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.
On this page
AI helps litigators most where the input is the record and the output stays inside the team: building chronologies, reviewing productions, preparing depositions and rehearsing argument against a hostile model. It hurts most where the output is a filed document containing an authority nobody opened. As of 14 September 2026, Damien Charlotin's database holds 2,041 decisions involving hallucinated AI material, 811 of them involving lawyers. The gains are real and the trap is well mapped.
The sanctions ledger
Every litigator should know a handful of these by name, because judges cite them back at counsel.
| Case | Court, date | What happened | Penalty |
|---|---|---|---|
| Mata v. Avianca | S.D.N.Y., 22 June 2023 | Six fabricated cases; lawyers asked ChatGPT to confirm its own citations | $5,000 joint sanction |
| Wadsworth v. Walmart | D. Wyo., 24 February 2025 | 8 of 9 cited cases fake; generated by Morgan & Morgan's in-house platform | $5,000 and pro hac vice revoked |
| Lacey v. State Farm | C.D. Cal., 6 May 2025 | 9 of 27 citations wrong; K&L Gates and Ellis George; CoCounsel, Westlaw Precision and Gemini all involved | $31,100 |
| Johnson v. Dunn | N.D. Ala., 23 July 2025 | Butler Snow lawyers; firm had an AI policy since 2023 | Public reprimand, removal from the case, bar referral |
| Ayinde v Haringey; Al-Haroun v QNB [2025] EWHC 1383 (Admin) | Divisional Court, 6 June 2025 | 5 fake authorities; 18 of 45 fictitious; heads of chambers and managing partners told they are responsible | £2,000 plus VAT wasted costs each; BSB and SRA referrals |
| Couvrette v. Wisnovsky | D. Or., 12 December 2025 | 15 non-existent cases and 8 fabricated quotations across three briefs; amended briefs quietly refiled | $110,204.38; case dismissed with prejudice |
| Whiting v. City of Athens | 6th Cir., 13 March 2026 | "Over two dozen" fake citations; counsel called the show-cause order void | $15,000 per attorney plus fees and double costs |
| State v. Lake | Nebraska Supreme Court, 15–16 April 2026 | 57 of 63 citations in a divorce appeal defective; AI use denied for weeks | Interim suspension |
| Sullivan & Cromwell (Prince Global Holdings) | Bankr. S.D.N.Y., April 2026 | AI-generated fake citations in an 8 April filing, caught by opposing counsel | Apology to Chief Judge Glenn; no fine |
| Cork v Smith [2026] EWHC 1199 (Ch) | Ch D, 4 June 2026 | Pinsent Masons associate; AI invented "Insolvency Rule 12.37(5)" and had flagged its own uncertainty | Referral to the SRA; firm bore extra costs |
The full list, with citations, is on the sanctions cases page. Three patterns run through it. Elite firms are not immune: Sullivan & Cromwell, K&L Gates, Butler Snow and Pinsent Masons are all in the database. Paid legal tools are not immune: Lacey involved CoCounsel and Westlaw Precision, Fletcher v. Experian (5th Cir., February 2026) traced 16 fabricated quotations to vLex and CoCounsel. And the response sets the penalty: in Gamez v. County of Fresno (E.D. Cal., April 2026) candid counsel who believed their tool had anti-hallucination protection were not sanctioned, while the Nebraska, Ninth Circuit (Lnu v. Blanche, June 2026) and Ontario (Ko v. Li) matters escalated after denial (Norton Rose Fulbright, June 2026).
Where litigators gain most
Set the filed document aside for a moment. The bulk of litigation time is spent on the record, and that is where the models are strong.
Chronology and case assessment. Give a frontier model the pleadings, the key correspondence and the witness statements and ask for a dated chronology with a source reference per entry, the gaps in the story, and the five facts each side most needs to prove. With a 1M-token context (about 555,000 words on Claude's current tokeniser) a mid-sized file fits in one pass. Quinn Emanuel's Christopher Kercher described the firm's approach as "treating Claude like a case team member" and feeding it chronological materials (Fortune, May 2026). Legora bought Wexler, a litigation fact-analysis tool, on 29 July 2026 for the same reason; Relativity's aiR for Case Strategy claims facts assembled up to 70% faster for early users. The output is a working document, never filed, and every entry carries a pointer you can click.
Deposition and transcript work. Summaries, inconsistency checks across a witness's deposition and prior statements, and outline generation from a chronology are tasks where the Vals benchmark found AI beating lawyers on transcript analysis as far back as February 2025. The deposition summaries workflow sets this out step by step. Consent and privilege questions around recording are covered by NYC Bar Formal Opinions 2025-6 and 2026-2; do not let a meeting bot into a witness preparation session without reading them.
Closed-universe research on the record. A Gemini Notebook (the renamed NotebookLM) or a Claude Project loaded with the exhibits answers "where does the contract say X" with a citation to the page, and cannot invent an exhibit that is not there. See the case-file assistant page for the setup and its limits.
Discovery: the Redgrave numbers
Generative review is now a defensible alternative to technology-assisted review, with a known trade-off. Redgrave LLP's study on a 45,000-document collection with 6% richness found Relativity aiR reached 87–88% recall against 64% for active-learning TAR, at 18.5 partner hours and $33,451 against 1,123 contractor hours and $57,405. The catch: aiR over-included, so precision was lower and second-pass volume doubled, and peers noted they had "never seen TAR 2 results this poor in real cases", so the baseline may flatter the comparison (eDiscovery Journal). The partner running it wrote 43 prompt iterations over 18 hours; the prompt is the work.
Courts are not demanding discovery about discovery. In Schulte v. LinkedIn (N.D. Cal., 30 June 2026) the court treated generative AI review as TAR and allowed search-term pre-culling without special scrutiny (Arnold & Porter). RelativityOne was renamed Relativity aiR on 16 September 2026 with agentic features included in platform pricing, and 57% of its new workspaces are now non-litigation, which tells you where the vendor thinks the growth is. The e-discovery page covers validation protocols and the precision problem in detail.
Drafting a brief without ending up in the database
The brief is the one document where the process must be rigid. This is the workflow that survives a show-cause order:
- Frame the issues and generate search strings with the model, with the instruction "Do NOT provide case citations."
- Run the searches in Westlaw, Lexis, vLex or CourtListener. Download every authority you intend to rely on.
- Feed the downloaded authorities back with "Work only with the authorities I have provided" and draft the argument section from them.
- For every citation in the draft, run the three checks: it exists (open it in the database, not in the chatbot); it supports the proposition (read the pinpointed page); its treatment is current (KeyCite or Shepard's). Log who did each check.
- Have a second person, not the drafter, check every quotation against the source. Misquotation is the fastest-growing category: 548 false-quote instances in the database as of September 2026.
- Never ask a model to verify its own citations. Raja Rajan was sanctioned a second time in April 2026 for "using AI to verify AI" and Charlotin's tracker names self-validation as an aggravating factor.
The citation-checking protocol formalises steps 4 and 5; the brief drafting page covers structure and tone. Clearbrief's Cite Check Report (December 2025) and similar tools add an audit trail, and California's 2026 guidance states the standard plainly: "No paper filed in court should contain citations that the attorney responsible for the filing has not personally read and verified."
Rehearsing oral argument against a hostile bench
This is the highest-value, lowest-risk use in litigation because nothing gets filed. The prompt:
You are a sceptical appellate panel of three judges. You have read the
attached opening brief (A), answering brief (B), reply (C) and the record
excerpts (D). Do not introduce any authority not in A–D; if you think an
authority is missing, say "the panel would want to know whether there is
authority on X" instead.
1. Identify the three weakest propositions in A and, for each, the record
cite in D or the argument in B that undermines it.
2. Draft the ten hardest questions the panel would ask appellant's counsel,
ordered from most to least damaging. For each, note the concession the
panel is fishing for.
3. Draft the best 60-second answer to each question using only A, C and D,
with pinpoints.
4. State which single concession, if made, loses the appeal.
5. Then switch sides and do 1–4 for appellee.
Be blunt. Do not soften. Do not summarise the briefs.
Run it, argue back, run it again with your answers pasted in. Opus 5 "verifies its own work without being told to" as of July 2026, so drop the old habit of adding a final verification instruction; it causes over-verification. The red-team playbook has variants for motions, mediations and negotiation.
Court orders and the judges' own AI use
The bench uses the same tools. A Northwestern and NYC Bar survey of 112 federal judges (published 27 March 2026) found 61.6% use at least one AI tool, with Westlaw AI at 38.4% and ChatGPT at 28.6%, and 45.5% had received no court-provided training (PlatinumIDS summary). Judges Neals (D.N.J.) and Wingate (S.D. Miss.) withdrew orders containing fabricated material in July 2025 and later admitted an intern used ChatGPT and a clerk used Perplexity; on 31 August 2026 a Fifth Circuit panel heard Mississippi's request to reassign Wingate's case because hallucinated citations persisted in the amended order (ABA Journal). Read the other side's brief and the court's draft order with the same suspicion you apply to your own: in Noland v. Land of the Free (Cal. Ct. App., September 2025) respondent's counsel was criticised for failing to notice the appellant's fabricated citations.
On disclosure, the map is patchy. Over 30 of 94 federal district courts had AI disclosure or certification requirements as of April 2026, copied from Judge Brantley Starr's May 2023 order; Judges Newman (S.D. Ohio) and Boyko (N.D. Ohio) prohibit generative AI except legal search engines and Microsoft products; Illinois's Supreme Court policy (1 January 2025) requires no disclosure and holds the filer accountable. The standing orders page tracks the current list; the short answer to "must I tell the court" is on the disclosure question page. Module 2 of the AI Lab for Lawyers course works through the research-and-verification loop above on a real brief, and Module 6 covers the court rules.
The one rule never to break
Read every case. Not the model's summary of the case, not the headnote, not the parenthetical: the case, at the page you are citing, in a database that cannot invent it. Every sanctioned lawyer in the ledger above owned an AI tool, most owned a paid one, and several worked at firms with written policies. What none of them did was open the source. The rest of the litigation toolkit, from chronology to discovery to the hostile bench, is yours to use aggressively, because none of it is filed without passing through that one step. Arbitration counsel face a related but different set of rules, covered on the arbitration page.
Frequently asked questions
What are the biggest AI sanctions against litigators so far?
Couvrette v. Wisnovsky (D. Or., December 2025) is the largest US penalty at $110,204.38 in fees and fines for fabricated citations and quotations, with the case dismissed. Whiting v. City of Athens (6th Cir., March 2026) fined two lawyers $15,000 each plus fees and double costs. The Nebraska Supreme Court suspended a lawyer in April 2026 after 57 of 63 citations in a brief were defective.
Do I have to disclose AI use to the court?
It depends on the court. As of April 2026 more than 30 of 94 US federal district courts had AI disclosure or certification requirements, starting with Judge Brantley Starr's May 2023 order. Illinois and New Jersey impose no disclosure duty. Two Ohio federal judges prohibit generative AI in their courtrooms except legal search engines and Microsoft products. Always check the judge's standing order.
Is generative AI document review accepted in discovery?
Yes, increasingly. In Schulte v. LinkedIn (N.D. Cal., June 2026) the court treated generative AI review like technology-assisted review and allowed search-term pre-culling without special scrutiny. Redgrave LLP's study found Relativity aiR reached 87–88% recall against 64% for active-learning TAR, at 18.5 partner hours instead of 1,123 contractor hours, though with lower precision.
Can AI help me prepare for oral argument?
Yes, and it is one of the safest uses because nothing the model produces is filed. Give the model your brief, the opposing brief and the record, and instruct it to act as a hostile bench: the five hardest questions, the weakest link in each argument, the concession the other side wants. Rehearse against that, then verify any authority it raises before you rely on it.
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
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
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
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.
10 min read
Use Cases
AI in E-Discovery: From TAR to Generative Review
Generative AI review beats TAR on recall (87–88% vs 64% in Redgrave's 2026 study) but loses precision. Tools, privilege, defensibility and a sampling plan.
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
Practice Areas
AI for Arbitration Practitioners: Submissions, Evidence and Tribunal Rules
AI for arbitration practitioners: what CIArb's 2025 guideline, tribunal procedural orders and disclosure rules mean for long records, witnesses and quantum.
11 min read