What “sealed” actually means
Plenty of apps promise privacy. Very few will tell you what happens to your words at the moment they ask a model for something. Here is exactly what happens to yours.
If an app uses AI and also promises privacy, one question matters more than the rest: at the moment it generates something, what exactly is sent, and what is not?
Most privacy pages answer an easier question instead — who we are, that we care, that we use industry-standard security. This one answers the actual question.
The seal is a filter, not a promise
Every entry carries one of three settings. Sealed means it is excluded from the request before that request leaves the app. Not redacted afterwards, not filtered by the model, not covered by a policy asking a provider to be careful — removed at the point the request is assembled, on your side, so it never travels.
That is a mechanical guarantee rather than a moral one, which is the only kind worth printing. A sealed entry is still there for you to read, and it shapes nothing that gets written back. That is the trade: total exclusion means the letters will not know about it either.
What does travel, and where
When you consent to generation, the minimum relevant unsealed text goes to Google Cloud Vertex AI, where the model runs. Google states that Vertex AI customer data is not used to train or fine-tune models without permission, although limited abuse-monitoring logs or short-lived caches may apply unless zero-retention controls are approved and configured. Private letters do not use search grounding, so nothing you wrote is turned into a web query.
Alongside that we keep usage metadata for each generation: which model, how many tokens, the estimated cost, whether it succeeded. That meter does not store letter text. It exists so the service can control its costs without reading your archive.
The uncomfortable part we are not hiding
Your account content is not end-to-end encrypted. It cannot be, because authorised server processes have to read unsealed content to do the things you asked for: write a letter that remembers last week, sync your Drawer to a new phone, interrupt with real-world resources when something you wrote suggests you are in danger.
An app offering those features while claiming end-to-end encryption is describing something that does not exist. We would rather say the true version. Your content is stored, it is readable by the service that serves it to you, it is never sold, never used for advertising, and you can export or delete all of it from Settings.
Deletion removes the authentication account, private state, entries, letters, Sea activity, reports and blocks. Encrypted backups and security logs may persist briefly under our providers’ own recovery schedules, and then expire.
How to interrogate any app, not just ours
- Ask what is excluded from a model request, and whether the exclusion happens before the request is sent.
- Ask whether content is used for training, and whether the answer also covers abuse-monitoring logs.
- Ask whether end-to-end encryption is claimed alongside features that require reading your content. Both cannot be true.
- Ask what deletion removes, what survives it, and for how long.
If a privacy page cannot answer those four in plain words, that is itself the answer.
This piece comes from the team behind Future Me Says — say what you can’t say elsewhere. a letter comes back..