Short answer: ChatGPT is as safe as the text you give it. The moment a client’s name, address or account number is in your message, it leaves your computer, travels to the provider’s servers and is stored there for a while. Whether that is acceptable depends on who the client is and what you owe them. Here is what actually happens, in order, and how to use the AI without the client ever being in the message.
What leaves your computer when you press Enter
Everything in the message box. The AI does not run on your computer: your text is sent to the provider, processed on their machines, and the answer comes back. Attachments follow the same route. A PDF you drop into the conversation is read on their side, not yours.
This is true of every mainstream AI, not just ChatGPT. Claude, Gemini, Mistral and the others work the same way. The product differences are in what the provider does with the text afterwards.
Who can read it, and for how long
Three things happen to a conversation once it arrives.
It is stored. Providers keep conversations so you can come back to them. Deleting a conversation removes it from your account, but providers generally keep a copy for a period, and may keep it longer when the law requires them to.
It may be reviewed. Most providers reserve the right to have staff read a sample of conversations, for safety checks and quality work. The wording varies; the right is almost always there.
It may be used to improve the AI. On consumer plans, conversations can be used to train future versions unless you turn that off in your settings. Business plans and API access usually exclude training by default, but the text is still stored and still readable under the provider’s terms.
None of this is hidden. It is in the privacy policy you accepted. The problem is not that providers are careless; it is that a provider is a third party, and your client never agreed to have their file read by one.
Why this matters more for some people than others
If you are drafting a holiday reply, none of this matters. If you are a lawyer, a doctor, an accountant, a notary or anyone who works under a duty of confidentiality, it matters a great deal.
A duty of confidentiality does not stop at the door of a chatbot. The provider’s terms of service are not a confidentiality agreement with you, and they are certainly not one with your client. Under GDPR in Europe, pasting a patient’s or client’s details into a tool that stores them abroad is a transfer you have to be able to justify. In most firms, nobody has written that justification down, because nobody realised the paste was a transfer.
The same goes for company data. A sales forecast, a customer list, an unreleased contract: the moment it is in the message, it is outside the company.
The settings that help, and their limits
Turn off training in your account settings. Use a business plan if your organisation has one. Delete conversations you do not need. Each of these reduces exposure, and you should do all three.
But none of them changes the basic fact: the text still leaves your computer with the client’s name in it. Settings decide what the provider may do with it. They do not stop the provider from receiving it.
The habit that solves it: the AI never gets the client
The AI does not need to know who your client is to do the work. It needs the structure of the problem: a lease, a diagnosis, a set of invoices, a dispute between two parties. The names, dates and amounts are what make the file confidential, and they are exactly what the AI can do without.
So the clean solution is to replace them before anything is sent. Marie Dubois becomes Claire Martin. The real IBAN becomes another IBAN with the same shape. The dates shift together so the timeline still makes sense. The AI reads a normal message about a person who does not exist, answers as usual, and the real details go back into the answer on your computer.
Done by hand, this takes longer than the work itself. Done automatically, it takes nothing: this is what Kaviar does, on your computer, before the message leaves. The AI you already use stays the same. It simply never sees what is sensitive.
A simple test before you paste
Ask yourself one question: if this message were printed and left on a train, would anyone be harmed, or would I be in breach of anything? If the answer is no, paste away. If the answer is yes, the client should not be in the message. Either take them out yourself, or use a tool that does it for you.
FAQ
Does ChatGPT keep what I paste?
Yes. Conversations are stored on the provider’s servers so you can return to them, and a copy is generally kept for a period after you delete them. Turning off training does not stop storage.
Is a business or enterprise plan enough to protect client details?
It helps: business plans usually exclude your conversations from training and add contractual commitments. But the text still leaves your computer and is still stored by a third party. For work under professional secrecy, that is usually the part you need to avoid.
Can I just remove the names myself before pasting?
You can, and for a short message it works. For a contract, a medical report or a bank statement, there are dozens of details to catch, and a single missed one is enough. That is why Kaviar does it automatically and shows you exactly what the AI received.
Is this only about ChatGPT?
No. Claude, Gemini, Mistral, Copilot and every other hosted AI receive your text the same way. The policies differ in detail; the route the text takes is identical.