Companion page to the newsletter issue "Privacy. Three settings to change tonight."
The newsletter gave you the short version and the three switches. This page is the whole map, plus the screenshots, plus the part about what to do if you work with other people's information.
Everything here was checked against what the companies themselves publish. Links are at the bottom so you can read the originals.
A note before you start. This page talks about how ChatGPT and Claude handle what you type into them. Settings and policies change often, so check the date at the bottom and confirm anything important against the company's own help pages. Nothing here is legal advice. If you handle regulated information as part of your job, your organization's rules come first.
You can read just the bold lines and get the whole thing.
What you type is scrambled on the way to the company's servers, so nobody in between can read it. Your coffee shop wifi cannot see it. This stop is fine, and it's the one people picture when they hear the word "encrypted."
The moment it arrives, your message becomes a row in a database with your account attached and the time you sent it. This happens on every plan, free and paid. The record is the thing that has a life after the conversation ends.
The model reads your message, writes a reply, and moves on. It does not file your chat away inside itself. This is the part most people get backwards. The model learned from enormous amounts of text long before you showed up, and it is not adding you to a file as you go.
So the worry "is another user going to see my chat" is almost always a no. The real question is what happens to the record from Stop 2.
Automated systems read every conversation looking for serious harm: weapons, fraud, child safety, self-harm. These are programs, not people. Most chats pass through and nothing happens.
If a chat does get flagged, Anthropic keeps it for up to 2 years, and keeps the score the scanner gave it for up to 7 years.
This is the stop almost nobody knows about. Companies hire people to read real conversations and rate them, which is how the answers get better.
In September 2026, 404 Media reported that OpenAI had brought on hundreds of contractors to do exactly this, in a program called Project Lily. OpenAI says it removes personal details first and admits that doesn't always work.