On 31 July 2026, a full bench of the Gauteng Division in Johannesburg handed down an appeal judgment that reads less like a ruling on a school-choice dispute and more like a case study in institutional design. According to GroundUp's 4 August report, Judge Ingrid Opperman found at least 11 fictitious or incorrect citations in the judgment of Acting Judge Themba Khaba, including a case, Lubbe v Volkswagen SA, that does not exist. She said the errors were most plausibly explained by AI "hallucinations" and referred the matter to the Legal Practice Council (LPC).
The panel of Judges Gregory Wright, Daphny Mahosi and Opperman still upheld Khaba's findings on the merits of a parental dispute over children's schooling. That detail matters. The outcome was right, but the paper trail supporting it was unreliable.
The strongest case for tougher rules
The argument for a firm regulatory response is serious. A judgment is not an ordinary document. It binds parties, becomes precedent and is presumed to rest on real authority. When a judge cites a case that does not exist, the litigants cannot tell whether the reasoning was genuinely derived from law or dressed up afterwards. Opperman herself said she would not hesitate to find rights violated had she been able to conclude that the reasoning was AI-generated, which shows how close the line is. Those who favour a formal ban or mandatory disclosure regime for judicial AI use argue that trust in courts is too fragile to risk on tools known to invent sources.
What the case actually shows
The case shows a verification failure, and South African courts have already been building law on verification failures. In Mavundla v MEC: Department of Co-Operative Government and Traditional Affairs KwaZulu-Natal, decided on 8 January 2025 in the Pietermaritzburg High Court, a notice of appeal cited non-existent cases including Pieterse v The Public Protector. The court dismissed the application, ordered the firm to pay costs from its own pocket and sent the judgment to the LPC. Judge Bezuidenhout called the conduct "irresponsible and downright unprofessional" and stressed that practitioners must verify every authority they cite.
Notice what did not happen in either case. No court said AI research tools should be prohibited. The remedy in Mavundla was personal cost liability and a professional-conduct referral. The remedy in Johannesburg was the same referral. That is the proportionate model: the tool is allowed, the human who signs the document owns every line.
The errors in the Johannesburg case were also caught by the ordinary machinery of adversarial litigation. Opperman detailed all 11 discrepancies, which lawyers for the appellant had raised and which a senior librarian at the Johannesburg Society of Advocates confirmed, according to GroundUp. Appeal works as a check when there is someone with standing and resources to run one. It is less reliable in the many matters where no one can afford to appeal.
The accountability gap
The more consequential finding concerns who can discipline whom. Opperman noted that the Judicial Service Commission has no authority over acting judges because it does not appoint them, and that "the easy answer to misconduct by an acting judge is simply not to renew their appointment," as GroundUp reports. Acting judges remain enrolled legal practitioners, so the LPC, which describes its mandate as regulating the professional conduct of attorneys and advocates, is the only body with jurisdiction. The Judicial Service Commission and its Judicial Conduct Committee handle complaints about judges, but that machinery does not reach this situation.
This is a structural problem that predates generative AI. Acting appointments are a routine way to relieve court backlogs. If the only sanction for careless work is quietly not being asked back, then the incentive to check a draft against a law report is weak. Nothing in the reporting suggests Khaba was acting in bad faith, and the LPC has yet to investigate, so no finding of misconduct exists. The point is that the current framework would not easily produce one through the judicial route.
A proportionate response
Policy should follow the evidence of where the failures occurred. Three measures fit.
- Verification as an explicit duty for judicial officers. Practitioners already carry it after Mavundla. The Office of the Chief Justice can issue a practice directive extending it to acting and permanent judges alike, with no legislation required.
- Closing the acting-judge gap. The Judicial Service Commission's complaints framework, or a defined protocol with the LPC, should specify who investigates an acting judge's conduct on the bench, so that referral to a professional regulator is not the accident of an enrolment status.
- Institutional tooling, not prohibition. Courts can give judges access to verified citation databases and drafting tools that link to actual SAFLII or law-report sources. A blanket ban would push use underground, where nobody sets any controls at all.
Overreaction carries real costs. South African courts face heavy backlogs, and drafting assistance is one of the few technologies that can help judges write faster. Rules that treat any AI use as misconduct would forfeit that benefit to punish a failure that verification prevents. Rules that ignore the problem would leave litigants with judgments they cannot trust.
The Johannesburg bench has, in effect, done the sensible thing: it corrected the record, protected the outcome that was substantively sound, named the accountability gap and pointed to the regulator with jurisdiction. The task now falls to the judiciary's administrators to make that gap smaller before the next hallucinated authority appears in a ruling where nobody appeals.