AI Agents for Lawyers

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.

By Dr. Niklas SchmidtUpdated 13 November 202612 min read
On this page
  1. Why a model is a useful opponent and a poor judge
  2. The five-round protocol
  3. Round 1: the opposing-counsel prompt
  4. Rounds 2 and 3: the sceptical bench and the cross-examination
  5. Two models against each other
  6. A worked example: the late-notice defence
  7. Rules that keep red-teaming out of Charlotin's database
  8. What to do this week

Red-teaming a legal argument with AI means making a model attack your position before opposing counsel, a judge or the other side of the table does. Done in five rounds, with attack and defence in separate sessions, no citations allowed and a second model as referee, it surfaces the weak premise or missing fact you have stopped seeing. The prompts and a worked example are below.

Why a model is a useful opponent and a poor judge

Frontier models are good at one half of adversarial work. Asked for the five strongest arguments against a motion they produce five, ranked and without sparing your feelings; the pattern is standard enough that prompt guides for lawyers list "adopt the role of opposing counsel" alongside role prompting and IRAC-style reasoning (The Legal Prompts, 2026). The other half, deciding which attack matters, they do badly, for three reasons.

First, they follow the framing. A model shown "my strong argument" will find it strong; shown the same argument as "the position under attack" it will find the holes. Guides on legal prompting report that structured prompts beat conversational ones and positive instructions beat prohibitions (LegalFly, 17 Sep 2026); the same mechanism means the model inherits whatever confidence you put in.

Second, you will over-trust it. Scientific American's May 2026 analysis of why lawyers keep filing fabricated cases blames "cognitive surrender", and reports a Nature study from February 2026 in which people who liked technology performed worse when told advice came from AI (Scientific American, May 2026). The bias that puts phantom citations into briefs will also make a confident reassurance from the model feel like a second opinion.

Third, its legal premises are unreliable even when its logic is sharp. In the HAQQ benchmark of 5 June 2026, 24% of answers across 300 tasks cited or misapplied law that did not support the claim. A red-team round that announces "the clause is unenforceable as an unfair term" has produced a lead, not a finding.

So the protocol keeps the model in the role it is good at: it attacks, you judge, a second model referees, and every authority comes from your research log. Quinn Emanuel's Christopher Kercher told Fortune that the firm's approach was "treating Claude like a case team member" (Fortune, 12 May 2026); the version here is a case team member who is never allowed to cite a case.

The five-round protocol

Each round is a fresh chat, so nothing the model said as your opponent leaks into what it says as your junior. Budget ninety minutes for a motion-sized argument and an afternoon for a trial theory.

RoundThe model playsYou give itIt returnsYou do
1 AttackOpposing counselThe facts, the position stated neutrally, the governing rules in your own wordsFive ranked attacks, each naming the fact or premise it targets, plus the single fact that would end the argumentMark each attack: answerable, needs evidence, or fatal
2 BenchA sceptical judge or arbitratorThe Round 1 brief plus its outputSix questions the bench would put to you, with the answer it expects and the answer it fearsDraft your answers; list the ones you cannot yet give
3 WitnessCross-examining counselThe witness statement or chronologyTwenty questions in sequence, each with the admission it seeksPrepare the witness; pull every document the questions rely on
4 DefenceYour own junior, fresh session, ideally a second modelThe Round 1 attacks only, not your marksA rebuttal to each attack with the evidence it needsCompare with your own marks; note where the model is more optimistic than you
5 ScoreA referee, second modelAttacks and rebuttals side by sideA score out of ten per point with "why not higher"Decide what to fix, what to concede and what to settle

Rounds 1 and 4 are the pair that matters; the rest refine them. The prompting patterns behind all five rounds, role, structure and three-turn iteration, are covered in Module 2 of the AI Lab for Lawyers course, and the prompt engineering guide explains why each part of the prompt below sits where it does.

Round 1: the opposing-counsel prompt

The prompt follows the standard five-part structure of role, task, sources, format and constraints. Two details do most of the work: the position is described without the words "my client", and citations are banned outright.

ROLE
You are lead counsel for the claimant in a commercial dispute. You are experienced,
unsentimental and paid to win. You have no loyalty to the position described below;
your job is to dismantle it.

TASK
Produce the five strongest attacks on the position under attack. Rank them from
most to least dangerous. For each attack state: (a) the attack in two sentences;
(b) the specific fact, document or premise it targets; (c) what evidence or
admission would make it succeed; (d) what the defending side would say in reply,
and why that reply is weak.
Then state the single fact which, if established, ends the position entirely.

SOURCES
Work only from the facts and rules below. Do NOT provide case citations, statute
section numbers or quotations from authorities. Where an attack depends on a legal
proposition, state the proposition in plain words and flag it as
[PROPOSITION TO VERIFY].

FACTS
[Paste the chronology. Dates, documents, who said what to whom. No characterisation.]

POSITION UNDER ATTACK
[State it neutrally, e.g. "The claim is barred because the buyer did not give
written notice within 30 days of discovering the defect, as clause 14.2 requires."]

GOVERNING RULES, IN OUR WORDS
[Two to five sentences, e.g. "Clause 14.2 requires written notice of any claim
within 30 days of discovery. The contract is governed by English law. Clause 14.2
is not expressed as a condition precedent."]

FORMAT
Numbered list, 1 to 5, then a final heading "The fact that ends it". Plain prose,
no bullet fragments. Maximum 600 words.

CONSTRAINTS
Do not soften any attack. Do not write "arguably" or "it could be said". Do not
summarise the position back to me. If the facts are insufficient for an attack,
say which fact is missing rather than inventing one.

Three notes. Paste the chronology, not your skeleton argument; a skeleton carries your framing, and the model will attack the framing instead of the facts. Give the rules in your own words so the model has nothing to hallucinate against. And keep the "single fact that ends it" line; it produces the most useful sentence of the exercise more often than seems reasonable.

Rounds 2 and 3: the sceptical bench and the cross-examination

Round 2 swaps the role and task lines and appends the Round 1 output. It asks the judge for questions, not a ruling, because a model's ruling is worthless and its questions are not.

ROLE
You are the judge hearing this application. You have read the papers once, you are
short of time and you are irritated by advocacy that avoids the difficult point.

TASK
List the six questions you would put to counsel defending the position under
attack, in the order you would ask them. For each: the question as you would say
it from the bench; the answer counsel hopes to give; the answer you suspect is
true. Do not decide the application. No citations.

Round 3 works on the record rather than the argument. Give it the witness statement or chronology and ask for twenty questions in sequence, each paired with the admission it seeks and the document it rests on. A grounded tool helps here. Gemini Notebook, renamed from NotebookLM on 16 July 2026, answers only from the sources you load, and Ernest Svenson's case-file workflow is a ready-made Round 3: load everything, ask for a one-page issues brief tied to passages, generate a chronology, then build cross-examination outlines (Attorney at Work, 21 May 2026). His limits apply: 50 sources per notebook free and 300 paid, no open-web research, Workspace or enterprise accounts only. The Gemini Notebook page covers the setup and the transcript summaries workflow does the same for deposition text.

Whatever produces the questions, tabulate them before the witness sees them:

QuestionAdmission soughtDocument it rests onIf the witness disagrees
"You received Mr Haddad's email of 3 March, didn't you?"Receipt of the first complaint on day 12Email, bundle tab 14Server logs; the reply sent from the same inbox on 4 March
"Nobody at your company wrote to reserve its position until 21 April?"No reservation of rights for seven weeksAbsence in the correspondence bundleSearch the service team's ticketing system before trial, not after
"The laboratory report you rely on was commissioned on day 20, after the flaking was already being discussed internally?"Discovery preceded the reportPurchase order for the test, bundle tab 22The engineer's calendar and the internal thread that prompted the order

For negotiators the bench becomes a counterparty. Replace the Round 2 role with the other side's general counsel, tell the model what that person's board has been promised, and ask for the six concessions it would demand and the walk-away point it would hide. The arbitration page has the tribunal version, with questions from three arbitrators of different legal traditions.

Two models against each other

One model attacking and defending is talking to itself. Two models is closer to a moot. As of September 2026 the practical pairings are Claude Opus 5 (1M context, $5 per million input tokens, released 24 July 2026), GPT-6 Astra (public to paid ChatGPT users since 4 September 2026, about 1.05M context, $10 per million) and Gemini 3.1 Pro (1M context, $2 per million, the cheapest for bulk attack rounds). The Vals Index of 15 September 2026 ranks Claude Fable 5.1 first, Opus 5 second and GPT-6 Astra third across 58 models, a gap too small to matter for this exercise.

StepModel AModel BWhat you learn
Same Round 1 prompt to bothAttacksAttacksAttacks both find are almost certainly real; attacks only one finds go on your list to judge yourself
Round 4 defenceDefends against B's attacksDefends against A's attacksEach model rebuts an argument it did not write, which removes the self-agreement problem
Round 5 refereeScores B's attack against A's defenceScores A's attack against B's defenceTwo independent scores per point; a gap of more than three points means re-reading the facts

Two cautions. GPT-6 Astra's "recurrent depth" architecture obscures some of its reasoning, so ask it for its reasons explicitly. And a consumer plan may train on what you paste unless you have opted out, so pair models on the tiers your firm has contracted; the Claude versus ChatGPT comparison sets out which plans train as of September 2026.

A worked example: the late-notice defence

Suppose your client supplied industrial coating to a buyer whose contract requires written notice of any claim within 30 days of discovering a defect. The buyer's formal letter of claim arrived 47 days after an internal buyer email in which its engineer, Mr Haddad, reported that the coating was flaking. Your position: the claim is barred by clause 14.2.

Put through Round 1 with the chronology and the neutral statement above, the scenario invites attacks of this shape:

  1. Discovery is not suspicion. The engineer's email records flaking, not a conclusion that the coating was defective; the 30 days run from discovery of a defect, and the buyer will say that came with the laboratory report on day 39. Targets the premise that day one is the email. [PROPOSITION TO VERIFY]
  2. The day-12 email. If the engineer copied your client's account manager on any message about flaking, that message is arguably written notice whatever its heading. Targets the fact that the first notice was the formal letter.
  3. Waiver by conduct. Your client's service team discussed the problem by telephone for three weeks and sent a technician without reserving its position. Targets the absence of a reservation of rights.
  4. Not a condition precedent. If clause 14.2 does not make time of the essence, late notice may sound in damages for the delay rather than bar the claim. Targets the governing rule as you stated it. [PROPOSITION TO VERIFY]
  5. Statutory control of the clause. Depending on the governing law and whether the terms were standard, a clause that extinguishes a claim for a defect may face a reasonableness test. Targets the enforceability premise. [PROPOSITION TO VERIFY]

The fact that ends it: an email from anyone at the buyer to anyone at your client within the 30 days that mentions flaking.

Your marks might read: 1 answerable, 2 needs evidence, 3 needs evidence, 4 fatal if true, 5 answerable. Round 2 then produces the question a judge would ask first, "Where in the bundle is the earliest written communication from the buyer about this coating?", and Round 3 turns it into the cross-examination table above. The exercise has done its job before any authority is touched: attack 2 sends someone to the account manager's mailbox, and attacks 1, 4 and 5 go to the research log for verification in Westlaw, Lexis or the national database, after which the verified authorities go back to the model with the instruction to "work only with the authorities I have provided" (The Legal Prompts, 18 Feb 2026).

PointAttack strength (Model A)Defence strength (Model B)Referee scoreWhy not higher
2 Day-12 email847Depends on a document nobody has yet found
4 Condition precedent766Turns on drafting the referee has not seen
3 Waiver674Three weeks of telephone calls is thin for waiver without a written concession

Read the referee's "why not higher" column twice. It tells you which points turn on a document you have not yet found, which on wording you have not yet quoted, and which are simply weaker than the attacking model made them sound.

Rules that keep red-teaming out of Charlotin's database

Damien Charlotin's database counted 2,041 court decisions involving hallucinated material by 14 September 2026, 811 of them involving lawyers. Adversarial prompting adds another place where a plausible authority can enter the file. In Withers v. City of Aberdeen (N.D. Miss., 8 Jun 2026) a trial was cancelled after both sides filed hallucinated citations (HAQQ, 11 Jun 2026). Five rules:

  1. No citations in any round. Propositions are flagged, verified in a research database with the three checks of existence, support and treatment, and only then handed back. The verification protocol is the full version; the hallucinations guide explains why paid legal tools need it too.
  2. Run it under counsel's direction on a business tier. In United States v. Heppner (S.D.N.Y., bench ruling 10 Feb 2026, opinion 17 Feb 2026) Judge Rakoff held that a defendant's consumer-Claude research memos were neither privileged nor work product, "because Claude is not an attorney" and the terms allowed logging and disclosure, while signalling that privilege might attach where counsel directs the use (Harvard Law Review Blog, Mar 2026). Warner v. Gilbarco (E.D. Mich., Feb 2026) reportedly went the other way for AI-assisted lawyer work product, treating the programs as "tools, not persons". In practice: the lawyer runs the rounds on a plan that does not train, labels the output as work product prepared in anticipation of litigation and files it in the matter. The privilege page has the detail.
  3. Check the standing orders. Over 30 of the 94 US federal district courts had AI disclosure or certification requirements by April 2026. Red-team notes are not filed, but a rebuttal paragraph that migrates into the brief is; the standing orders page lists who requires what.
  4. Assume the bench is doing the same. In a survey of 112 federal judges published on 27 March 2026, 61.6% used at least one AI tool and 28.6% used ChatGPT (Platinum IDS, 2026). The judge you simulated in Round 2 may run your brief through the same model. Under the UK judiciary's guidance of 31 October 2025, judges remain responsible for everything produced in their name, and the same is true of you.
  5. Log it. Prompt, model, date, and what changed in the argument. The record shows the tool tested judgment rather than replaced it, and it feeds the firm's prompt library with the rounds that worked.

What to do this week

  1. Take one live argument, write the chronology without adjectives and state the position under attack in one neutral sentence.
  2. Run Round 1 in a fresh session on your firm's business-tier model and mark the five attacks.
  3. Run Round 4 in a second fresh session, on a different model if you have one, and compare its rebuttals with your marks.
  4. Send every flagged proposition to research with the three-check rule before it goes anywhere near a draft.
  5. File the transcripts as work product in the matter, with a two-line note of what changed.
  6. For the next hearing, add Round 2 the day before and count how many of the six questions the judge actually asks.

Frequently asked questions

Can AI really argue the other side of a legal case?

Yes, and it is one of the tasks models do best, because generating counterarguments needs breadth rather than judgment. Asked for the five strongest attacks on a position, a frontier model produces five, ranked, with the fact each one targets. What it cannot do reliably is decide which attack matters or supply the authorities for it; that stays with you and a research database.

Which AI model is best for red-teaming a legal argument?

As of September 2026 the practical choices are Claude Opus 5, GPT-6 Astra and Gemini 3.1 Pro, all with roughly a million tokens of context. The Vals Index of 15 September 2026 ranks Claude Fable 5.1, Opus 5 and GPT-6 Astra as the top three, but the useful trick is not picking one; it is giving the same brief to two of them and treating the attacks both find as confirmed weaknesses.

Is it safe to put my case strategy into ChatGPT or Claude to stress-test it?

Only on a tier that does not train on your inputs and only when a lawyer directs the exercise. In United States v. Heppner Judge Rakoff held that a defendant's research memos from consumer Claude were not privileged, while signalling privilege might attach where counsel directs the AI use. Claude Team, Enterprise and API usage and ChatGPT Business and Enterprise are not used for training; consumer plans require an opt-out.

How is AI red-teaming different from using AI for legal research?

Red-teaming asks the model to attack reasoning and evidence you supply; research asks it to find law, which is where hallucinated citations come from. The protocol here bans citations in every round and has the model flag any legal proposition it relies on for verification. Authorities then come from Westlaw, Lexis or your national database, and only verified ones are handed back to the model.

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