Private AI browser: what the phrase means and how to check it
Short answer
“Private” is used for several different claims, from “we do not train on your data” to “your data never leaves the device”. Only the last is checkable from outside, and it is the only one that survives a change of policy.
Four claims that sound alike
These are ordered from weakest to strongest. Vendors rarely say which one they mean.
- “We do not sell your data.” Says nothing about who holds it or how long.
- “We do not train on your data.” A commitment about one use, revocable in a terms update.
- “We do not retain your data.” Better, and still unverifiable from outside.
- “Your data never leaves the device.” Verifiable in five minutes with a network monitor.
Why the last one is different in kind
The first three are promises about conduct on someone else's infrastructure. You cannot audit them, their enforcement depends on a regulator noticing, and they can change with notice you will not read.
The fourth is a property of the software on your machine. It is false the instant a packet goes out, which means anyone can falsify it and nobody has to take it on faith.
How to check any browser's claim
Open a network monitor filtered to the browser's process, ask the assistant a question about a long page, and watch what goes out. A hosted assistant produces an obvious outbound request carrying the page.
Then try it with the network off. An assistant that still answers is running locally; one that errors was never local.
What privacy here does not cover
Browsing is still browsing: the sites you visit see your requests, and non-URL text in the address bar goes to a search engine. Fingerprinting is not solved by anybody.
Prompt injection is unaffected by where the model runs. A page can still try to talk to your assistant.