The usual AI horror story has counsel in the dock. This time the fake case was in the judgment.
We have all read the stories by now. Counsel files heads, the other side cannot find a case, the court goes looking, and a very uncomfortable costs hearing follows. Legal Twitter enjoys it. The rest of us quietly go and check our own footnotes.
FJL v TGO, decided by a full bench in Johannesburg on 31 July 2026, turns the story round. The wrong citations were not in anyone’s heads. They were in the judgment of the acting judge whose order was on appeal.
What happened
The facts are the familiar, sad stuff of the family court. Divorced parents, two children at a private school, a father who stopped paying the fees. He went to the High Court to have the children moved to a cheaper school. The mother counter-applied to keep them where they were. The acting judge found for the mother, and then ordered that his order operate while the father appealed. The father used his automatic right of appeal against that second order, and the full bench heard it as a matter of extreme urgency.
Counsel for the father argued that the acting judge had made errors of reasoning, including errors possibly caused by a misapplication of AI. Wright J (Mahosi J concurring) held that the order was correct and declined to “enter the lists” about AI (at [18]). Opperman J agreed with the order but got there by a different route, and it is her separate judgment that every advocate should read.
The table
Shortly before the hearing, the father’s team filed a note listing 11 discrepancies in the acting judge’s judgment. The mother’s team accepted that they were accurately described. So did the senior librarian of the Johannesburg Society of Advocates, who checked them (at [32]). Opperman J set them out in a table. It is uncomfortable reading, because the errors are so varied. A real Eastern Cape judgment, cited as a Johannesburg judgment with the wrong reference. A real SCA case cited under the citation of a different case. A Knoop judgment cited for prospects of success, when that Knoop judgment does not mention prospects of success and the other one does. Party initials muddled, with the citation belonging to a different case altogether. A passage presented as a quote from the Constitutional Court, which was in fact the Constitutional Court quoting an older High Court judgment, and one sentence in the quote that does not appear in that older judgment. And Lubbe v Volkswagon SA [2023] 4 All SA 289 (C), cited in five paragraphs of the judgment. It does not exist.
The judgment below recorded, in some places, that counsel for the mother had given the court these references. Both teams agreed she had not (at [32]). None of them was in her heads (at [34]). Spare a thought for junior counsel whose name appears in a judgment next to a case that does not exist.
Opperman J was careful here, and so should we be. She said the most plausible explanation for the fictitious Lubbe reference was an AI hallucination (at [35]). She made no finding that AI was used (at [37]). The acting judge has not been heard.
Why it matters
The line that has stayed with me is this one:
“Names of non-existent cases are not law.”Opperman J at [62]
The point she goes on to make is simple and, I think, the whole of the matter. Law has to be verifiable. Other people must be able to find it and check it. The lawyers must check their sources before they use them, and the public must be able to trust the court to do the same (at [62]).
She went further. Had she been able to conclude that the reasoning was machine-generated, and that the real decision-maker was a machine, she would have held that the father’s right to have his dispute decided by a court under section 34 of the Constitution had been infringed (at [66]).
In the end the appeal failed anyway. Opperman J put the fake and wrong authorities aside, applied the law as it actually is, and reached the same result (at [67] and [104]). The children stay at their school. But she sent the judgment to the Legal Practice Council for investigation, making clear that her remarks were observations to be investigated, not findings (at [122]).
What it means for the rest of us
Look at that table again. Only one of those errors is a case that does not exist. The rest are real cases with the wrong citation, the wrong court, the wrong quote, or cited for something they do not say. A quick “does this case exist?” check would have caught Lubbe and missed everything else.
That matches my own experience. When Gemini scrambled my quotes in 2025, every case it cited was real, and I had them all on my computer. What it invented were the words.
So the check has to go further than existence. Open the judgment. Find the passage. Compare it word for word. Read what comes before and after. Then ask whether the case says what you are citing it for. I have put that routine on the checking page.
One more detail from the judgment. The Johannesburg High Court already warns its acting judges, in their induction, to use AI with extreme circumspection and not to rely on it for citations or summaries of cases. And Opperman J noted that South Africa does not, yet, have its own guidelines (at [39]). That is part of why I have started drafting proposed guidelines for advocates. Comments are welcome.
Source on file: FJL v TGO (case no 2025/220239), High Court, Gauteng Division, Johannesburg, 31 July 2026 (Wright, Opperman and Mahosi JJ). Paragraph references are to the judgment as handed down.
Leave a comment