On 15 October 2026, Acting Judge Sean Snyman of the Johannesburg Labour Court stayed a CCMA arbitration. The reasons were handed down on 23 October. They record that the arbitrator, Commissioner Menzi Fakude, had based a jurisdictional ruling on a Constitutional Court authority that does not exist. The citation, '[2023] ZACC 40', points to a different case, one about immigration and spousal visas, according to GroundUp's report. Snyman wrote that it appeared Fakude 'must have relied on AI' to substantiate his reasoning and never checked it.
This is the second South African judicial-decision failure of this kind in three months. It points to a policy conclusion that is narrower than either 'ban AI' or 'carry on regardless'.
What the court actually decided
The applicants were the South African Post Office, which is in business rescue, and its two business rescue practitioners. Six former employees had been dismissed, and the CCMA was due to arbitrate their unfair dismissal claims on 15 October. The applicants had received Fakude's ruling on 5 August 2026. In it he held that business rescue did not stop the arbitration going ahead. The practitioners filed a review application on 2 October and sought an urgent stay.
The reasons are a more layered judgment than the headlines suggest. The court's summary records that the moratorium in section 133(1) of the Companies Act applies to unfair dismissal proceedings under the Labour Relations Act. Absent the practitioners' consent or leave of the High Court, the CCMA could not proceed. It also records that the business rescue practitioners had not been joined, which was a further error of law. Only after that does it hold that basing the decision on 'non-existent Constitutional Court authority' without verifying the research is material irregularity and misconduct justifying immediate intervention.
The order is a stay pending review. It is not a final setting-aside. GroundUp reports that Snyman said the workers could pursue their claims once the Post Office exits business rescue. The fabricated citation was not the only defect. But it is the one that exposed a decision-maker who had not read the authority he called decisive.
This is not a one-off
On 31 July 2026, a Full Bench of the Gauteng Division in Johannesburg dismissed an appeal in a children's school-fees dispute. The judgment records as a headline issue the 'consequences of discrepancies in cited judgments' by the acting judge below, and which body should receive a report, the Judicial Conduct Committee or the Legal Practice Council. GroundUp reported that Judge Ingrid Opperman detailed at least 11 fictitious or inaccurate citations. She wrote that she made no finding on whether AI was used, and that 'jumbled together quotes from cases, some identified, some not, are not law'. She said she would send her judgment to the Legal Practice Council. The judiciary's own AI policy is still in draft.
Snyman's judgment begins by noting that the Labour Court had already faced the problem in Molawa v Matjhabeng Local Municipality, where a disciplinary chairperson relied on unverified citations. The warnings about fabricated authority have so far been addressed mostly to lawyers. Two of these three episodes involve decision-makers.
The strongest case for tighter rules
The argument for prohibition deserves a fair statement. A judgment or arbitration award is an exercise of public power. The parties cannot see how it was produced, and they cannot cross-examine the research. If a decision-maker can outsource reasoning to a system that invents authority, the legitimacy of the decision is in doubt. And if the invented authority is cited to a party who cannot check it, that party is harmed. A categorical rule against generative AI in adjudication is easy to state, easy to police and easy to explain to the public.
Why a duty to verify is the better rule
The judgments themselves point somewhere more precise. Neither Fakude's ruling nor the acting judge's was undermined by the use of a tool. Each was undermined by the absence of a check that a competent reader would perform in minutes: pull the judgment and read it. Snyman put the standard directly. An arbitrator who uses AI for research is compelled to verify that the judgment exists and to read it to confirm that it decided what the AI says it decided.
That suggests three design principles for the pending judiciary policy.
- Regulate the output, not the tool. A rule that every authority cited in reasons must have been personally read is enforceable. It also catches the identical failure when the source is a hurried clerk, a bad textbook or a misremembered case. Opperman's remark that she made no finding on AI use shows why this matters: the harm is the same whatever the cause.
- Separate administrative from adjudicative use. Transcription, scheduling and summarising bundles carry little risk of invented law. Drafting reasons carries a lot. A blanket ban would forgo efficiency gains on the first to prevent harms in the second, in a system whose backlogs are well documented.
- Reach the quasi-judicial layer. The CCMA, disciplinary chairpersons and other tribunals decide very large numbers of disputes. A policy written only for High Court judges would miss the setting where this most recent failure occurred.
The draft policy has not been published, so its content cannot be judged. What can be said is that the cases to date show a failure of professional method. They do not show a failure of technology that only prohibition can solve.
What to watch
Two questions matter more than any single sanction. The first is whether the Legal Practice Council's concurrent jurisdiction, which Opperman described, produces a clear accountability route for acting judges, whom the Judicial Service Commission does not appoint. The second is whether the CCMA, which was not itself a party to the fault but whose commissioner was, issues internal verification rules before the judiciary's policy arrives. Both can be done now, without waiting for legislation and without restricting the research tools that make small tribunals more productive.
The lesson of October 2026 is that adjudicators are answerable for every authority in their reasons, however it was found. Policy that says so plainly, and attaches consequences, protects litigants without banning a useful technology.