INVENTED LAW: WHAT HAPPENS WHEN THE JUDGE CITES A CASE THAT DOES NOT EXIST

For close to two years the profession has been having the same uncomfortable  conversation about artificial intelligence, and it has always had the same shape in that a practitioner under deadline pressure asks an AI tool for authority; the tool obligingly  invents a case; the fiction survives into heads of argument; a judge notices. Mavundla and Northbound Processing are the known examples. In each, the court was the party  that caught the error. 

A full court of the Gauteng Division, Johannesburg has now delivered a judgment in  which the court itself is the subject of the enquiry. In FJL v TGO, delivered on 31 July  2026, Opperman J found eleven discrepancies in the citations relied on by an acting  judge in the court below — including a case that appears not to exist at all — and  referred the matter to the Legal Practice Council. The judgment is marked reportable  and of interest to other judges and deserves to be read in full. 

How the case came to light 

The underlying dispute was a hard-fought fight between divorced parents over where  their two children should be schooled. The father applied urgently to move them from  an expensive private school to a public school; the mother counter-applied to enforce  the private school obligations under the 2019 divorce order, including arrear fees of  R277 078.50. The acting judge dismissed the father's application and granted the  mother's counter-application. When the father's application for leave to appeal  triggered automatic suspension of that order, the mother obtained relief under section  18(3) of the Superior Courts Act 10 of 2013, which the father appealed as of right under  section 18(4). 

Shortly before the appeal, the father's legal team filed a supplementary note identifying  discrepancies in the acting judge's citations. The verification that followed is what  separates a serious allegation from an insinuation in that the respondent's team  agreed the discrepancies were accurately described, the senior librarian of the  Johannesburg Society of Advocates confirmed them, and Opperman J independently  instructed a court researcher to verify them before the hearing.

The resulting table reveals three distinct failure modes rather than a list of  typographical slips. In some entries the correct case name is attached to the citation  of an entirely unrelated matter. In others the name and citation are correct, but the  quoted passage is not in that judgment — a quotation from one case spliced onto a  quotation from another, uncited, case. In the starkest category, name and citation  correspond to nothing at all. The example Opperman J singled out is Lubbe v  Volkswagon SA [2023] 4 All SA 289 (C), relied on in five separate paragraphs of the  judgment under appeal. It does not exist. 

Critically, these references did not come from counsel. Junior counsel for the  respondent below had not put any of them before the court, notwithstanding that the  judgment in places recorded that she had. With counsel excluded as the source,  Opperman J concluded that the most plausible explanation, certainly for the fictitious  Lubbe reference, is AI “hallucination”. 

Two judgements, two approaches 

The appeal court was not unanimous on how to handle this. Wright J, with Mahosi J  concurring, dismissed the appeal on the merits and expressly declined to enter the AI  debate. Opperman J agreed with the order but not the route, and wrote separately. 

Her reasoning on why she could not stay silent matters as much as her conclusions.  She dealt with the discrepancies at face value and deliberately did not raise them  privately with the acting judge, on the basis that any explanation belongs before the  body responsible for regulating conduct — not in a chambers conversation away from  the scrutiny of the litigants. That is a point about institutional transparency as much as  about AI. She was equally careful about what she was not deciding, she made no  finding that AI was in fact used, and recorded that she could locate no reported  authority anywhere — South Africa, the UK, Canada, Australia or the United States — in which a judge had been found to have used AI in judicial reasoning. 

Why this is not a technicality 

The most valuable part of the judgment is Opperman J's explanation of why fabricated  authority is a constitutional problem rather than a housekeeping one.

Her starting point is disarmingly simple. Names of non-existent cases are not law, and  quotations jumbled together from cases — some identified, some not — are not law  either. From there she develops a point about the nature of legal reasoning itself. An  essential quality of law is that its verifiability in that others must be able to find it and  check it. In that sense the exercise is academic and scientific — the same process  using the same ingredients ought to be reproducible by peers. Universities prescribe  referencing conventions precisely so scholarship can be relied upon. Confidence in a  legal system is built the same way, on predictability and verifiability, which she  identifies as the foundations of trust. 

This is where erosion of public confidence becomes concrete rather than rhetorical. A  judgment is not merely an instruction to the parties. Drawing on Stuttafords Stores  (Pty) Ltd v Salt of the Earth Creations (Pty) Ltd, Opperman J reaffirms that reasons  explain to the parties, and to a public with a legitimate interest in open courts, why a  case was decided as it was — and that this discipline is what curbs arbitrary judicial  decisions. Remove the ability to check the sources and you remove the discipline.  What remains looks and reads like a judgment, but nothing in it can be tested. 

Section 34 of the Constitution guarantees everyone the right to have a dispute decided  in a fair public hearing before a court, and its purpose, per Lesapo v North West  Agricultural Bank, is to protect the judicial process itself. Opperman J's conclusion is  uncompromising in that, had she been able to find that the reasoning was AI-generated  and the true decision-maker a machine, she would not have hesitated to find the  appellant's section 34 rights infringed. Diverting decision-making to a robot breaches  the oath of office and denies a party the right to be heard at all. A body of precedent  salted with untraceable citations does not simply contain errors — it degrades the raw  material every future court and practitioner must work with. 

The distinction that actually matters: assistant versus decision-maker 

Public commentary tends to collapse into one question — should judges and lawyers  use AI at all? Opperman J's survey shows that's the wrong question. The line every  serious jurisdiction has settled on is between AI as an assistant and AI as a decision maker. On the permitted side sit summarisation, research assistance, editing and  drafting support — on one condition: the judge verifies the output and remains  personally responsible for the final judgment. 

Lord Justice Birss of the England and Wales Court of Appeal, who disclosed using  ChatGPT to summarise an area of law while drafting a judgment, is the standard  illustration, useful as a drafting aid, never as a substitute for verification. Australia's  Federal Court and the Supreme Court of Victoria take the same position, as does the  Supreme Court of India in Pooja Ramesh Singh v Jammu and Kashmir Bank Ltd 2026  INSC 668, which went further and held that a decision resting on hallucinated material  is no decision at all, to be set aside even where only a trace of fabricated material  entered the process. Opperman J did not go that far — she reconsidered the section  18(3) application afresh rather than setting the order aside — but the direction of travel  elsewhere is instructive.

What the judiciary is already doing 

This is not being met with silence. The Gauteng Division's induction programme  already cautions acting judges against relying on AI for citations or unverified sources,  and the 2026 South African Judiciary Conference resolved to finalise a draft AI policy  defining permissible and prohibited judicial uses, with initial deployment limited to  efficiency tools such as transcription and case management. Notably, the resolutions  also commit to preserving the underlying legal research and reasoning skills judicial  work depends on — a reminder that the long-term risk is not only today's fabricated  case, but skill atrophy through disuse. 

The accountability gap 

A significant part of the judgment addresses a structural problem AI has simply  exposed, an acting judge's misconduct has no clear home. The JSC Act and Code of  Judicial Conduct extend to acting judges on paper, but the Judicial Service  Commission plays no role in appointing them, and section 177's removal mechanism  was built for permanent judges only. In practice, the only "sanction" has been to simply  not renew the appointment — no finding, no public record, no protection for future  litigants. Opperman J's answer is that an advocate or attorney appointed as an acting  judge never stops being an enrolled practitioner, so the Legal Practice Council's  disciplinary jurisdiction is continuous and survives the acting term. She held this to be  the most effective forum available, running concurrently with the JSC's, while making  clear that acting judges ought, as a matter of principle, to be accountable to the Commission too. That, she said, is the law as it should be — Parliament has not made  it so. 

Why it is worse when it is the judge 

There's a natural instinct to treat a judge's error more charitably than a practitioner's.  Opperman J takes the opposite view, when a practitioner cites fabricated authority,  opposing counsel and the judge are both there to catch it. When the fabrication  originates in the judgment itself, those filters have already been passed — there is no  scrutiny left to apply before the error becomes a binding order. 

Practical implications 

For practice, verification is non-delegable and doesn't end at your own filing — these  discrepancies were caught because a legal team checked every authority against the  judgment, then had the findings independently confirmed before raising them. An AI drafted paragraph and an AI-decided paragraph look identical from outside, which is  exactly why responsibility can never shift to the tool. And if you're the one raising an  issue like this, do it the way this case did — independent verification, agreement  between teams, and openness rather than private correspondence. The substantive  law here didn't change; what changed is who now has to answer for how a judgment  was reasoned. 

How we can help 

Cases like this one are a reminder that the law is moving quickly on AI, and that getting  sound, verified advice matters more than ever — whether you're navigating a dispute,  drafting a contract, or simply want peace of mind that your legal affairs are in capable  hands. At Gittins Attorneys Incorporated, our team combines rigorous legal research  with the pace and value our clients have come to expect. If you have a legal matter  you'd like to discuss, we'd welcome the opportunity to assist. Contact us today to  arrange a consultation.