14 Lesson 14 of 15
What happens to what you type
Two separate questions get mixed up here: what the model does with your words, and what the company does with them.
Whatever I type goes into the AI. It feels like feeding a machine. Type something private and it is soaked up for good into a thing that millions of other people also talk to.
The question is not whether the model keeps it. It is whether it left your machine.
See it in the lessonSplit the question in two, using the three stores from the memory lesson.
What the model does with it
Your words go into the conversation store. They are read again each turn, they fall out of the window when it fills, and they are gone when you start a new chat. They do not go into the training store. Nothing you type today changes what the model knows tomorrow, and no other user can ask it about your conversation.
What the company does with it
This is the part that really matters for privacy, and it has nothing to do with how language models work. Your conversations are usually sent to a company's servers and stored there. Staff or contractors may read them, and they may be used to train future versions. Which of these apply depends on the product, your plan, your settings and your country. Free plans use conversations for training more often than paid or business plans.
So “the model won't remember this” is true, but mostly beside the point. The app may save it as a memory about you, and the company may keep the whole chat. The question is not whether the model keeps it. It is whether it left your machine.
A rule you can actually keep
Complicated rules get dropped. This one holds up: treat the chat box like a message to a colleague you do not know well, at a company you do not work for. Not a diary, not a search bar, and not a secure system.
In practice, that means no passwords or card numbers. No details that identify other people (their health, their money, their situation), because that is not yours to hand over. Nothing covered by a confidentiality agreement at work, and nothing from a system you are not allowed to copy out of. Your own private matters: only if you have checked the settings and are happy with them.
If you need the help but not the exposure, remove the identifying parts. “A 34-year-old colleague called Priya at my firm” becomes “a colleague”. The advice will be just as good, because the details you removed were not doing any work.
Two settings worth finding once, in whatever tool you use: whether your conversations are used for training, and whether saved memory is on. Both are usually in the settings under privacy or data, and you can usually change both.
The diagram shows both halves on one screen. Send a message and follow where it goes. Then start a new chat, and notice what empties and what does not.
Read this diagram as text
Each message goes from your device to the company's servers. There it lands in two places: the conversation the model reads, and a stored log. Starting a new chat empties the conversation but keeps the log. The model's training store is never touched. If training use is on, the log feeds a future version of the model, drawn separately. If saved memory is on, a short note about you is kept and carried into new chats. The line under the buttons reports all of this after every action.
Strip it before you paste it
Paste in something you were about to send to a chatbot. This finds the details that usually add nothing to the request: email addresses, phone numbers, long numbers, dates and names. It runs only in your browser, so nothing is sent anywhere. It will not catch everything, so read the result before you use it.
Write your own one-line rule for what you will not paste in. Make it about your own life, not general. Then go and find those two settings.
Worth flagging: How companies handle data changes often, and varies by product, plan and country. Do not take anything in this lesson as current fact about any particular tool. Check the one you use.