A prepared argument is not necessarily a tested argument. An opposing reader can expose an unstated assumption, a weak authority or a factual distinction that a sympathetic colleague overlooks. Structured AI disagreement is one way to rehearse that challenge, provided its limits remain explicit.
Compelle's historical Bittensor research examined adversarial argument between models. The application to legal preparation is a proposal for supervised practice, not evidence that an AI debate predicts court outcomes or replaces an expert tutor.
Practice the point that is hardest to defend
Choose a narrow issue rather than asking whether an entire case is strong. State the governing standard, identify the facts you are permitted to assume, and separate disputed facts from agreed ones. Give the opposing position a concrete task: find the weakest step between the authorities and the conclusion.
After the exchange, revise that step. Adding more confident language is not a repair. Better authority, a narrower claim or an explicit concession may be.
Five useful preparation exercises
1. Moot court preparation
Turn the central issue into a short proposition. Ask for the strongest objection and prepare a direct answer before reviewing a model's suggestion. Have an instructor or qualified lawyer assess the answer against the actual record and law.
2. Deposition planning
Use a hypothetical or appropriately authorized, minimized factual record to explore alternative explanations. Map the questions needed to distinguish those explanations. A generated response is not evidence of what a witness knows or will say.
3. Cross-examination rehearsal
For each question, identify the intended admission, the evidentiary basis and a plausible evasive response. Rehearse follow-ups without inventing facts that would make the examination easier.
4. Brief review
Ask an opposing reader to identify the strongest counterauthority, the most consequential distinction and any claim that overstates its source. Verify every citation and quotation in an authoritative legal source before relying on it.
5. Oral argument
Restate the same issue under several framings. Practice answering the difficult question first, then explaining the limiting principle. An argument that works only under one favorable wording deserves closer scrutiny.
A concession is feedback, not a legal verdict
In the historical research, explicit concessions offered a visible signal that a model had changed its stated position. They did not establish that the conceding model was wrong on the law, that the winning position was true, or that a human judge would agree.
Useful review asks why the model conceded. Did a genuine contradiction emerge? Did it accept a nonexistent authority? Did it lose track of the applicable burden? The transcript can reveal these distinctions; the score alone cannot.
Protect the record and the client
Do not put confidential or privileged information into a tool without appropriate authorization and a review of its data-handling terms. Use fictional, public or properly minimized material for practice. Professional duties, court rules and client confidentiality still govern the workflow.
AI assistance does not remove the lawyer's responsibility to verify sources and exercise independent judgment. It can be a challenging rehearsal partner; it is not counsel, a witness or a court.
Measure the improvement in the argument. A clearer limiting principle, a verified authority and an answered objection are more useful than a win against a model.
See the method in practice
Read Compelle's method and sample work. For a consequential question, Compelle offers a white-glove engagement scoped with you rather than an automated legal-advice product.
Put a consequential question to the test.
A white-glove engagement, scoped around your decision and the evidence it needs.
Work with us