The Honourable Mr Justice Claude – generative AI in the courtroom

Dario Milo, Partner at Webber Wentzel and co-head of the firm’s AI Disputes unit.

---

It was 1999. I was completing my legal training. The partner I was working for, the formidable Peter Reynolds, was appointed to arbitrate a complex commercial dispute. I served as his assistant at the arbitration, performing the same role as judicial clerks do for judges.

It came time to hand down his award. Bright-eyed and bushy-tailed, I did a first draft of the award based on Reynolds’ views of the dispute and his direction as to how the award should be crafted. Reynolds spent time reviewing, restructuring and making substantive edits to the draft and the award was issued. The losing party was furious, even more so when they received Reynolds’ detailed fee narration and realised that I had prepared the initial draft of the award.

This was a reviewable gross irregularity and arbitrator misconduct, the losing company claimed – Reynolds had delegated his decision-making to me, his lowly clerk. The company sought a review of the award under the Arbitration Act.

The case went all the way to the Supreme Court of Appeal (SCA) where Reynolds was vindicated. The Court held that the parties to the arbitration had agreed that his assistant (yours truly) would do substantive work for the arbitrator; this would clearly cover research and could include the preparation of a first draft of the award, under the arbitrator’s direction. “There could be no breach of [Reynold’s] duty as arbitrator if he availed himself of assistance expressly or impliedly agreed upon.”

I was reminded of this episode when reading what is now the leading judicial discussion in the common law world on the risks of a judge delegating decision-making to AI, the judgment of Justice Ingrid Opperman at the end of July in a family law case, FJL v TGO.

The underlying facts concerned whether two children should be removed from their present private school and enrolled at a less expensive public school. An acting judge ruled in favour of the mother that the children should stay put. The father’s application for permission to appeal was refused by the judge and he then sought permission to appeal from the SCA.

The acting judge ruled that while this application for leave to appeal was pending, the children should not be removed from the private school. It was this last ruling that went on urgent appeal to the full court of three judges – including Justice Opperman.

All three judges sided with the mother – while the father pursued his application for permission to appeal before the SCA, the children could not be moved.

Justice Opperman went further, addressing the elephant in the courtroom – the acting judge’s judgment contained 11 discrepancies: cases which had the wrong citation, incorrect quotations from cases and a reference to a non-existent case. “The most plausible explanation, certainly for the fictitious reference, is that it is a product of the use of artificial intelligence and what has been dubbed ‘hallucinations’,” said Justice Opperman. While she made no finding on whether AI was actually used, the judge went on to discuss the implications of the use of AI on court decisions.

Judge Opperman correctly framed the issue as one implicating section 34 of the Constitution, the right to have a dispute that can be resolved by application of law decided “in a fair public hearing before a court or, where appropriate, another independent and impartial tribunal or forum”.

“A judgment must reflect the Judge’s considered view, rather than the output of a process that lacks transparency and in which the genesis of the conclusion remains concealed or obscure. Judicial authority must remain human, accountable and transparent.” If “the true decision-maker was a machine”, continued Justice Opperman, there would have been a contravention of section 34.

She noted that her judgment would be forwarded to the Legal Practice Council to investigate whether the acting judge committed any professional misconduct.

This reasoning chimes neatly with the Constitutional Court’s leading case on section 34, in the dispute between Vodacom and “Please Call Me” inventor Nkosana Makate: “The court must bring its reasoning to bear on …material issues and reach a conclusion on them.”

This duty of proper consideration will not be discharged where a judge effectively outsources the decision-making to AI as opposed to using it to enhance efficiency – for research, summation, editing and possibly drafting support, always subject to the need to verify its output.

While section 34 does not apply to private arbitrations, arbitrations must also be conducted fairly. Given this fairness obligation and the Arbitration Act’s grounds of review based on misconduct or gross irregularity, there can be little doubt that a court would set aside arbitration awards where the arbitrator has effectively delegated decision-making to their AI tool.

As the SCA stated in the case in which I was the assistant: “In no circumstances may the assistant be allowed to usurp the decision-making function of the arbitrator... Ultimately the question to be asked, and answered, is whether the arbitrator exercised his own judgment in deciding the issues.”

This is a live issue. It was reported in April this year that the Quebec Superior Court had set aside an arbitrator’s award on the basis that his use of AI – again evident through hallucinated case authorities – effectively meant the decision had been delegated to the AI tool.

One of the differences between arbitrations and court proceedings is that arbitrations are based on consent. So as long as there is transparency and informed consent, the parties can always agree to have AI involved in a more substantive manner in resolving their dispute. By way of example, the American Arbitration Association introduced an AI Arbitrator in November 2025 for low-value construction disputes where no witnesses are required. The AI tool reviews the submissions, summarises them (which the parties check for accuracy) and then prepares a draft award which a trained human arbitrator reviews and finalises.

AI has already altered dispute resolution practice around the world. There are cases in the United States and United Kingdom which make it clear that asking an AI tool for legal advice does not protect that legal advice from discovery in litigation. Courts and tribunals in those countries have held that uploading legally privileged information into a non-enterprise AI tool means confidentiality and legal privilege are lost.

A US court has ordered that prompts used by an expert to select certain documents from a dataset should be disclosed to the opposing party. And two weeks ago, in a class action dispute involving McDonald’s, an Australian Federal Court judge, Michael Lee, ordered the parties to investigate and report to the court on using AI technology “to turn an enormous body of primary information into something intelligible enough to permit the parties, and ultimately the Court, to understand what it shows.”

This is a brave new litigation world. There may not yet be a place on the bench for The Honourable Mr Justice Claude, but, in the words of Justice Lee: “The objective is to use technology to perform work which does not itself require human judgment, so that professional and judicial judgment can be concentrated upon the matters which do.”