FJL v TGO (GJ)[1] came before a full court in Johannesburg as an extremely urgent appeal under section 18(4) of the Superior Courts Act 10 of 2013, against the enforcement pending appeal of an order keeping two children in their private school. Shortly before the hearing, however, the appellant filed a supplementary note identifying 11 discrepancies[2] in the judgment of the acting judge below. Some real cases carried incorrect citations, some real citations contained quotations stitched together from other uncited cases and one authority, cited five times, did not exist at all. The senior librarian of the Johannesburg Society of Advocates confirmed the list of errors, as did a court researcher.
Opperman J, writing separately in concurrence, made no finding on how the errors arose. He held only that the most plausible explanation for the fictitious Lubbe reference “is that it is the product of the use of Artificial Intelligence (AI) and what has been dubbed ‘hallucinations’”.
The entire concurring judgment makes for fascinating and insightful reading, but it is paragraph 66 on which we wish to focus:
“Had I been able to conclude that the reasoning in the judgment under consideration was AI-generated and the true decision maker was a machine, I would not have hesitated to conclude that the appellant’s section 34 rights were infringed. To divest the decision-making responsibility to a robot, breaches not only the oath of office, which Acting Judges take, but it also denies a party the right to be heard by a Court as intended in terms of section 34.”
This is, as far as we can establish, the first South African judicial statement that adjudication by machine would be unconstitutional. The statement is obiter and therefore not binding precedent. However, in our view, it is likely to carry significant persuasive value in AI-related cases and should be ignored at an AI-user's own peril.
The statement raises important questions. Does section 34 of the Constitution[3] in fact contain a right to a human judge? And, if it does, how would a litigant ever enforce it? The answer to the first question, we suggest, is yes, based on a straightforward interpretation of our law. The answer to the second is where things get interesting.
Section 34 guarantees a fair public hearing “before a court or, where appropriate, another independent and impartial tribunal or forum”. Section 165(1) vests the judicial authority of the Republic in the courts. Section 174(1) then provides that “[a]ny appropriately qualified woman or man who is a fit and proper person may be appointed as a judicial officer” while Schedule 2 requires that person, on appointment, to swear to “administer justice to all persons alike without fear, favour or prejudice, in accordance with the Constitution and the law”. Our Constitution clearly prescribes that a judicial office is held by a person, under oath.
In Mphahlele[4] at paragraph 12, Goldstone J in the Constitutional Court held that the rule of law requires judges “not to act arbitrarily and to be accountable”, that furnishing reasons is “the manner in which they ordinarily account for their decisions” and that reason-giving “is a discipline which curbs arbitrary judicial decisions”. The Constitutional Court reaffirmed that discipline in Stuttafords[5] and then went further in Vodacom (Pty) Ltd v Makate[6], stating that “[t]he duty of proper consideration is an integral component of the fair hearing right” (paragraph 45).
If section 34 requires the court itself to do the considering, a judgment whose reasoning was produced elsewhere is constitutionally suspect. Makate concerned a court failing to consider material at all rather than the provenance of its reasoning, but Opperman J makes the same point in plainer language at paragraph 65: “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.”
The right cannot mean, however, that a judge may never use the technology. The judiciary has itself resolved to adopt AI for transcription, translation, case management and document summarisation, with a draft AI policy to follow (2026 Judiciary Conference resolutions, recorded at paragraphs 50 to 56 of the judgment). The practitioner cases Mavundla[7] and Northbound Processing[8] have established that legal work may be assisted by AI but that responsibility for it may not be delegated to AI.
But where does one draw this line? Is it with the research memorandum, the first draft or the model used, in Opperman J’s phrase, as a “verifier of last instance”? Where, precisely, does assistance end and adjudication begin?
Opperman J’s phrase “the true decision maker was a machine” is a useful starting point. A judgment would be machine-made not only where an AI model writes it but where the judge draws so heavily on the machine’s output that the consideration Makate requires happened somewhere other than in a judicial mind.
The difficulty is that the right to a human judge is difficult, if not impossible, to vindicate. Consider how the discrepancies in FJL v TGO came to light: not through some court-driven process but through the appellant's legal team auditing the citations and taking the results to the librarian. Yet even with the errors tabulated, confirmed and undisputed, Opperman J could reach only “the most plausible explanation” and expressly made no definitive finding on whether AI had been used.
The evidence needed for any stronger finding is out of reach. The confidentiality of judicial deliberation in South Africa is well established. No provision is made for disclosure in respect of a judge’s drafting process. Mphahlele tells us that judges account through their reasons rather than their processes. Judges do not disclose the contributions of their clerks or researchers in judgments. A litigant alleging machine adjudication must rebut the presumption articulated in SARFU[9] that judges will honour their oath of office, displaceable only by cogent evidence. Yet the most conspicuous place such evidence could come from, the judicial drafting process, is shielded from disclosure and scrutiny.
The judicial AI guidance for England and Wales, refreshed in October 2025,[10] records that judges are “not generally obliged to describe the research or preparatory work which may have been done in order to produce a judgment”, AI included. After all, is a language model not just another pair of hands? New Zealand’s guidance for judicial officers[11] is blunter, stating that “[y]ou do not need to disclose use of a GenAI chatbot.”
On the other hand, an AI model’s fabrications in the judicial context produce errors that present as law and so disable the judge’s ordinary review mechanism. Other jurisdictions appear to appreciate that the unfettered use of AI in the judicial context may compromise the rule of law. The Council of Europe’s December 2025 guidelines[12] state that parties “have the right to know which elements of a judicial resolution have been assisted by generative AI”. As far as we can tell, however, there is no binding instrument anywhere that obliges a judge to disclose the use of AI in producing a judgment.
South Africa may be about to become the exception. According to media reports on the Office of the Chief Justice’s draft AI policy,[13] judicial officers would be required to disclose any material contribution of generative AI to a court document, specifying the tool used and its contribution. If the reporting is accurate, provisions mandating disclosure would address the "unprovable suspicion". A court on appeal or review, the Judicial Conduct Committee or the Legal Practice Council, in the case of acting judges, would not need to reconstruct what happened in chambers but rather test what the judge said happened. The section 34 constitutional right emphasised by Opperman J in paragraph 66 of FJL v TGO would become more readily enforceable in the AI context once a disclosure obligation exists. Where no consequences attach to the undisclosed use of generative AI, however, the risk of hidden use persists. When hidden use comes to light, as it did in FJL v TGO, litigants, the public and the interests of justice will be served only where offending judges are readily held to account.
FJL v TGO makes no finding against the acting judge, who has not yet been heard. Opperman J was at pains to route the issue to the proper forum, in this case the Legal Practice Council, with the Judicial Service Commission enjoying concurrent jurisdiction, rather than address it directly in the judgment or privately.
Opperman J's judgment is nonetheless a powerful reminder that the section 34 right to be heard by a court requires that a person is listening. Absent a reasoned human judgment, what is the purpose of a judge at all?
[1] Unreported case no 2025/220239, 31-7-2026
[2] The Court’s list at para 32 lists twelve discrepancies
[3] Constitution of the Republic of South Africa Act 108 of 1996
[4] Mphahlele v First National Bank of SA Ltd 1999 (2) SA 667 (CC)
[5] Stuttafords Stores (Pty) Ltd v Salt of the Earth Creations (Pty) Ltd 2011 (1) SA 267 (CC) at paras 10-11.
[6] Vodacom (Pty) Ltd v Makate and Another 2025 (6) SA 352 (CC)
[7] Mavundla v MEC: Department of Co-Operative Government and Traditional Affairs, KwaZulu-Natal 2025 (3) SA 534 (KZP)
[8] Northbound Processing (Pty) Ltd v South African Diamond and Precious Metals Regulator [2025] ZAGPJHC 661
[9] President of the Republic of South Africa v South African Rugby Football Union 1999 (4) SA 147 (CC) at para 40
[10] Courts and Tribunals Judiciary Artificial Intelligence (AI) Guidance for Judicial Office Holders 31 October 2025, available at https://www.judiciary.uk/wp-content/uploads/2025/10/Artificial-Intelligence-AI-Guidance-for-Judicial-Office-Holders-2.pdf
[11] Courts of New Zealand Guidelines for Use of Generative Artificial Intelligence in Courts and Tribunals 7 December 2023, available at https://www.courtsofnz.govt.nz/assets/6-Going-to-Court/practice-directions/practice-guidelines/all-benches/20231207-GenAI-Guidelines-Judicial.pdf
[12] The Council of Europe's European Commission for the Efficiency of Justice (CEPEJ) Guidelines on the use of Generative Artificial Intelligence for Courts 19 December 2025, available at https://rm.coe.int/cepej-2025-18final-en-draft-guidelines-on-the-use-of-generative-ai-for/48802a4ad1
[13] The draft itself has not been published
