This is a scripted illustration. The suggestions are scripted, not model output, the person is fictional, and nothing is sent anywhere. The real prototype is a Firefox extension that needs local models.
Discretion helps people share less before they send, and it has two parts. One part scrubs personal details from a text field before you send it. The other part adds context from your local notes on click. As a prototype, it is a Firefox extension, but it could be built into a browser, work as a keyboard application, or stay an extension. Ideally the model is private, such as local model compute.
Chat encourages people to share as much as possible to get the best response, so they give model providers a lot of personal information, often without noticing how revealing their statements are.
Anyone who uses AI chat or AI search has this problem, whether or not they are logged in. Chat rewards sharing as much as possible to get the best answer, so people reveal personal information to model providers. This includes both PII and statements that can be used to profile their habits or activities. People are often not aware of how revealing their statements can be.
It matters because the information is shared before the person can weigh the trade-off.
Discretion acts at the moment of sharing, before the text is sent. One part scrubs personal details from what you are about to send. The other part adds useful context from your own local notes, so you share what helps, not everything.
Ideally the model that does this is private, running on your own device, so the check doesn't create a new place for your data to go. As a prototype it is a Firefox extension. It could also be built into a browser or work as a keyboard application.
If people can see and control what they share at the moment they send it, they will share less personal information with AI services, and still get good answers by drawing on their own notes. If that holds, tools that sit between a person and an AI service, running privately on their own device, become part of how the web works.