What "being used for training" actually means

This is the sentence that worries people the most, and often for the wrong reasons. Let's look at what it really covers.

What it isn't

It isn't storage: the model doesn't keep your conversation in a drawer to pull it out for someone else. Training adjusts internal settings, by the billions, from huge amounts of text. Your message weighs a speck of dust in that.

And no one can ask it "what did Philippe write to you yesterday". That's not how it works.

What it still is

A model can accidentally memorise a passage seen many times or that is very distinctive. This is documented: researchers have pulled fragments of training texts back out. The risk is low for an ordinary conversation, real for something highly identifiable, like a key or a rare number.

And part of the conversations is reviewed by humans to assess the quality of responses. That's the most concrete point.

The right way to decide

Don't reason in terms of "is this risky", but in terms of "could this text appear elsewhere without bothering me". A recipe: no problem. A relative's medical file: no.

The good news

It's almost always disableable, with a single checkbox, and the service keeps working exactly the same. The next lesson shows where to find it.