South Africa does not need a new statute before AI that processes personal information is regulated. The Protection of Personal Information Act 4 of 2013 already applies. If an identifiable person can be singled out from what goes into a model, what comes out of it, or what is kept in logs, prompts or vendor systems, the eight conditions for lawful processing apply.
There is still no dedicated South African AI Act. King V (released 31 October 2025; apply-and-explain for financial years beginning on or after 1 January 2026) tells governing bodies to oversee data, information, technology and emerging technologies, including AI. That is oversight of the organisation. It is not a substitute for POPIA, and it is not a licence to treat a software platform as "governance."
This article is the local baseline. Later pieces cover incomplete application of POPIA, literacy, cross-border flows, the EU AI Act, banks with African operations, and when a European quality management system becomes relevant.
POPIA applies to the processing of personal information entered in a record by or for a responsible party who is domiciled in the Republic, or who is not domiciled here but uses means in the Republic (unless those means are used only to forward information through the Republic). Processing includes collection, receipt, recording, organisation, storage, updating, retrieval, use, dissemination, merging and destruction.
AI does not escape that list by being called inference or "the model." Typical processing includes:
The responsible party is the one who determines the purpose of the processing. That is usually the South African company, bank, organ of state or practice that decided to use the tool. The vendor is often an operator. Getting that split wrong is how organisations file the duty on the wrong desk.
The responsible party must ensure the conditions are met. "The model decided" is not an answer. Neither is "the vendor's terms allow it" if you determined the purpose.
Personal information must be processed lawfully and in a reasonable manner that does not unreasonably infringe privacy. Collection should be from the data subject where practicable. Minimisation still applies: a hiring model cannot process personal information without a valid legal basis.
The purpose must be specific, explicitly defined and lawful. "Analytics" or "AI improvement" is not a purpose until it is specified. Further processing of a service archive to train or fine-tune a model is a new purpose. It needs its own analysis.
Further processing must be compatible with the original purpose, subject to the Act's tests. Compatibility is not assumed because the same vendor holds the data.
Information must be complete, accurate, not misleading and updated where necessary. A score treated as a fact about a person is a quality problem if it is an inference the organisation cannot explain or correct.
The responsible party must take reasonably practicable steps to ensure the data subject is aware of the purpose, the information collected, the fact of an operator, and intended transfers. A notice that never mentions ranking, profiling or an offshore model does not describe the processing that is actually happening.
Section 19 requires reasonable technical and organisational measures against loss, damage, unauthorised destruction and unlawful access. Those measures include identifying reasonably foreseeable risks, putting safeguards in place, verifying them, and updating them. For AI, the assets include training sets, prompts, embeddings, vendor logs and model-access accounts - not only the on-premise server. Operators must be bound in writing to maintain security measures.
Access, correction and related rights apply to information the organisation holds, including, where they are personal information, scores and inferences. If the information sits only in a vendor cloud, the responsible party still has to give effect to the right.
Section 71 adds a further control where automated decision-making has legal or other substantial consequences, subject to the Act's qualifications. Ranking or auto-rejecting people can fall into that territory. A human who only ticks the model's list is not a safeguard.
Special personal information and children's information keep their extra restrictions. An inference that reveals health, biometric data or a child's information does not become ordinary data because a model produced it.
Every public and private body has an Information Officer. For a private body that is the head, or a person duly authorised. Registration with the Information Regulator is required. The role is not a title on a website.
Regulation 4(a) requires the Information Officer to ensure a compliance framework is developed, implemented, monitored and maintained. That framework includes governance arrangements for compliance, assignment of roles, the legal and organisational context, an inventory of personal information, documentation of processing, identification of outsourced processing, minimisation, risk assessment and treatment.
Regulation 4(b) requires a personal information impact assessment so that adequate measures exist to comply with the conditions. A PIIA for AI processing has to look at inferences, operators, training use, rights-handling and residual risk to data subjects - not only at encryption.
If those two duties are not done, later talk of "AI programmes" has no local legal floor.
An operator processes personal information for the responsible party. The operator may do so only with the responsible party's knowledge or authorisation, and must treat the information as confidential. The written contract must require the operator to establish and maintain Section 19 security measures.
An AI platform that hosts prompts, fine-tunes on customer tickets, or sends data to subprocessors is inside that definition whenever it processes for your purpose. If it also determines its own purposes - for example using your data to improve its own independent product - it may be a responsible party for that further processing. The assessment has to say which.
Cross-border transfers are Chapter 9. They are the subject of a later article. The baseline point is only this: putting personal information into a model hosted outside the Republic is a transfer analysis, not a procurement footnote.
It does not ban AI.
It does not make every spreadsheet an AI system.
It does not, by itself, impose EU high-risk duties or require EN 18286.
It does not move residual accountability from the responsible party to a dashboard.
It does mean that a South African organisation already has a statutory file to complete before it scales scoring, generative tools or vendor models on personal information: purpose, minimisation, notices, operator contract, security, rights path, PIIA, and an Information Officer who can show the framework is alive.
King V expects the governing body to oversee data, information and technology so that the organisation can pursue its strategy without ignoring ethics, privacy, security and accountability for automated outcomes. Human oversight and override, proportionate to risk, sit in that expectation.
The board evaluates, directs and oversees what the organisation may do with AI that affects people and information. Management directs how that is delivered: inventory, contracts, PIIAs, competence, monitoring. A platform can hold the records. It cannot sit in the board's chair, and it cannot discharge POPIA.