Artificial intelligence in a South African magistrates' court — uses, limits, and what must still be human
What has arrived, whether we asked or not
Officers are already using large language models to draft a first structure of reasons, to locate a section, or to summarise a long transcript. Practitioners are doing the same. The Department and the OCJ have spoken, in circular language, about digital transformation. None of that is a practice directive that tells a magistrate when a sentence generated by a machine may be spoken as the court's own. South Africa does not yet have a dedicated Artificial Intelligence Act. What we do have is the Constitution, the Promotion of Access to Information Act, the Protection of Personal Information Act 4 of 2013, the Electronic Communications and Transactions Act, the Legal Practice Act, and the Code of Judicial Conduct. POPIA still treats a charge sheet, a child's name and a complainant's address as personal information. Uploading a bundle to a public model is not research. It is a disclosure.
The case for careful use
Used as a clerk, not as a colleague, a model can shorten the hour before a reserved judgment finds its spine. It can list the section 60 factors, remind the court of the Children's Courts Rules that commenced in March, or produce a first chronology from a messy affidavit. For a rural court with a thin library and an uneven network, that is not a toy. It is a possible equaliser — if, and only if, every proposition is checked against the statute and the file. Translation tools, used with an interpreter and not instead of one, can help a court see whether a written statement matches what was said. Case-management dashboards can show which part-heards have gone quiet. None of that replaces the finding.
A file that lives in two places. The judgment must still have a name under it.
The case against unexamined use
Models invent citations. They flatten living customary law into a paragraph that sounds finished. They carry the bias of the data they were trained on — including older South African reporting that did not always see every person in the room. They cannot hear a complainant who has gone quiet, or see that an unrepresented accused has not understood the question. A court that signs a machine's fluency as its own reasons has not given reasons. It has given a font. There is also a Republic-shaped problem. Load-shedding, a portal that will not open, and a generator that has not been fuelled are already section 34 and section 165(4) issues. An AI strategy that assumes always-on connectivity is a strategy written for a different country. The present of the roll is still a person who walked to the building.
A working method
• Do not paste a charge sheet, a child's particulars, or an unredacted transcript into a public model. POPIA did not take Heritage Day off.
• If a tool helped you find a section, cite the section, not the tool. Check every case name before it leaves chambers.
• Reasons remain yours. If you cannot explain a paragraph without the machine, delete the paragraph.
• An admission from an unrepresented accused should still be slow, repeated, and written. No model should stand between that person and the record.
• Tell the senior if the court is being asked to rely on an AI-generated bundle the court itself cannot verify.
A note from Neelan Karikan. Authority stems from action. A machine can fetch a section. It cannot sit still in a frightened city, or write an amount as if a child might one day read the page. That work is still ours — and it is a form of living heritage.