Home / Insights / ChatGPT and POPIA

ChatGPT and POPIA: what your team can and can't paste.

Pasting information about a person into an AI tool is "processing" under POPIA, and sending it to an overseas provider is a cross-border transfer. Neither is banned. Both come with conditions. Here is the plain-language version.

The thirty-second version

POPIA never mentions AI, but it governs personal information wherever that information goes. If a prompt contains information about an identifiable person, pasting it into an AI tool is processing under the Act, and your business stays responsible for it. Whether that processing is lawful depends on what the information is, which tier of the tool you use, and the safeguards around it.

What counts as personal information

More than most people think. Under POPIA it includes a person's name together with contact details, ID and account numbers, employment and salary details, location data, correspondence, and even opinions or assessments about a person. Two categories carry stricter rules: special personal information (health, biometrics, race, religion, among others) and children's information. If a prompt would let someone identify a real person, assume POPIA applies to it.

The traffic-light test

This is the test we put in client AI policies, and it covers almost every everyday case.

Green: paste freely

  • Your own drafting with no personal information in it: proposals, emails, marketing copy with names removed
  • General questions ("how does a letter of demand work in South Africa?")
  • Public information, published prices, your own website copy
  • Anonymised patterns: "a client in logistics with 40 drivers", not the client's name

Amber: depends on your tool tier

  • Internal business information with no personal information in it: strategy notes, draft pricing, process documentation
  • Not a POPIA question, but a confidentiality one. Fine on a business tier with data processing terms; unwise on a free consumer account

Red: never, on any tier, without specific safeguards

  • Client or staff personal information: names with contact details, ID numbers, account numbers
  • Payroll, HR records, disciplinary matters
  • Health or other special personal information, and anything about children
  • Passwords, access keys, and anything under an NDA

Why the tool's tier matters

The same product behaves differently at different tiers, and the difference is contractual. Ask two questions of any AI tool: does the provider train its models on what your staff paste, and what does your contract say about your data?

Business tiers of the major tools (ChatGPT Team and Enterprise, Claude for Work, Copilot for Microsoft 365) commit by default not to train on your business data and give you admin controls and data processing terms. Consumer tiers vary, their settings change, and a free account creates no contract between the provider and your business at all. That is why our policy template approves tools by name and tier, not by brand.

The overseas question

Almost every AI provider processes data outside South Africa, which makes a prompt containing personal information a cross-border transfer under section 72 of POPIA. Such transfers are permitted on several grounds; the one most businesses rely on in practice is that the overseas recipient is bound by an agreement that upholds protection comparable to POPIA's conditions. A business-tier contract with data processing terms is how you get that agreement. A free consumer account gives you nothing to point to.

One thing does not change with geography: your business remains the responsible party. The provider processes on your behalf; accountability stays home.

Five practical rules

  1. Put the red list in writing and make it the hardest rule in your AI use policy.
  2. Buy a business tier for anyone using AI on real work. The cost difference is small; the contractual difference is the whole point.
  3. Strip identifying details before pasting when a real example is needed.
  4. Keep consumer accounts for general questions only, never client information.
  5. Make honest reporting safe, so a wrong paste surfaces the same day.

This article is practical guidance from implementation work, not legal advice. For a formal opinion on your POPIA position, speak to a privacy attorney. The Information Regulator publishes guidance at inforegulator.org.za.

Common questions

Is it illegal to paste client information into ChatGPT?

Not automatically. It is processing under POPIA, and processing is lawful only when it meets the Act's conditions: a proper purpose, minimality, security safeguards, and a lawful basis for any cross-border transfer. On a consumer tool with no contract and no safeguards, you are unlikely to meet those conditions. On a business tier with data processing terms, you often can.

Is the free version of ChatGPT safe for business use?

Treat consumer tiers as unsafe for personal or confidential information by default. Their data controls vary and change, and they come with no contract between the provider and your business. They are fine for general questions and drafting that contains no personal or confidential information.

Does POPIA apply if the AI company is overseas?

Yes. Your business remains the responsible party regardless of where the tool runs, and sending personal information to an overseas provider is a section 72 cross-border transfer. The practical mechanism most businesses rely on is a business-tier contract in which the provider commits to comparable protection.

Related reading

Want this handled properly, once?

The AI Adoption program sets up the policy, the right tool tiers and contracts, and a working first pilot, so your team gets the benefit without the exposure.

Book a call if you want one, or just send the form. We reply within one business day. POPIA-compliant.

Explore AI Adoption →