and Gemini. And the model cannot be the checker: a Cherry Hill lawyer was reportedly sanctioned a second time on 27 April 2026, $5,000, after "using AI to verify AI", and Charlotin's database treats self-validation as an aggravator. Why models invent authorities is covered in AI hallucinations in legal practice; this page is about what to do with the output.
Charlotin's tracker listed 2,041 decisions on 14 September 2026, 811 involving lawyers: 1,690 instances of fabricated case law, 854 of misrepresented citations and 548 false quotations (AI Hallucination Cases). The middle category, a real case cited for something it does not say, is the one an existence check misses entirely.
The five checks and what each one catches
Reviews of the Q1 2026 sanctions describe verification as three checks, existence, support and treatment, and quote courts calling it a "nondelegable duty" that AI "in no way abrogates" (David Kemp's Q1 2026 sanctions review). Five is safer, because quotations and jurisdiction fail independently of the other three.
| Check | The question | Where you look | What it catches | Recent failure |
|---|---|---|---|---|
| 1. Existence | Does this name and citation exist? | Westlaw, Lexis, CourtListener, BAILII, RIS, juris, EUR-Lex, the court's own site | Fabricated case law (1,690 instances in Charlotin) | Couvrette, 15 invented cases; OGH 14 Os 95/25i (7 October 2025), invented Austrian Supreme Court decisions |
| 2. Quotation | Are the quoted words verbatim at the pinpoint? | Full-text search inside the opinion | False quotations (548) | Fletcher, 16 fabricated quotations from paid tools |
| 3. Support | Does the holding, on these facts and in this posture, carry your proposition? | The passage and the paragraphs around it | Misrepresented authority (854); HAQQ found 24% of 3,000 answers misapplied real law | Cork v Smith, real rule number, invented content |
| 4. Status | Reversed, overruled, distinguished, superseded by statute? | KeyCite, Shepard's, ICLR or BAILII citator, RIS decision history | Stale law presented as current | Anything decided after the model's training cutoff |
| 5. Fit | Right jurisdiction, court level, date and statute version? | Opinion header; statute version in force at the relevant date | Wrong-forum authority; Vals found every tool lost 11–14 points on multi-jurisdiction questions | Cross-border memos |
In Johnson v. Dunn (N.D. Ala., 23 July 2025) the court "ran independent searches for each allegedly fabricated citation, to no avail" before reprimanding three Butler Snow lawyers; the firm had an AI policy dating from June 2023. Policy is not protocol.
The protocol, step by step
- Freeze the draft and list every authority: cases, statutes, regulations, secondary sources and URLs. Include the ones you are sure of; fabrications often reuse real party names with invented citations.
- Name a checker who is not the author. On a solo desk, check in a separate sitting with the draft closed, so each authority is read cold.
- Search the database twice per authority: by citation and by party names. A miss on both is a fail; a hit on the names with a different citation is a fail pending re-check of the correct case.
- Paste every quoted passage into the opinion's full-text search. Verbatim or fail. Record the paragraph or page.
- Read the passage and its neighbours. Note the posture, holding or dictum, majority or dissent. Confirm it carries the proposition in your draft, not a neighbouring one.
- Run the citator: KeyCite or Shepard's in the US, the ICLR or Westlaw UK citator in England and Wales, RIS decision history in Austria, juris or beck-online in Germany, Curia for CJEU authority.
- Confirm jurisdiction, court level and date, and for statutes the version in force at the relevant time.
- Record the result in the verification table with initials, database and time.
- Treat a fail as a strike. Never ask the model for a replacement; research the point afresh and, if nothing supports it, change the argument.
- The signing lawyer reads the table before filing. California's Practical Guidance, updated 14 May 2026, puts it plainly: no paper filed in court should contain citations that the responsible attorney has not personally read and verified.
The Divisional Court in Ayinde v Haringey [2025] EWHC 1383 (Admin) held that freely available generative AI tools "are not capable of conducting reliable legal research" and addressed heads of chambers and managing partners directly (judgment, 6 June 2025). Step 10 is how a supervising lawyer discharges that responsibility; the background is in the Ayinde explainer.
Time budgets
Verification costs about what the research would have cost without AI, because reading is the slow part and AI does not remove it. The figures are a working budget, not a study.
| Step | Per authority |
|---|---|
| 1. Existence | 1–2 min (two searches) |
| 2. Quotation | 2–3 min (full-text search) |
| 3. Support | 5–10 min (reading in context; the irreducible core) |
| 4. Status | 1–2 min (one citator run) |
| 5. Fit | 1 min |
| Total | 10–18 min; a 15-authority brief is three to four hours |
The lawyers in the Vals research study of 2025 averaged about 23 minutes per question without AI. Oregon courts have priced fabrications at $500 per fake citation and $1,000 per fake quotation (ComplexDiscovery, 6 April 2026). Fifteen minutes an authority is cheap, and it is billable: ABA Formal Opinion 512 permits billing actual time spent.
The verification table
One table per filing or memo, saved with the matter. It is the audit trail if a court asks questions and the training record if a regulator asks about AI literacy.
| # | Authority as cited | Proposition | 1. Exists (database, date) | 2. Quote verbatim (pinpoint) | 3. Supports | 4. Status | 5. Fit | Result | Checker, time |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Ayinde v Haringey [2025] EWHC 1383 (Admin) | Hamid jurisdiction where fake authorities are cited | Yes (BAILII, 19 Sep 2026) | Yes, sanctions paragraph | Holding; Divisional Court; on point | No negative treatment | England and Wales; High Court; June 2025 | PASS | AB, 09:40 |
| 2 | "Insolvency Rule 12.37(5)" (the provision filed in Cork v Smith) | Power to extend time | Not found as cited (legislation.gov.uk, 19 Sep 2026) | n/a | n/a | n/a | n/a | FAIL: strike, research afresh | AB, 09:52 |
| 3 | [authority as given by the model] |
A cell is "Yes" only if the checker opened the document, and a row is PASS only if all five cells are filled. Keep a research log beside the table: tool and model, date, the prompt, the raw output saved unedited. Courts price candour. In Gamez v. County of Fresno (E.D. Cal., April 2026) no sanction followed because counsel was candid; in Whiting v. City of Athens (6th Cir., 13 March 2026) two lawyers who refused to answer the court's seven questions and called the show-cause order "void" were fined $15,000 each plus fees and double costs. The wider case list is in AI hallucination sanctions.
Where the tools fit, and where they stop
Several products automate parts of the protocol. Used correctly they shorten checks 1 and 4; used instead of checks 2, 3 and 5 they reproduce the Lacey and Fletcher problem with a better-looking audit trail.
| Tool (as of September 2026) | What it does | Checks it helps | What it does not do |
|---|---|---|---|
| Thomson Reuters CoCounsel Legal (US GA 20 August 2026) with Deep Research Verify and Westlaw Quick Check | Runs citations against Westlaw and KeyCite; flags bad law in an uploaded brief | 1, 4 | Read the passage for support; catch a real case cited for the wrong point |
| Lexis+ with Protégé (US GA 24 February 2026) | Shepard's-grounded verification inside drafting workflows | 1, 4 | Confirm quotations against the opinion text |
| Clearbrief Cite Check Report (16 December 2025) | Audit-trail report for partners; ranked highest, 40.5 of 50, by the Nevada State Bar AI Work Group according to BriefCatch (LawSites) | 1, 2 in part, 4 | Replace the reader on check 3 |
| Claude for Legal plugins (open source, Apache 2.0) | Tag any citation not checked against a connected research tool such as CourtListener or Trellis as [verify] | Flags what to check | Verify anything without the connector |
| UPenn's open-source URL checker | Confirms cited URLs resolve | 1, for web sources | Confirm the page says what the model claims |
The limits are measured, not theoretical. Stanford's RegLab found Lexis+ AI hallucinating 17% of the time and Westlaw AI-Assisted Research 33% in 2024. HAQQ's June 2026 benchmark of 3,000 answers put the lowest citation hallucination rate at 3% (GPT-5.5), the highest at 64% (Mistral Large), and found 24% of all answers citing or misapplying law that did not support the claim (HAQQ). A University of Pennsylvania study of 3 April 2026 found deep-research agents fabricating 3–13% of URLs, more than plain models because they cite more (arXiv). Grounding lowers the rate; a court needs zero. Tool detail is in AI legal research in 2026 and the reviews of CoCounsel Legal and Lexis+ with Protégé.
When a citation fails
| Failure | Action | Never |
|---|---|---|
| Not found under citation or party names | Strike it; research the proposition in the database; if nothing supports it, the argument changes | Ask the model for "a real case that says the same thing" |
| Found, quotation not verbatim | Replace with the exact words and pinpoint, or paraphrase without quotation marks and re-run check 3 | Leave quotation marks around words the court did not use |
| Found, does not support the proposition | Strike it; this is the misrepresented-authority category, 854 instances in Charlotin and the hardest to spot | Shade the proposition until it fits |
| Negative treatment | Strike, or cite with the treatment disclosed if the point survives | Cite it clean |
| Already filed | Write to the court and opposing counsel at once, explain what happened, file corrected papers | Blame a paralegal, deny AI use, or "correct" with new AI output |
The last row has its own case law. In Ko v. Li (Ontario, 2025) the lawyer's false account of who drafted the factum, not the ChatGPT errors, produced a criminal-contempt referral, and the Nebraska Supreme Court's interim suspension of an Omaha lawyer in April 2026 followed weeks of denying AI use in a brief where 57 of 63 citations were defective. Check the other side's authorities with the same table: in Noland v. Land of the Free (Cal. Ct. App., 12 September 2025) respondent's counsel was criticised for missing the appellant's fabricated citations (reported by LawSites), and Sullivan & Cromwell's April 2026 apology to a bankruptcy judge followed errors that Boies Schiller, opposing counsel, had found.
A prompt that keeps the model out of the citation business
Fewer unverified citations means less verification. The pattern now standard in legal prompting guides is to take issue-framing and search strings from the model, run the search yourself, and feed verified authorities back (The Legal Prompts, 18 February 2026).
Stage 1 — issue framing (before any database search)
You are assisting a [jurisdiction] litigator. Do NOT provide case citations,
statute section numbers or quotations from any source. If you believe an
authority exists, describe it in words ("a 2019 Court of Appeal decision on
limitation in professional negligence") and mark it [TO SEARCH].
Task: identify the legal issues raised by the facts below, the elements of
each claim or defence, and the most useful search strings for [Westlaw /
Lexis / BAILII / RIS], with Boolean connectors.
Facts: [paste]
Output: numbered issues; for each, elements, factual gaps, three search
strings, and any [TO SEARCH] authorities described in words.
Stage 2 — drafting (after the verification table is complete)
Work only with the authorities listed below, which have been read and
verified. Cite them by the short names given. Do not add any authority,
quotation or section number that is not in this list. If a proposition
needs support that the list does not provide, write [NO AUTHORITY IN LIST]
and continue.
Verified authorities:
1. [short name] — [citation] — [proposition it supports] — [pinpoint]
2. ...
Task: [draft the section].
Two notes on current models. Claude Opus 5, released 24 July 2026, "verifies its own work without being told to" according to Anthropic's release notes, so a closing instruction to "double-check all citations" now produces a paragraph of reassurance rather than a check; the model has no database to check against either way. And no prompt makes a general model a research tool: the Divisional Court's finding in Ayinde stands, and whether AI can research without hallucinating has a short answer. The right place to hold the Stage 2 list across a matter is a project workspace, described in setting up Claude Projects for a matter.
What to do this week
- Copy the verification table into the matter template and add "verification table attached" to the filing checklist.
- Run the protocol retrospectively on the last brief or memo that used AI. The result is your baseline error rate.
- Set the rule that no authority enters a draft until its row reads PASS, and that the model is never asked to check itself.
- Decide who checks: a second lawyer where the team allows it, a separate cold sitting where it does not.
- Check whether any court you appear before has an AI certification requirement; the list is in court standing orders on AI disclosure.
- Add the protocol to the firm's AI policy as an annex. Butler Snow, Morgan & Morgan and Gordon Rees all had policies when they were sanctioned; none contained a table.
Module 2 of the AI Lab for Lawyers course runs this protocol live on a real research task, including what the failed rows look like. Judge Rankin's line from Wadsworth v. Walmart (D. Wyo., February 2025) remains the shortest statement of the standard: a fake opinion is not existing law.
Frequently asked questions
How do I check whether an AI-generated case citation is real?
Open a primary database such as Westlaw, Lexis, CourtListener, BAILII or Austria's RIS and search by the citation and separately by the party names. If neither search returns the case, it does not exist. Do not accept a Google snippet, a secondary summary or the AI's own assurance. Then read the opinion to confirm it says what the draft claims.
Can I ask ChatGPT or Claude to verify its own citations?
No. The model has no database to check against, so it will answer with the same pattern-matching that produced the citation. In April 2026 a New Jersey lawyer was reportedly sanctioned for a second time after using AI to verify AI, and Damien Charlotin's tracker treats asking a model to self-validate as an aggravating factor. Verification happens in a research database, by you.
How long does it take to verify AI legal research properly?
As a working budget, allow 10 to 18 minutes per authority for all five checks, most of it spent reading the relevant passage for support. A brief with 15 authorities is therefore three to four hours of checking. Tools that automate the existence and citator steps can cut that by a third, but the reading cannot be delegated.
Do paid legal research tools remove the need to verify?
No. Stanford's RegLab measured hallucination rates of 17% for Lexis+ AI and 33% for Westlaw AI-Assisted Research in 2024, and 2026 sanctions have involved CoCounsel, vLex and in-house firm platforms. Grounded tools cut the rate and shorten the checks; the duty to open the source and read it stays with the lawyer who signs the filing.
What should I do if I find a fake citation after filing?
Correct it immediately, in writing, and explain what happened. Courts consistently treat candour as mitigating and evasion as aggravating. In Gamez v. County of Fresno no sanction was imposed because counsel was candid, while in Whiting v. City of Athens two lawyers who refused to answer the court's questions were fined $15,000 each plus fees and double costs.
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.
Related reading
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
Regulation & Ethics
AI Hallucination Sanctions: The Cases Every Lawyer Should Know (2023–2026)
AI hallucination cases reached 2,041 in Charlotin's database by September 2026, 811 involving lawyers. Every sanction from Mata to Couvrette's $110,204, with lessons.
12 min read
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.
11 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
Regulation & Ethics
Ayinde v Haringey Explained: The UK's Leading AI Hallucination Judgment
Ayinde v Haringey [2025] EWHC 1383 (Admin) explained: the fake cases, the Divisional Court's sanctions ladder, the duty on heads of chambers and how the SRA built on it.
11 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