AI Agents for Lawyers

Use Cases

Drafting Briefs and Motions with AI Without Ending Up in Charlotin's Database

AI brief drafting is safe only as a five-stage workflow: grounded research, outline, drafting from verified authorities, every case read, court rules checked.

By Dr. Niklas SchmidtUpdated 13 November 202610 min read
On this page
  1. The database, and how lawyers get into it
  2. Stage 1: research, and the authority you never generate
  3. Stage 2: outline and argument with grounded drafting tools
  4. Stage 3: verification, or Couvrette, Whiting and the 57 of 63
  5. Stage 4: the court's own rules
  6. Stage 5: candour when something gets through
  7. Before your next filing

AI can draft a brief or motion faster than any associate, and as of September 2026 it has also put more than 800 lawyers into a public database of court decisions about fabricated authority. The difference between the two outcomes is a five-stage workflow: grounded research, an outline, drafting only from verified authorities, a lawyer reading every citation, and a check of the court's own AI rules before filing.

The database, and how lawyers get into it

Damien Charlotin's AI Hallucination Cases database listed 2,041 decisions worldwide on 14 September 2026: 1,395 in the United States, 217 in Canada, 111 in Australia, 69 in the United Kingdom. Pro se litigants account for 1,174; lawyers for 811; judges for 32. The counts went from roughly 200 in mid-2025 to 712 by the end of 2025, 1,598 by 9 June 2026 and 2,041 now, or about eight new decisions a day (Charlotin). US courts imposed over $145,000 in penalties in the first quarter of 2026 alone (ComplexDiscovery).

Read enough of the entries and a pattern appears. Nobody was sanctioned for using AI. They were sanctioned because a citation the model produced was never checked, and then, in the worst cases, because they denied it, refiled quietly or asked the model to confirm its own work. Norton Rose Fulbright's June 2026 review of the case law puts it plainly: courts hold that existing rules suffice (FRCP 11, FRAP 38, 28 U.S.C. § 1927, inherent power), and "candour mitigates, evasiveness aggravates" (Norton Rose Fulbright). The sanctions cases page catalogues them; this page uses five of them as the structure for a workflow.

Stage 1: research, and the authority you never generate

The first failure happens before drafting starts, when the model is asked to find the law. In Cork v Smith [2026] EWHC 1199 (Ch), a Pinsent Masons associate asked an AI to research insolvency law and it produced "Insolvency Rule 12.37(5)", which does not exist. The tool had flagged that its answer might be inaccurate. The judge criticised a "cavalier attitude", the firm was referred to the SRA and agreed to bear the extra costs.

The control is a rule, not a tool: authorities enter the brief only through a citator-backed database. Use a general model for issue framing and search strings with citations forbidden, run the search in Westlaw, Lexis, Clio Work or your national database, and export the results with pincites. The AI legal research page sets out that routine and the tools behind it; the output of Stage 1 is a verified authorities list, and nothing else goes into Stage 3.

Stage 2: outline and argument with grounded drafting tools

Once the authorities are fixed, the model earns its keep. The tools that ship with a citator behind them as of September 2026:

ToolWhat it does for a briefGroundingDate
Westlaw Brief Builder (in CoCounsel Legal)Builds argument sections from Westlaw research, KeyCite-checked; Deep Research VerifyWestlaw, Practical LawNext-generation CoCounsel GA in the US 20 Aug 2026 (Thomson Reuters)
Lexis+ with Protégé, Agentic DraftingDrafts from Lexis content with Shepard's Verify trust markers; Workrooms with dual approvalsLexisMay 2026; Legal Intelligence Engine 24 Aug 2026
Claude for Legal, litigation pluginBrief sections, chronology, claim charts; every citation not checked against a connected research tool is tagged [verify]Whatever connector is attached: CourtListener, Trellis, CoCounsel, MidpageRepo state 18 Sep 2026 (GitHub)
Astra for LawDrafting on a 230M-URL index of US lawCourtListener/Free Law Project index, no citator17 Sep 2026; days old, few independent tests
General model (Claude, ChatGPT, Gemini) with the verified list attachedOutline, argument structure, counter-arguments, plain-English rewriteOnly what you attachAny

Speed is real. Katy Young of Ad Astra Law Group, a six-lawyer San Francisco firm, told Above the Law in February 2026 that a 45-page complaint that used to take two days now takes her two and a half hours (Above the Law). The CoCounsel review and Protégé review cover the two incumbents' drafting agents in detail.

The drafting prompt for a general model, whichever surface you use:

You are drafting the argument section of a [motion type] in [court], for the [moving/opposing] party. Attached: (1) the verified authorities list with pincites, (2) the statement of facts, (3) the outline.

Rules:
- Cite ONLY authorities from the verified list, using the pincites given. If a proposition needs support and no listed authority provides it, write [no source] and stop; do not supply a case, statute or quotation from memory.
- Quote only text that appears verbatim in the attached extracts. Paraphrase everything else.
- Follow the outline order. One heading per outline point.
- Keep each paragraph under 120 words. No adverbs in the first sentence of a paragraph.
- After the draft, list every citation used with the outline point it supports, so I can check each one.

Then run the adversarial pass before anyone verifies anything: "Act as opposing counsel. Give the five strongest attacks on this draft, and for each say which paragraph it targets and whether the weakness is factual, legal or structural." The red-teaming playbook has the full set of prompts. Note that Claude Opus 5 (24 July 2026) "verifies its own work without being told to", so prompts written in 2025 that add "then verify your citations" now produce over-verification, and in any event that self-check is not your verification.

Stage 3: verification, or Couvrette, Whiting and the 57 of 63

Three decisions define the standard.

Couvrette v. Wisnovsky (D. Or., order 12 Dec 2025). Three briefs in a $12 million family-winery elder-abuse suit contained 15 non-existent cases and 8 fabricated quotations. Magistrate Judge Mark D. Clarke dismissed the case with prejudice, called the conduct "a notorious outlier in both degree and volume", and ordered $110,204.38 in total: about $80,000 of the defendants' fees plus fines of about $15,000 and $14,000 on the two lawyers (ABA Journal). The largest US sanction to date, and the aggravating fact was that amended briefs had been filed quietly rather than the error confessed.

Whiting v. City of Athens (6th Cir., 13 Mar 2026). The first federal appellate fines: $15,000 per attorney plus appellees' fees and double costs, for "over two dozen" bad citations. The lawyers had refused to answer the court's seven questions and called the show-cause order "void".

Nebraska Supreme Court (15 Apr 2026). Interim suspension of an Omaha lawyer whose divorce-appeal brief had 57 of 63 citations defective; he denied using AI for weeks before admitting it (WOWT).

The rule that would have prevented all three is one sentence from California's 2026 practical guidance: "No paper filed in court should contain citations that the attorney responsible for the filing has not personally read and verified" (Wilson Elser). The citation-checking protocol turns it into steps; the artefact that proves it happened is this table, attached to the draft that goes to the signing lawyer:

Citation as it appearsExists (database, date)Pincite supports the proposition (Y/N; paragraph read)Quotation verbatim (Y/N/none)Treatment (KeyCite/Shepard's flag)Checked byDate

Tools can populate the first and fifth columns. Clearbrief's Cite Check Report (launched 16 December 2025) is designed as an audit trail for partners and was ranked highest, 40.5 out of 50, by the Nevada State Bar's AI work group according to BriefCatch (LawSites); Westlaw Quick Check and BriefCatch's validation do similar work. None of them fills the third column. That is the lawyer reading the paragraph, and it is the column that catches the misgrounded citation, a real case cited for something it does not hold, which Charlotin logs 854 times.

Two cautions. Fletcher v. Experian (5th Cir., Feb 2026) fined counsel $2,500 for 16 fabricated quotations that came from paid legal tools, vLex and CoCounsel, and the court remarked the problem "shows no sign of abating"; grounding lowers the rate, it does not remove the reading. And a lawyer who asks the model whether its citations are real has not verified anything, as the second sanction on a New Jersey lawyer in April 2026 for "using AI to verify AI" confirms.

Stage 4: the court's own rules

Before filing, three checks that differ by forum:

  • US federal courts. As of April 2026 more than 30 of the 94 district courts had AI disclosure or certification requirements, starting with Judge Brantley Starr's order in the Northern District of Texas in May 2023. Separately, over 300 individual federal judges have standing orders; the two numbers count different things. Judges Newman (S.D. Ohio) and Boyko (N.D. Ohio) prohibit generative AI outright except for legal search engines and Microsoft products, an exception that quietly permits Copilot (AI court disclosure map). Illinois requires no disclosure under its Supreme Court policy of 1 January 2025, and the Seventh Circuit is described as the anti-disclosure circuit. The standing orders page tracks the list; the direct answer is for the colleague who asks at 5pm.
  • England and Wales. Ayinde v Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin) set out the ladder from public admonition through wasted costs, contempt and referral to the police, and addressed heads of chambers and managing partners directly. The SRA's warning notice of 17 August 2026 reported 42 AI-misuse reports in a year, and a UK tracker counts 76 cases. In SRA v Kumar the Solicitors Disciplinary Tribunal struck off a registered foreign lawyer in August 2026 for using fabricated authorities in his own defence and then "correcting" them with more (Legal Futures). The Ayinde explainer has the judgment in full.
  • Austria and Germany. The Austrian OGH dismissed an AI-drafted Nichtigkeitsbeschwerde without reaching the merits (14 Os 95/25i, 7 Oct 2025) because it cited invented OGH decisions, pointing to the lawyer's personal responsibility under § 9 Abs 1 RAO. AG Köln (312 F 130/25) treated invented BGH citations as a potential breach of § 43a Abs. 3 BRAO. There is no disclosure duty to the court, but the sanction for a bad citation is that the filing fails.

Stage 5: candour when something gets through

The last stage is the one nobody plans for. Four 2026 outcomes show how much the response matters:

  • Gamez v. County of Fresno (E.D. Cal., Apr 2026): no sanction, because counsel was candid immediately and had believed the tool carried "anti-hallucination" protection.
  • Sullivan & Cromwell (Apr 2026): a partner apologised to Bankruptcy Judge Martin Glenn (S.D.N.Y.) for fabricated citations in an 8 April filing flagged by opposing counsel; the firm's review "did not identify the inaccurate citations generated by AI"; no fine (Bloomberg Law).
  • Alabama Supreme Court (24 Apr 2026): the appeal was dismissed outright, $17,200 in fees ordered, and the lawyer may not file in that court without a co-signing attorney, after his apology for hallucinated cases cited two more non-existent cases in the same paragraph (Bloomberg Law).
  • New Mexico Supreme Court (Aug–Sep 2026): $5,000 and a bar referral for a murder-appeal brief containing witnesses ChatGPT had invented. Justice C. Shannon Bacon to counsel: "Do you watch the news? Do you listen to the radio?" The lawyer said he had expected "a bulletproof summary" (Reuters).

The protocol: as soon as a bad citation is found, file a notice of errata with a corrected brief, state that AI was involved if it was, explain the verification failure, and offer to bear any costs the error caused. Do not amend quietly, do not blame staff, and do not wait for the show-cause order. Three of the four outcomes above turned on exactly that.

Before your next filing

  • Put the verified authorities list and the seven-column table into the matter template so that every brief starts with them.
  • Save the drafting prompt above with the "[no source]" rule, and tell associates that any citation the model adds on its own is unverified by definition.
  • Look up the judge's standing order and the court's local rule on AI before the first draft, not the night before filing.
  • Decide now who signs the certification. FRCP 11 makes it personal.
  • Write the errata protocol into the litigation manual while nobody needs it.

The litigators' page places this workflow next to discovery review and deposition work, and Mata v. Avianca, three years on explains why the first sanction in 2023 still describes most of the 2026 cases. Module 6 of the AI Lab for Lawyers course walks through these sanction decisions and the filing controls that follow from them, so that the five stages become habit rather than a checklist read once.

Frequently asked questions

Can I use AI to draft a legal brief or motion?

Yes, and many litigators do, but only as part of a workflow where the authorities come from a citator-backed database, the model drafts from that verified list, and the signing lawyer reads every cited case at the pincite. Used as a research tool that invents its own citations, AI drafting has produced more than 800 court decisions sanctioning or admonishing lawyers worldwide.

What is Charlotin's database of AI hallucination cases?

Damien Charlotin, a research fellow at HEC Paris, maintains a public database of court decisions in which a party relied on hallucinated AI material. On 14 September 2026 it listed 2,041 decisions worldwide, 1,395 of them in the United States, with 811 involving lawyers and 32 involving judges. It is an undercount, because only decisions that find or clearly imply hallucination are included.

Do I have to tell the court I used AI to draft a motion?

It depends on the court. As of April 2026 more than 30 of the 94 US federal district courts had AI disclosure or certification requirements, and over 300 individual federal judges had standing orders, while Illinois requires no disclosure and the Seventh Circuit is sceptical of it. Check the judge's page before filing, and assume the answer differs from the last court you were in.

What happens if my AI-drafted brief contains a fake citation?

Sanctions in 2026 range from public reprimand to six-figure fee awards, dismissal of the client's appeal and interim suspension, under existing rules such as FRCP 11 and the court's inherent power. Courts consistently penalise denial and delay more than the error itself; lawyers who disclosed promptly have escaped sanction, while those who called a show-cause order void paid $15,000 each.

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

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

Regulation & Ethics

Court Standing Orders on AI: Where You Must Disclose or Certify

Court AI disclosure rules in 2026: over 30 of 94 US federal districts and 300+ judges require certification, two Ohio judges ban it, Illinois requires none.

10 min read

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

Playbooks

Red-Team Your Argument with AI: Adversarial Prompts for Litigators and Negotiators

Red-team your legal argument with AI in five rounds: prompts that make the model argue the other side, play a sceptical judge and score your case before filing.

12 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

Practice Areas

AI for Litigators: Research, Drafting, Discovery and the Sanctions Trap

AI for litigators pays off in chronologies, discovery and deposition prep; it costs $110,204 when a brief cites invented cases. Where to gain and the one rule.

8 min read