The Witness Who Holds One Fact Back

Dubai should let software play the witness and keep a human in the judge's chair. The first is a training tool the region's young advocates have never had. The second is a grading decision, and grading is where legal education quietly decides what kind of lawyer it is producing.

1. What opened in Hall 6

The software arrived on Monday at the Dubai AI Festival. Legaline Platform, a UAE-based legal-tech company, launched Meta Legal Space on 26 October 2026, describing it as the first AI-powered mock trial in the Middle East and North Africa. A trainee picks litigation, arbitration or mediation, then picks a legal framework: UAE onshore, DIFC or ADGM. She then argues in front of an AI judge, an AI clerk, AI witnesses and AI experts. Opposing counsel can be another model or a classmate.

The design detail worth reading slowly sits in the witness specification. Each witness carries a preset background, a level of honesty, a level of composure under cross-examination, and one fact the witness will not volunteer. The trainee has to dig it out by questioning, and answers stay inside the case file. The launch cases were written by practising lawyers: a supply-contract breach with two witnesses and an expert report, a construction defects dispute, an unfair-dismissal claim before the DIFC Small Claims Tribunal. After each session the trainee gets a report scoring argumentation, procedural compliance and objection handling, and the best scores go on a leaderboard. Registration is free.

Jouslin Khairallah of KH Legal, a founding member of the platform, put the gap plainly in the release: "Trainees in this region have never had a place to rehearse a hearing." The point is right, and anyone who has watched a first-year associate freeze on her first objection knows the cost of the missing rehearsal. Treat the "first in MENA" claim as the vendor's own framing; it has not been independently checked.

2. The rulebook the trainee will graduate into

The simulator lands inside a profession that already counts its training. Under Dubai Executive Council Resolution No. 240 of 2015, the Continuing Legal Professional Development bylaw, a practising legal consultant must earn 16 CLPD points per filing period, at least 8 of them from mandatory activities the Legal Affairs Department specifies, with no single elective worth more than 4. Non-compliance runs to a 15-day notice and, eventually, referral to the regulator.

What the Department now designates as mandatory is telling. Its current CLPD course catalogue lists "Responsible Use of AI for Legal Professionals", "AI Governance: Legal Risk & Compliance" and "Ethical AI and Algorithmic Governance" among the mandatory activities, alongside elective courses on specific chatbots. Dubai has made AI literacy compulsory for working lawyers. It has not yet said anything about AI in advocacy training.

The courts have been clearer about duties than about pedagogy. The DIFC Courts' Practical Guidance Note No. 2 of 2023 tells parties to declare AI use early, to verify outputs against case law and statute before relying on them, and reminds signatories that a false statement made without honest belief engages the contempt rule. And on 29 September 2026, at the UN Crime Congress in Abu Dhabi, the UAE Minister of Justice launched ten guiding principles for AI use by law firms, with one line that should be printed above every simulator: "No matter how advanced technology becomes, professional responsibility cannot be delegated."

Put the pieces side by side and the pattern is visible.

In every row but the first, a person or a public body owns the sign-off by default. That asymmetry is the subject of this piece.

3. Who grades the rehearsal

An AI witness that withholds a fact is a sensible design. An AI judge that rules on an objection under DIFC procedure and then scores the trainee on "procedural compliance" is a different sort of machine, because it is making legal determinations and teaching them as correct.

The evidence on how often legal models get the law wrong is sobering. Varun Magesh, Daniel Ho and colleagues at Stanford tested the leading commercial legal research tools, the retrieval-backed products from LexisNexis and Thomson Reuters, and found they hallucinated between 17 and 33 percent of the time. Those are well-funded systems built specifically to avoid invented law. A courtroom simulator ruling in real time on hearsay, leading questions and relevance will sometimes be confidently wrong, and a trainee who loses a point to a wrong ruling learns the wrong rule. A leaderboard then rewards whoever adapted best to the model's mistakes.

Legaline has already built the remedy into the product: a licensed practitioner from its network can preside and score a session live. The question for Dubai's regulators is which of the two modes gets to count.

4. The strongest case against simulators

The most serious objection comes from outside law. Hamsa Bastani and colleagues ran a field experiment with nearly 1,000 high-school maths students in Turkey, published in PNAS in June 2025. Students given an unrestricted GPT-4 interface scored 48 percent better during practice. Those given a tutor version with guardrails scored 127 percent better. When access was removed for the exam, the unrestricted group did 17 percent worse than students who never had the tool, and the tutor group came out roughly level with them. The authors' reading is that unfettered access can harm learning, because practice performance and acquired skill come apart.

That finding argues against any AI rehearsal space whose scores look like progress. I take it seriously, and I think it cuts in a narrower direction than it first appears. Meta Legal Space looks more like the guardrailed tutor than the open chatbot: the witness does not hand over the answer, the trainee has to extract it, and the case file bounds what the model can say. The design pushes effort onto the student, which is exactly what Bastani's tutor did. Where the study bites is the scoreboard. Bastani's practice grades overstated learning by a wide margin. A simulator's session scores are practice grades.

In the oral examinations I have sat in on as a visiting professor, the pattern I remember best is the strong written candidate who came apart when an examiner interrupted with a question she had not rehearsed. Composure under interruption was the skill missing from her file. A good simulator can drill it. Only a human examiner could tell whether she had it.

So my position is specific. I would refuse to sign off on any CLPD accreditation, university credit or bar-admission credit for simulator sessions scored only by the AI judge. I would accept, today, accreditation for sessions where a licensed practitioner presides, with the AI cast playing every other role.

5. The bench is being trained too

The same question is arriving on the other side of the courtroom. At Riyadh's second International Conference on Judicial Training, Egyptian judge Mohamed Samir Marai presented a digital platform simulating litigation and judgment drafting, with pre- and post-assessment and automatic tracking of professional-development hours, and insisted it operate under "mandatory human verification". In Abu Dhabi, the UAE Ministry of Justice says its five AI agents now support five of seven case-file review stages, cutting review from days to hours, while the draft judgment stays non-binding until a judge approves it. Those time savings are the Ministry's own figures and have not been independently measured.

The natural accreditor sits between these two worlds. The DIFC Courts' 2026-2030 strategy commits to deeper engagement with law firms, academic institutions and professional bodies to develop the next generation of court professionals. A roster of DIFC-registered practitioners willing to preside over simulator sessions would cost the courts little and settle the grading question.

The library at a Dubai law school stays open late during exam weeks. Somewhere in it this term, a second-year student will spend an hour cross-examining a construction expert who knows the concrete was poured two days early and will not say so unless asked the right way. On the third attempt she will ask it the right way, and the expert will say it. The next morning a practitioner from Jumeirah Lake Towers will sit in the judge's chair on the same file, interrupt her on the second question, and watch what she does with her hands.

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan for Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.