POPIA and AI, a practical guide for South African organisations

South Africa has no AI-specific law. The Draft National AI Policy published on 10 April 2026 was withdrawn on 26 April 2026, and an expert review panel is helping the Department of Communications and Digital Technologies redraft it. Until a new policy and any law follow, the Protection of Personal Information Act 4 of 2013 is the main binding rule for AI that uses personal information. Section 71 on automated decisions gets most of the attention and is covered in full on the POPIA section 71 page. This guide covers the rest of POPIA, which reaches every AI system that touches personal information.

Know which AI uses personal information

POPIA’s conditions apply to processing, so the first job is knowing where AI processes personal information. That covers models you build, AI features inside software you already use, and generative AI tools staff adopt on their own. For each, record the purpose, the personal information it uses and where it comes from, whether special personal information or children’s information is involved, where the processing happens, and whether the system makes or shapes decisions about people.

The AI system register guide covers what each entry should hold, and the shadow AI guide covers how to find the tools nobody told you about.

Check automated decisions against section 71

For each system that decides things about people, ask whether the decision has legal consequences or a substantial effect, whether it is based solely on automated processing and whether that processing is intended to build a profile. If all three are true, the decision is only allowed under a section 71(2) exception. Where the contract exception is used, section 71(3) requires the person to be able to make representations and to receive enough information about the underlying logic to do so.

Give notice that covers the AI

Section 18 requires a responsible party to take reasonably practicable steps to make sure the data subject knows what information is collected, the source where it is not collected from them, the purpose, and whether it will be transferred to a third country and with what protection. An AI system that uses personal information for a new purpose, or sends it to an offshore model provider, can make an existing privacy notice inaccurate. Review notices when AI is introduced, not only when the law changes.

Secure the AI and bind the vendors

Section 19 requires appropriate, reasonable technical and organisational measures to secure personal information. That means identifying reasonably foreseeable internal and external risks, keeping safeguards against them and updating them in response to new risks. For AI, foreseeable risks include personal information leaking through model outputs, prompt injection, training data exposure and staff pasting information into public tools.

Most AI is bought, so section 21 matters as much as section 19. Where a vendor processes personal information for you, it is an operator, and a written contract must require it to maintain the section 19 measures. The operator must tell you immediately if it has reason to believe personal information was accessed by an unauthorised person. The vendor AI risk guide covers what to ask an AI supplier.

Check prior authorisation and transfers

Section 57 requires prior authorisation from the Information Regulator before some kinds of processing, including processing unique identifiers for a new purpose with the aim of linking them with other parties’ information, processing information on criminal behaviour or unlawful conduct for third parties, credit reporting, and transferring special personal information or children’s information to a country without adequate protection. An AI fraud or credit model may fall into one of those categories.

Section 72 governs transfers outside South Africa, which most hosted AI services involve. A transfer is allowed where the recipient is bound by a law, binding corporate rules or binding agreement that gives adequate protection substantially similar to POPIA, where the data subject consents, where it is necessary for a contract with the data subject or in their interest, or where it benefits them and consent is impractical.

Report security compromises

Section 22 requires the responsible party to notify the Information Regulator and the data subject as soon as reasonably possible where there are reasonable grounds to believe personal information has been accessed or acquired by an unauthorised person. Since 1 April 2025 the Regulator has taken security compromise reports through its eServices portal. An AI incident that exposes personal information is a security compromise, so AI incidents need a route into the breach process. The AI incident guide covers how to run one.

How Aicura supports the work

Aicura is an AI governance platform. Discovery and the AI register build the inventory, data provenance traces which information feeds which system, assessments record purpose, people affected and risks, vendor due diligence covers operators, and incident tracking covers security compromises. Aicura does not assess against POPIA. It gives you the record POPIA’s conditions depend on, and the decisions stay with you.

Aicura is hosted on AWS in Sydney, Australia, so using it involves transferring personal information out of South Africa. Section 72 governs that transfer, and our data processing agreement sets out the binding safeguards that cover it.

A note on this page

This is Aicura’s reading of POPIA as it applies to AI. It is not legal advice and it is not the Act itself. Read the Act as published by the Information Regulator, and take advice on how it applies to your organisation.