Rehearsal Before the Bench

For as long as there have been courtrooms, the first time most people stood in one was also the first time they were tested by one. A young advocate's debut cross-examination, a litigant's first contested hearing, a newly appointed judge's earliest ruling from the bench, each unfolded live, with real stakes, on the record. The dress rehearsal, where it existed at all, was a luxury rationed to those who could summon senior colleagues, hire a mock jury or book a moot court. Now a new category of tools is changing that arithmetic. Agentic AI personas, software that convincingly plays the judge, the hostile witness, the opposing lawyer or the examining bench, let people rehearse the proceeding before they live it.

This is not a story about robots replacing lawyers or algorithms deciding cases. It is a story about practice: about who gets to prepare, how realistically, and at what cost. The implications run straight through the most strained part of the justice system, where the overwhelming majority of people who walk into a courtroom do so without a lawyer at all.

The Old Way: Practice Was a Privilege

The legacy model of courtroom preparation was built around scarce human time. Advocacy training leaned on the moot court, a simulated argument before volunteer judges, precisely because the skills it teaches cannot be read from a book. Guides for oral advocacy stress that the craft is learned by drilling: anticipating the question you most fear, answering without notes, holding a core theory under pressure, and treating the bench with composure when interrupted (Stetson University College of Law). One competition guide puts it plainly: students "can drill by asking questions over and over again so that when they get to the competition, they will have built up a memory bank of responses" (Lewis & Clark Law School).

The problem was never the value of the rehearsal. It was the supply. A moot requires judges, opponents, a fact pattern and a room. Deposition and trial training historically depended on senior attorneys volunteering hours to play the witness and critique the performance, an apprenticeship model that legal commentators describe as effective but expensive and unevenly distributed (Daily Journal). Junior lawyers at well-resourced firms got many reps; everyone else got few.

For people without lawyers, the rehearsal simply did not exist. And those people are not a fringe. Drawing on samplings of state court records, an estimated 75 percent of civil and domestic-relations cases involve at least one self-represented litigant, and roughly 25 percent involve no lawyer at all (Fordham Law Review). National Center for State Courts "landscape" studies found that 76 percent of civil, non-family cases and 72 percent of family cases involved one or more self-represented parties, a stark reversal from 1992, when 95 percent of general-jurisdiction civil cases had attorneys on both sides (International Journal of Online Dispute Resolution). More than thirty million people a year appear without counsel in America's state and county courts (Fordham Law Review).

The consequences compound. In some courts, eviction and child-support dockets are almost entirely lawyerless: one analysis found 98 percent of tenants in eviction cases and 95 percent of parents in child-support cases were unrepresented in a major state system (Legal Services Corporation). Researchers studying these "primarily lawyerless" courts conclude that traditional adversary procedure "has largely disappeared" and that the experience "amplifies inequality and human suffering" (New York County Lawyers Association). Globally the gap is vast: the World Justice Project estimates that roughly 5.1 billion people, about two-thirds of the world's population, face at least one unmet justice need (World Justice Project).

The Shift: Personas You Can Practice Against

The breakthrough is not that AI can argue a case. It is that AI can play a role convincingly enough to be a useful sparring partner. The clearest evidence comes from an adjacent profession that has spent two years testing the same idea: medical education, where AI "standardized patients" stand in for human actors during clinical-skills training.

In a randomized trial of 56 medical students, the group that practiced history-taking with a generative-AI simulated patient scored significantly higher on a structured clinical exam than those who practiced with traditional role-play, 86.79 versus 73.64 on a 100-point scale (BMC Medical Education / PubMed). A separate prospective study found an AI-powered simulator produced role-play responses that were medically plausible in more than 99 percent of cases, with feedback that reached "almost perfect" agreement with a human rater (JMIR Medical Education / PubMed). A randomized gynecology-clerkship trial similarly found students using an AI patient achieved significantly higher completeness and organization in their interviews (PubMed).

The lesson that transfers to courts is precise: availability and reps. Students consistently rated AI practice as freer from time and environmental constraints and easier to self-pace (PMC / BMC Medical Education). A synthetic opposing counsel does not need to be scheduled. A synthetic judge will hear the same opening twenty times without fatigue. For a discipline whose entire pedagogy is "drill the question until the answer is automatic," that is a structural change.

The machine doesn't have to be a better advocate than a human. It only has to be available when the human mentor isn't.

Courts and the legal world are already signaling demand. In the National Center for State Courts' 2024 public poll, support for courts using AI to help people was striking: 71 percent backed using AI to break down legal jargon and make information more accessible, and support rose to 77 percent among people who had represented themselves (National Center for State Courts). On the practitioner side, one 2025 industry survey reported over 95 percent of legal professionals expect generative AI to become central to their workflow within five years (Thomson Reuters survey, via LawSites).

What It Looks Like Now

In practice today, agentic panels show up in four distinct courtroom-adjacent workflows, each generic to the category rather than any one platform.

1. Advocate training and deposition prep

An advocate uploads a closed set of case materials and conducts a live, spoken cross-examination or deposition against AI personas that play the witness, opposing counsel and even the court reporter, then receive structured feedback on structure, timing and objection-handling (Suffolk University Journal of High Technology Law). The model is explicitly framed as a supplement to, not a replacement for, supervised practice with senior attorneys (Above the Law).

2. Judicial-education moot benches

Judicial training already operates at scale, the National Judicial College alone reports that roughly 10,000 judges study with it each year across some 90 programs (National Judicial College). Agentic personas let educators run repeatable bench scenarios, an emotional self-represented party, an evasive expert, so judges can rehearse procedural-fairness skills that courses such as "Procedural Fairness and Empathy for Self-Represented Litigants" are designed to build (National Judicial College).

3. Helping self-represented litigants rehearse

This is the highest-leverage use. A litigant can practice answering a judge's questions, telling their story within time limits, and hearing what an opposing lawyer might say, before the hearing where it matters. Procedural-justice research is unambiguous that how people experience a hearing, whether they feel heard and treated with dignity, shapes their trust in the outcome (Supreme Court of Ohio). Yet only 27 percent of the public believes courts effectively help people self-represent (National Center for State Courts).

4. Readiness and fairness diagnostics

Because every rehearsal is recorded, the same systems can flag where a person is unprepared, missing a key fact, exceeding time, misreading the bench, turning practice into a measurable readiness signal rather than a one-shot gamble.

The Hard Limits: Realism, Fairness and Over-Reliance

Every promise here carries a matching risk, and honest assessment requires naming them. The first is realism's ceiling. In the medical analogue, AI role-play sometimes lagged human practice on the most demanding interpersonal dimensions: in one pilot, students rated peer role-play higher for handling unexpected situations and for the "appropriate tension and responsibility" of a live encounter (PMC / BMC Medical Education). A synthetic judge will not replicate the unscripted human pressure of a real bench, and over-training against a predictable persona can breed false confidence.

The second risk is substantive error. If a litigant or advocate lets an agentic system drift from rehearsal into legal advice, accuracy becomes a live hazard. Stanford researchers found general-purpose language models hallucinate on specific legal queries between 58 percent and 88 percent of the time, and at least 75 percent of the time when asked about a court's core holding (Stanford Law School). Even purpose-built legal research tools with retrieval grounding hallucinated between 17 percent and 33 percent of the time in a follow-up Stanford study (Stanford / RegLab). A practice partner that occasionally invents the law is fine for drilling delivery; it is dangerous if mistaken for counsel.

That leads to the third risk: the line between rehearsal and the unauthorized practice of law. A tool that helps someone practice how to present is categorically different from one that tells them what their rights are. The same NCSC poll that found enthusiasm for AI in courts also found the public divided on letting non-lawyers handle legal matters, 60 percent supportive but 26 percent worried about "mistakes and inadequate representation" (National Center for State Courts). Courts' own AI guidance bodies have begun flagging related concerns, including the threat AI-generated material poses to public trust in proceedings (National Center for State Courts).

The Next Few Years

Three shifts look likely over the next three to seven years. First, rehearsal becomes a standard pre-hearing step rather than an elite advantage, embedded in self-help portals that the public already supports courts building, and in clinical and judicial-education curricula where the evidence for AI role-play is strongest. Given that adoption of generative AI among legal professionals more than doubled in a single year (LawSites), the practice layer will follow the research layer fast.

Second, the realism gap narrows but never closes. Model generations have already cut legal hallucination rates meaningfully, one analysis traced general-model error from 88 percent down to 58 percent across generations, with grounded tools lower still (AI Law Librarians). Persona realism will improve similarly. But the most valuable rehearsals will be hybrid: AI for volume and availability, humans for the unscripted pressure and judgment that machines still approximate poorly.

Third, governance hardens around the rehearsal/advice boundary. Expect court-rule guidance, disclosure norms and bright-line separations between tools that help people perform and tools that purport to tell them the law. The institutions that publish today's justice-gap and procedural-fairness research, courts' own statistical bodies, judicial colleges and access-to-justice commissions, are the ones most likely to set those lines.

Conclusion

The courtroom has always rewarded the prepared and punished the unready, and preparation was distributed as unequally as everything else. AI-persona mock proceedings do not change what a hearing demands. They change who gets to rehearse for it. If the category holds its discipline, practice not advice, supplement not substitute, human judgment in the loop, it could narrow one of the oldest and quietest inequities in justice: the gap between those who walk into court having done it before, and those facing it cold for the very first time.

Sources

1. Fordham Law Review, self-representation rates in civil and domestic-relations cases
2. Fordham Law Review, 30+ million appear without counsel annually
3. International Journal of Online Dispute Resolution, NCSC landscape study figures
4. Legal Services Corporation, The Justice Gap report
5. New York County Lawyers Association, "primarily lawyerless" courts
6. World Justice Project, Measuring the Justice Gap (5.1 billion)
7. Stetson University College of Law, Oral Argument: The Essential Guide
8. Lewis & Clark Law School, Appellate Moot Court Guide
9. Daily Journal, The future of litigation training
10. BMC Medical Education / PubMed, randomized GPT history-taking trial (86.79 vs 73.64)
11. JMIR Medical Education / PubMed, LLM simulated patient, 99% plausibility
12. PubMed, randomized gynecology AI simulated-patient trial
13. PMC / BMC Medical Education, AI chatbot vs peer role-play OSCE pilot
14. National Center for State Courts, State of the State Courts 2024 poll
15. LawSites, Thomson Reuters generative-AI survey
16. LawSites, AI adoption among legal professionals doubled
17. Suffolk University Journal of High Technology Law, AI deposition practice
18. Above the Law, AI deposition simulator
19. National Judicial College, ~10,000 judges trained annually
20. National Judicial College, procedural-fairness course for SRLs
21. Supreme Court of Ohio, Procedural Fairness for Self-Represented Litigants
22. Stanford Law School, legal hallucination 58 to 88%
23. Stanford RegLab, purpose-built legal tools hallucinate 17 to 33%
24. AI Law Librarians, hallucination rates improving across model generations
25. National Center for State Courts, AI-generated evidence and public trust