South Africa's Office of the Chief Justice (OCJ) has released a Draft Policy on Artificial Intelligence for the Judiciary that draws one of the clearest lines yet between what AI may and may not do inside a courtroom. The draft, circulated for comment in 2026, expressly prohibits AI from performing any adjudicative function — evaluating evidence, assessing witness credibility, interpreting or applying the law, generating substantive legal reasoning, or determining bail, sentence, or risk. AI is confined to administrative and support tasks — scheduling, case-flow management, document summarisation, transcription, translation, and formatting — and even then only under human verification (IOL).
The Problem the Policy Is Answering
This is not abstract risk management. South African courts have already been burned. In Mavundla v MEC: Department of Co-Operative Government and Traditional Affairs KwaZulu-Natal [2025] ZAKZPHC 2, the KwaZulu-Natal High Court found that seven of the nine authorities cited by counsel were entirely fabricated by generative AI — including a non-existent Pieterse v The Public Protector. Judge Bezuidenhout called the conduct "irresponsible and downright unprofessional" and referred the practitioners to the Legal Practice Council (Mondaq).
Months later, in Northbound Processing (Pty) Ltd v South African Diamond and Precious Metals Regulator (Gauteng High Court, 30 June 2025), Acting Judge Smit discovered fictitious citations generated by a tool called "Legal Genius." Even though counsel admitted the error and apologised unreservedly, the court issued a mandatory referral to the Legal Practice Council, warning that "save in exceptional circumstances, admonishment alone is unlikely to be a sufficient response" (Cliffe Dekker Hofmeyr). Two cases, two referrals, one unmistakable judicial message: AI output is not authority, and persuasive writing is no substitute for verification.
The Strongest Case for the Ban
The case for prohibiting AI from adjudication is serious and deserves to be stated plainly. Judicial reasoning is not a text-prediction task; it is the exercise of constitutional discretion by an accountable human officer. As District Court Magistrate Lizelle Cloete argued in De Rebus (June 2026), generative systems risk producing "standardised outputs" that narrow independent thinking, and many run as proprietary "black boxes" whose training data can replicate the very inequities a constitutional court exists to remedy (De Rebus). In criminal justice, an opaque risk-assessment model determining bail or sentence would offend the rule of law before it ever produced a wrong answer. A litigant has a right to a decision authored by a judge who can be questioned, appealed, and held to account — not by a model that cannot explain itself. On that, the OCJ is right.
Why the Draft Is Proportionate, Not Reflexive
What makes this policy commendable rather than merely cautious is that it bans a function, not a technology. It does not prohibit judges from using AI; it prohibits AI from being the judge. Administrative drudgery — transcription, translation across South Africa's eleven official languages, summarising voluminous records — is precisely where AI can expand access to justice in an under-resourced system, and the draft explicitly invites that use to "reduce delays, lower administrative costs, and expand access to judicial services."
The draft also avoids the trap of static rules. It imposes an explainability requirement — any deployed system must produce intelligible, scrutable outputs, with non-compliant systems suspended — and demands that tools be tested for racial, gender, linguistic and socio-economic bias and piloted before deployment. It mandates disclosure of any material AI contribution to court documents, requires POPIA compliance for systems touching sensitive case data (POPIA), and commits to review at least every two years. That is the architecture of proportionate regulation: a firm prohibition on the one thing that cannot be delegated, paired with conditional permission for everything that can.
Where the Risk Now Lies
The danger going forward is not the prohibition but its boundaries. "Substantive legal reasoning" is easy to ban and hard to define. If a judge uses an AI tool to summarise a 400-page record, and the summary silently omits the one paragraph that decides the case, the line between "administrative support" and "adjudicative influence" has been crossed invisibly. Human-in-the-loop verification only works if the human can realistically catch the machine's error — and summarisation tools are seductive precisely because they save the reader from reading. The policy's explainability and disclosure requirements are the right instruments, but they will need teeth and training, not just text.
South Africa is also moving on a parallel national track: the Department of Communications and Digital Technologies opened its broader Draft National AI Policy for public comment in April 2026 (Baker McKenzie). The judiciary's decision to write its own rules — developed by the bench rather than imposed on it — is the correct instinct. Courts understand judicial independence better than any line department, and a profession-led standard is more likely to be respected than an external mandate.
The OCJ has done what good regulators rarely manage: respond to a genuine, demonstrated harm with a rule narrow enough to preserve the upside. The fabricated-citation cases proved the floor was needed. The challenge now is to police the grey zone without smothering the administrative gains that make AI worth having in a strained court system at all.