Is using ChatGPT or another AI tool a breach of POPIA?

Not by itself. It becomes a problem when personal information goes into a tool abroad without a ground under section 72, without an operator agreement under section 21, or without telling people under section 18. Business accounts with a data processing agreement, clear rules on what may be pasted, and an updated privacy notice cover most of it.

Does section 72 apply if the AI provider is a large, reputable company?

Yes. Section 72 is about where the information goes and on what basis, not about the size of the recipient. A reputable provider makes it easier to show adequate protection, usually through its data processing terms, but you still need to have checked.

Can a medical practice use AI to summarise patient notes?

Health information is special personal information. Transferring it to a country without adequate protection needs prior authorisation from the Information Regulator under section 57. The safer routes are a provider whose agreement gives adequate protection, removing identifying details first, or running the AI on infrastructure the practice controls.

Insights · POPIA & AI

POPIA section 72 and AI tools: when a prompt becomes a cross-border transfer

Pasting a client email into an AI assistant feels like using a search engine. If the service runs outside South Africa and the email names a person, the law treats it as sending that person's information abroad.

November 2026 4 min read ASi Imperium engineering

Short answer: Section 72 of POPIA lets you send personal information to a third party in another country only on one of five grounds, the most practical being a binding agreement that gives adequate protection, or the person's consent. Most popular AI services process prompts outside South Africa, so a prompt that contains personal information is caught. It is allowed, but you stay responsible for it, and health, biometric and children's information carry stricter rules.

What section 72 actually says

A responsible party in South Africa may not transfer personal information about a data subject to a third party in a foreign country unless one of these applies:

The recipient is bound by a law, binding corporate rules or a binding agreement that gives an adequate level of protection, substantially similar to POPIA's conditions for lawful processing, including limits on passing the information on again.
The data subject consents to the transfer.
The transfer is necessary to perform a contract between the data subject and you, or for steps they asked you to take before a contract.
The transfer is necessary for a contract concluded in the data subject's interest between you and a third party.
The transfer benefits the data subject, it is not reasonably practicable to get their consent, and they would be likely to give it if asked.

The full text is in section 72 of the Act published by the Department of Justice. Nothing in it mentions AI. It does not need to: it covers any transfer, and an AI prompt is one.

Why AI tools fall inside it

Three things make an everyday AI tool a section 72 question:

The prompt leaves the country. Most widely used AI services run their models in data centres in the United States or Europe. A prompt, an uploaded file, a meeting recording or a transcript is processed there.
The prompt often names people. A client's email, a patient's note, a CV, a call transcript, a spreadsheet of debtors: all of it is personal information under POPIA.
The provider is processing it for you. In POPIA's terms the AI provider acts as your operator. Section 21 requires a written contract that makes the operator keep the security measures section 19 requires of you. A free or personal account normally comes with consumer terms, not an operator agreement.

Section 18 adds a transparency duty. When you collect personal information, people must be told, where it applies, that you intend to transfer it to another country and what protection it will have there. A privacy notice written before your staff started using AI tools probably says nothing about this.

Where the rules get stricter

Health, biometric and other special personal information

Section 26 lists special personal information: religious or philosophical beliefs, race or ethnic origin, trade union membership, political persuasion, health or sex life, biometric information and criminal behaviour. Section 57(1)(d) requires prior authorisation from the Information Regulator before you transfer special personal information, or children's personal information, to a third party in a country without adequate protection. For a medical practice summarising consultation notes in a consumer AI account, that is the provision to read twice.

Decisions made by the machine

Section 71 says a person may not be subject to a decision with legal consequences, or that affects them to a substantial degree, based solely on automated processing that profiles them, including their performance at work, creditworthiness, reliability, health or conduct. An AI tool that scores job applicants or loan applications with no person reviewing the outcome sits inside that section, wherever the model runs.

What to put in place

None of this bans AI. It asks you to know where the information goes and to be able to show you thought about it. In practice:

Find out what is in use. Ask staff which AI tools they use and under which account. Personal accounts used for work are the usual gap.
Move work onto business accounts. Business and enterprise plans generally come with a data processing agreement, state where data is processed and how long it is kept, and exclude your data from model training. Consumer terms often do not. Read the current terms of the account your staff actually use; they change.
Write down what may never be pasted. Identity numbers, health information, children's information, banking details and anything under legal privilege are a sensible starting list.
Strip names where the task allows it. Many drafting and summarising tasks work just as well on text with names and numbers removed.
Keep the sensitive layer at home. For confidential documents, AI can run on infrastructure you control, so the information never leaves. See private AI.
Update your privacy notice. Say which categories of information may be processed abroad by AI services, and on what basis.
Check what comes back. AI output can be confidently wrong. In April 2026 the national government withdrew its draft AI policy after its reference list was found to contain sources that did not exist (SAnews). A person should check anything AI writes before it is filed, published or sent to a client.

What this costs if it goes wrong

Under section 109 the Information Regulator can impose an administrative fine of up to R10 million. The more likely cost is the conversation with a client who learns their information was pasted into a personal chatbot account, and the time spent working out what else was.

This is an engineering firm's reading of the Act to help you ask the right questions. It is not legal advice. For an opinion on a specific transfer, speak to an attorney.