Judge Faces Accountability for Fabricated Case Citations in AI-Assisted Ruling

Acting Judge Themba Khaba of the Johannesburg High Court must now answer to the Legal Practice Council after a judgment he issued was found to contain at least 11 problematic case citations, some of which appear to be entirely fabricated. The discovery cuts to the heart of judicial accountability and raises pointed questions about what happens when artificial intelligence enters the courtroom unchecked.

The problem came to light when three appeal judges reviewed a custody dispute ruling Khaba had handed down. The appeal panel upheld his core findings, but Judge Ingrid Opperman issued a separate judgment cataloguing serious flaws in how he had cited legal authorities. The errors were varied and troubling: cases that do not exist, passages quoted from cases that contain no such language, correct case names paired with wrong citations, and wrong citations paired with correct names.

One fabricated case stood out in particular. Lubbe v Volkswagen SA, Opperman noted, has no apparent foundation in law. The pattern, she concluded, was consistent with what AI systems produce when they generate plausible-sounding but false information, a phenomenon widely described as “hallucinations.” She stopped short of making a formal finding that AI was used, but said the evidence pointed strongly in that direction.

The broader concern extends well beyond a single flawed judgment. Opperman observed that very few judges anywhere have been publicly identified as using AI in their rulings. She cited Lord Justice Colin Birss of England and Wales, who disclosed using ChatGPT as a drafting aid, but was careful to note the distinction: Birss had verified all information himself and treated the tool only as a research helper, not as a decision-maker. That distinction matters enormously for the public.

If a judge delegates actual decision-making to a machine, Opperman stressed, that would breach the judicial oath and deny litigants their constitutional right to be heard by a human court. She acknowledged that Khaba must be given the opportunity to respond and that due process must be followed.

The case also exposed a significant gap in how the judiciary holds acting judges accountable. The Judicial Service Commission, which disciplines permanent judges, has no formal authority over acting judges because it does not appoint them. The only practical consequence for misconduct is non-renewal of an acting appointment, a measure that carries no public sanction and leaves no record of findings. Citizens who appear before an acting judge have no reliable mechanism to know whether that judge has previously been found to have acted improperly.

Opperman noted that this accountability vacuum has never been properly resolved through legislation or judicial precedent. An acting judge who is also a practicing lawyer or advocate, however, remains subject to the Legal Practice Council’s disciplinary powers. She identified the LPC as the most effective available forum to investigate the matter and indicated she would refer her judgment to the council accordingly.

She was clear that the LPC’s jurisdiction should run alongside, not instead of, the JSC’s authority. Once the LPC completes its investigation, the matter could still be referred to the JSC’s Judicial Conduct Committee if warranted.

The underlying dispute involved two parents arguing over their children’s schooling. The father, described as wealthy, had unilaterally stopped paying private school fees and filed an urgent application in November 2025 while boarding a business class flight to Europe. Khaba’s order preserved the status quo, allowing the children to remain at their current school. The appeal judges found his reasoning sound, even as they grappled with the citation errors. The children’s schooling, at least, was not disrupted.

What changed with Opperman’s judgment is the public record. The judiciary’s ongoing efforts to develop an AI policy now carry fresh urgency, and the question left open is a practical one: how will courts ensure that any use of AI in judicial work remains transparent, independently verified, and clearly subordinate to human reasoning, before the next flawed judgment reaches an appeal bench?