Whether ChatGPT is safe for your company's data depends mostly on the account your staff use. On a personal Free or Plus account, OpenAI can train its models on what an employee types, and your company has no contract with OpenAI. Under POPIA, your company carries that risk. On a company plan with a signed data processing agreement, used by staff who know what may go into it, the risk is much lower and you're in a position you can defend.
The risky case rarely starts with a decision anyone made. Let's say someone in finance pastes the debtors' age analysis into ChatGPT to draft reminder emails. Or someone in HR asks it to tidy up a disciplinary letter with the employee's name and ID number still in it. Both are trying to save half an hour, and in both cases personal information has left your company through an account IT has never seen.
This is already common in large South African companies, and World Wide Worx's SA AI Index 2026 found that 93% of them use generative AI and 75% report some "shadow AI" (tools staff use without the company's approval).
I'm not a lawyer and this article isn't legal advice; for your own situation, speak to your information officer or a POPIA attorney.
Is it safe to put company data in ChatGPT?
The account matters more than the brand. OpenAI treats personal and company plans differently on training, on how long it keeps conversations and on who can see them, as the table shows.
What OpenAI says it does with your data, by plan (checked 20 November 2026)
| Plan | Trains OpenAI's models? | How long it's kept | Who else can see it | Contract with your company |
|---|---|---|---|---|
| Personal (Free, Plus, Pro) | Yes by default; the user can switch it off | Until deleted, then gone within 30 days (security and legal exceptions) | OpenAI, under its own privacy policy | None; OpenAI's agreement is with the employee |
| Temporary chat | No | Up to 30 days, for safety | OpenAI | As for the account |
| ChatGPT Business | No, by default | Set by your admins | Your admins can view, export and delete chats | Data processing addendum |
| ChatGPT Enterprise | No, by default | Set by your admins | Your admins, through an audit log | Data processing addendum |
| API | No, unless you opt in | Up to 30 days for abuse monitoring | OpenAI, for abuse monitoring | Data processing addendum |
Switching training off on a personal account doesn't cover everything. Even with it off, a thumbs up or down on an answer can send the whole conversation for training, and switching it off doesn't delete anything already in the history.
On a Business plan, the company can read what staff type. According to OpenAI's enterprise privacy page, admins "can view, access, export, and delete end user conversations".
Microsoft follows the same pattern. Copilot on a personal Microsoft account trains on conversations unless you opt out, while Copilot on a work account falls under Microsoft's data protection addendum and doesn't train on prompts.
Google does too. The personal Gemini app lets human reviewers read a sample of chats, while Gemini in Google Workspace isn't reviewed or used for training outside your domain. With all three vendors, the exposure comes from staff doing company work on a personal account.
Does ChatGPT leak sensitive data?
ChatGPT has had exposures, but none of the known ones came from the model itself. In March 2023 a bug let some users see the titles of other users' chats and exposed payment details of 1.2% of Plus subscribers. In 2025 an opt-in setting let search engines index thousands of shared chats, until OpenAI removed it in August. That November an attacker broke into Mixpanel, an analytics supplier, and took names and email addresses of some users; OpenAI says no chats or passwords were exposed.
The best-known case, at Samsung, involved no flaw in ChatGPT at all. In April 2023 its engineers uploaded internal source code to ChatGPT. Samsung then banned generative AI on company devices, with a memo warning that breaking the rule could lead to dismissal. The code left the company because someone chose to paste it in, which is where most of the risk to your data sits.
What does POPIA expect of the employer when staff use AI?
POPIA puts the risk on the company, whichever tool the employee picked. The Act calls whoever decides why and how personal information is processed the responsible party, and for records handled at work, that's the employer. Pasting a client's details into a chatbot counts as processing. POPIA also protects information about identifiable companies, so a client business's details are covered as well as a person's.
Section 19 requires the responsible party to secure personal information through "appropriate, reasonable technical and organisational measures". That includes identifying the risks it can reasonably foresee inside the company. An employee pasting client records into a free chatbot is about as foreseeable as an internal risk gets in 2026.
Section 20 is the one that speaks to staff. Anyone processing personal information for the employer may do so "only with the knowledge or authorisation of the responsible party", and must keep it confidential. An employee who pastes client data into a personal ChatGPT account without permission is working outside that authority.
Section 21 asks for a written contract with anyone who processes personal information for you (POPIA calls them an operator). A personal ChatGPT account, signed up with a Gmail address and paid for on someone's own card, has no contract with your company at all. A Business or Enterprise plan with OpenAI's data processing addendum is the kind of contract the section has in mind; whether it's enough for your data is a question for your lawyer.
When personal information reaches someone who shouldn't have it, section 22 requires you to notify the Information Regulator and the people affected as soon as reasonably possible. The Regulator hasn't published a guidance note on AI, so these sections are what employers have to work from.
Section 109 allows fines of up to R10 million for failing these duties, and under section 99 a person can sue for damages "whether or not there is intent or negligence". In July 2023 the Regulator fined the Department of Justice R5 million for not complying with an enforcement notice issued after a breach linked to lapsed antivirus licences.
Is your data leaving South Africa when staff use ChatGPT?
It is, on every plan. For South African users, OpenAI's privacy policy names OpenAI OpCo, LLC in San Francisco as the controller, and says data is stored in the United States and other countries. OpenAI lets Enterprise and API customers choose where some data is stored, but when it announced those regions in November 2025, South Africa wasn't among them.
Section 72 allows personal information to go to a third party abroad only under set conditions. The two that matter here are a binding agreement (or law) giving substantially similar protection, or the consent of the person the information is about. A debtors' list pasted into a personal account is unlikely to meet either, because nobody asked the debtors and the only agreement is between OpenAI and the employee. On a company plan, the data processing addendum is that agreement, and your lawyer can judge whether it protects your data well enough.
For a company that wants its data to stay in the country, CBOS builds private AI systems where the data is stored and processed in South Africa, so section 72 doesn't come into it.
Can my employer see if I use ChatGPT, and can I be fired for it?
On a company account, yes. ChatGPT Business lets admins view and export conversations, and Enterprise gives them an audit log. On a personal account your employer can't open your history, but company laptops and networks are often monitored, so visits and uploads can show up in its security tools.
Using ChatGPT at work isn't misconduct in itself. Breaking a clear company rule about it, or disclosing confidential or personal information, can be. Under the Code of Good Practice: Dismissal, in force since September 2025, an employer has to show that a rule existed and that you knew it or should have known it. It also has to show the rule was applied consistently and that dismissal was a fair sanction.
There's no reported South African case about ChatGPT that I know of, but the closest one points the same way. In NSFAS v CCMA, reported in September 2026, an accountant was dismissed for forwarding nine confidential work emails to her personal Gmail.
The CCMA ordered her reinstatement, but the Labour Court found the commissioner had asked the wrong question. He had looked at whether the emails reached anyone else, when the charge was that she had moved confidential information onto a personal platform outside NSFAS's control. A personal ChatGPT account is a personal platform in much the same way.
The same test protects staff. Without a written rule and some training, a dismissal for pasting something into ChatGPT is much harder to justify, so a clear policy serves the employee as well as the company.
What should you do if you've pasted confidential information into ChatGPT?
Deal with it the same day, using these five steps:
- Stop adding to the conversation, then delete the chat and any shared link to it. OpenAI schedules deleted chats for permanent deletion within 30 days, unless it must keep them for security or legal reasons.
- On a personal account, switch off "Improve the model for everyone" under Settings, then Data controls. This only changes what happens from now on.
- Don't rate the answer with a thumbs up or down, because feedback can send the conversation for training.
- If you pasted a password or an API key, change it straight away.
- Tell your manager or the company's information officer what you pasted, where and when.
Whether the paste needs a section 22 notice is the information officer's decision. No Regulator guidance yet says whether a paste into a consumer AI tool counts as a security compromise. Telling them early is far better for you than having it found later.
What should an AI use policy cover?
An AI use policy turns those sections of the Act into rules a staff member can follow on a Tuesday afternoon without reading the Act. It doesn't need to be long, but it should answer these six questions:
- Which AI tools are approved, and on which accounts? Company accounts only, never personal ones.
- What may never go into any AI tool? ID numbers, bank details, payroll, and special personal information such as health records.
- Who approves a new tool, and who signs the agreement with its provider?
- Who checks AI output before it goes to a client, a regulator or the board?
- Who do staff tell, and how quickly, when something has gone into the wrong tool?
- How will staff be trained, and how often will the policy be reviewed?
The fourth question is the easiest to skip, and the government showed this year why it shouldn't be. On 26 April 2026 the Minister of Communications and Digital Technologies withdrew the draft National AI Policy. Fictitious sources, apparently generated by AI, had been found in its reference list. No personal information was involved, but it was the failure a policy is meant to prevent, found in the document meant to set the country's policy.
A policy that bans AI without offering anything in its place tends to move the habit onto personal phones, where nobody can see it. Staff use ChatGPT because it saves them real time, and a ban asks them to choose between the policy and their workload. It's a pattern CBOS sees with distributors. IT, quite reasonably, turns down a general AI tool that could read every branch's database, and meanwhile staff paste exports into public chat tools anyway.
Is there an AI that doesn't store your data?
Very little stores nothing, so the useful question is who controls what it keeps. ChatGPT's temporary chat isn't used for training, but OpenAI may keep it for up to 30 days. Deletion has legal exceptions, and they've been used. From mid-2025 a US court order in the New York Times case forced OpenAI to keep consumer chats indefinitely, deleted ones included, until it ended on 26 September 2025.
More control comes with a company plan where your own admins set retention, such as ChatGPT Business or Enterprise, Copilot on a work account, or Gemini in Workspace. Even then, the data sits with the vendor, and with ChatGPT it sits outside South Africa.
The most control comes when the company owns the system, because then it decides where its data is kept and who can see it. CBOS builds companies their own private AI on their own data, stored and processed in South Africa, and no outside AI model trains on it. Each build has enterprise guardrails and permissions set per user, so staff get answers only from documents they're allowed to open.
CBOS's page on a private AI assistant built on your own data explains how permissions, privacy, data boundaries, traceability and human oversight are set for each use case. The assistant is designed and built like any of the bespoke systems CBOS builds, around the company's own processes and records.
What should you put in place before your staff use AI?
The sections above come down to seven steps, in roughly this order:
- Find out which AI tools staff already use, and on which accounts.
- Confirm who your information officer is.
- Choose the approved tools and move staff onto company accounts.
- Sign the provider's data processing agreement, and have your lawyer check it against section 72.
- Write the AI use policy, including what to do after an accidental paste, and train staff on it.
- Switch off model training on any account where it's a setting.
- Give staff an approved assistant for the work they were using public tools for.
If you're still deciding where AI is worth using at all, my earlier article on the everyday, operational AI worth putting into a business covers that first.
With those steps in place, the debtors' list from the opening goes into a tool your company chose, and a discovery call with CBOS is a good place to begin choosing it.
Frequently asked questions
Is it safe to discuss business ideas or plans with ChatGPT?
On a personal account, OpenAI holds the conversation under its own privacy policy and may train on it, so treat anything you type as having left the company. Unreleased plans, pricing and client names belong on a company plan or a private system your company controls.
Is ChatGPT confidential?
Not in the way a conversation with your lawyer is. On a personal account, OpenAI holds your chats under its own privacy policy and can train on them by default. On a company Business workspace, your own admins can read and export conversations.
Does ChatGPT use what my staff type to train its models?
On personal accounts it can, unless the user switches off "Improve the model for everyone" under Data controls. OpenAI says it doesn't train on ChatGPT Business, Enterprise or API data by default.
Is ChatGPT Enterprise secure?
OpenAI encrypts Enterprise data at rest and in transit, has a SOC 2 Type 2 audit report and doesn't train on Enterprise data by default. It also signs a data processing addendum. What staff may put into it is still a question for your policy.
Is ChatGPT POPIA compliant?
Not by itself, because POPIA places the duty on your company as the responsible party, whichever tool staff use. On a personal account your company has no contract with OpenAI, so nothing meets section 21 or gives section 72 the agreement it looks for when personal information goes abroad. A Business or Enterprise plan with OpenAI's data processing addendum gives your lawyer an agreement to assess, though the data still leaves South Africa.
Can I upload confidential documents to ChatGPT?
Not to a personal account, unless your employer has said you may. Uploaded files follow the same training and retention settings as the chat. Section 20 of POPIA lets staff process personal information only with the employer's knowledge or authorisation.
What should I do if I pasted confidential data into ChatGPT?
Delete the chat and any shared link the same day, then tell your manager or the company's information officer what you pasted, where and when. On a personal account, switch off "Improve the model for everyone" under Data controls and don't rate the answer, because feedback can send the conversation for training. Change any password or API key you pasted. Whether it needs a section 22 notice is the information officer's decision.
Can I be fired for using ChatGPT in South Africa?
You can be, if you broke a clear company rule about AI or disclosed confidential or personal information, but using ChatGPT at work isn't misconduct in itself. Under the Code of Good Practice: Dismissal, your employer has to show the rule existed, that you knew or should have known it, that it was applied consistently and that dismissal was a fair sanction. Without a written rule and training, a dismissal is much harder to justify.
How long does ChatGPT keep my data?
Saved chats stay until you delete them, and deleted chats are permanently removed within 30 days, with security and legal exceptions. Temporary chats may be kept for up to 30 days. On Business and Enterprise, your admins set retention.
What should you never put into ChatGPT?
On a personal account, leave out ID numbers, bank or card details, passwords and API keys, health and HR records, and clients' personal information. Unreleased financial results and anything under a confidentiality agreement belong on the same list.
Does South Africa have an AI law or AI policy?
South Africa has no AI Act. The draft National AI Policy was published for comment in April 2026 and withdrawn later that month, and Cabinet approved its withdrawal for a full rework. Until a new one appears, POPIA and your contracts govern AI use at work.
Is Copilot or Gemini safer than ChatGPT?
The account matters more than the vendor. Personal Copilot and the personal Gemini app keep conversations for 18 months by default, while their work versions don't train on your data and come with a data processing agreement. A work account on any of the three is safer than a personal account on any of them.