The South African Labour Court has stayed disciplinary proceedings against three employees of the Matjhabeng Local Municipality after lawyers for the employees presented prima facie evidence that the chairperson of the hearing may have relied on hallucinated AI-generated legal authorities when making key rulings.

The employees’ lawyer showed that, apart from one cited authority, the legal sources referenced by the hearing chairperson in a June ruling either did not exist or said nothing close to what the chairperson claimed they had said. The challenge targeted two rulings made by the chairperson and raised serious questions about the integrity of the disciplinary process.

In her judgment, Judge K Allen-Yaman stopped short of finding that artificial intelligence had written the chairperson’s reasoning. However, she warned that if an AI tool had generated the citations, “the ruling cannot stand,” and said blind reliance on a chatbot would be unacceptable under standards already set by South African courts. The term “hallucination” in this context refers to the tendency of large language models to generate fluent, confident-sounding text — including case names, citations, and legal principles — that is fabricated, as the models predict plausible language rather than retrieve verified facts.

The case comes barely three months after South Africa’s Department of Communications and Digital Technologies was forced to withdraw its Draft National Artificial Intelligence Policy after News24 found that a portion of its 67 academic references were fabricated or pointed to journals that had not published the cited work. The document had already been approved by Cabinet and gazetted for public comment before the accuracy of the scholarship supporting it was checked. Two senior officials in the department were later named in a Rapport investigation and given until the end of April to provide explanations.

The episode was not the only recent citation controversy involving South African government institutions. The Department of Home Affairs faced a similar issue not long before, after journalists identified more than a hundred invented references in the reference list of the Revised White Paper on Citizenship, Immigration and Refugees. The department argued that the fabricated citations were limited to a standalone list and did not affect the substance of the policy. That defence echoed the argument made in the AI policy case: incorrect sources, but correct conclusions.

South Africa’s judiciary has been confronting the same problem since at least January 2025, when the KwaZulu-Natal High Court’s judgment in Mavundla v MEC for Co-Operative Government and Traditional Affairs exposed heads of argument drafted with ChatGPT by a candidate attorney. The submission included an invented case, “Hassan v Coetzee,” based on a legal principle that appeared plausible but was entirely fabricated.

Since then, the Johannesburg High Court has referred lawyers to the Legal Practice Council for using fake legal authorities in Northbound Processing v South African Diamond and Precious Metals Regulator. In that matter, one judge demonstrated in court how easily ChatGPT could invent cases by running live prompts.

South Africa faces repeated AI citation scandals

The cases do not point only to isolated recklessness by individual practitioners or officials. They also reflect a broader institutional failure: relying on tools that cannot be trusted to verify the accuracy of the material they generate. Generative AI models are designed to produce fluent and confident text, not necessarily accurate text. A fabricated Constitutional Court judgment can read like a real one until someone attempts to locate and verify it.

South African courts and government departments are discovering, through repeated scandals, that sounding correct and being correct are different things. The Legal Practice Council and the Labour Court have both signalled a strict approach to the use of false authorities, and the Matjhabeng case shows that the problem has moved beyond junior lawyers working under deadline pressure.

The issue now extends into quasi-judicial decision-making, where rulings can affect workers’ jobs and livelihoods. If a disciplinary chairperson can rely on a hallucinated citation to justify a ruling involving three employees, the same risk may arise anywhere a public official is expected to exercise independent judgment rather than reproduce output generated by a chatbot.

International data indicates that the risk is increasing. Damien Charlotin’s database has recorded more than 1,300 confirmed AI-hallucination cases in courts worldwide: One of the earliest and most widely reported was the 2023 US case Mata v. Avianca, in which a federal judge in New York sanctioned two lawyers who submitted a legal brief containing six non-existent cases generated by ChatGPT, one of whom had used the tool to verify citations the other had produced. Research published in Nature found that the share of academic papers containing at least one fabricated citation rose roughly ninefold between 2024 and 2025.

South Africa is not an outlier in the wider problem of AI-generated false citations. It is one of the clearest examples of how quickly institutional trust in AI-assisted work can move ahead of institutions’ ability to verify it.