There are a couple of obvious routes to improvement. You don't ever substantiate the idea that secrets are "revealed" because a transformer is given user prompts. An actual illustrative example is needed. Why this is different than the search box should also be explained.

In essence, this is just another example of a service being offered in return for the surveillance of the user, which in this case is relatively closely limited to the input. But running the model by presenting input to an instance running elsewhere, like any remote service, is replaceable. I don't use a speech-to-text bot to make my audio transcriptions remotely, though that would be very "convenient." I run whisper.cpp on one of my own computers, so the actual recording of my speech in class is never transmitted to a third party (though I will make it available to the public anyway, through the web). That model running in my own computer also transcribes my day-to-day voice notes, which are indeed private, privileged, and definitely not something to be given to a third-party listener of any kind. But I don't have to. If I wanted to be running Mistral in a laptop, or one of the Meta LLaMa products, I could be doing that too. Using slightly different software slightly differently changes the entire nature of the situation and completely solves the problem. That's important to keep in mind, and to help the reader understand.