Auto Tagging sorts the sources you follow — newsletters, feeds, YouTube channels — into topics like Design or Finance and kinds like Company Updates or Store. To do that, Neuplace asks an AI model, Google's Gemini, to classify each source. That means something leaves Neuplace. This post is about exactly what, and why we drew the line where we did.

It is off until you turn it on

Auto Tagging starts switched off for every account. Turning it on opens a confirmation screen that tells you, before anything happens, what will be sent and to whom. Only confirming that screen switches it on, and we record when you agreed.

Your consent is checked again every time a source is about to be classified, not just once at the start. If you turn Auto Tagging off while some sources are still waiting in line, those requests stop before reaching Gemini.

What Gemini receives

Each request describes one source, and only these four things about it:

Its name, as shown in Neuplace. For an email newsletter, that is the sender name carried on the message — which is sometimes a person's name.
Its web domain.
What kind of source it is: an email sender, a feed or a YouTube channel.
Its public description, the short blurb the publisher wrote about itself.

What Gemini never receives

Your emails' contents. Their subject lines. Sender email addresses. Your account, or anything that says who you are or that you follow this source. A request describes a publisher, not a reader.

That boundary is built into how the feature works, not just written into a policy. The classifier is handed a single, small set of those four fields and nothing else; there is no place for an email body or an address to slip in.

Where the description comes from

The description is the most useful clue, so it is worth being precise about it:

For email newsletters, it is read from the publisher's own public website.
For newsletters on Substack or beehiiv, the publisher's site is the platform itself, so we read the publication's own public page there. To find that page, Neuplace uses the part of the sender's address before the "@" — so Substack or beehiiv does see that name in the request, as anyone opening that public page would. This happens only for those two platforms, never for any other address. Neuplace doesn't keep the page address it builds for this lookup, and doesn't log it.
For feeds, it comes from the feed itself.
For YouTube channels, it comes from YouTube, read with your Google sign-in.

Before a description is stored or sent anywhere, we strip email addresses and phone numbers out of it. Some channel descriptions carry contact details; classifying a channel does not need them.

A closed list, and a preference for "I don't know"

Gemini cannot invent tags. It must pick from a fixed list of 40 topics and 7 kinds, or answer "unknown" — and we tell it plainly that a wrong label is worse than none. It also runs at temperature zero, which keeps its answers as consistent as possible instead of creative.

We use Gemini's paid service, under which Google does not use the data we send to train its models. Our privacy policy covers this in section 5.

Tags belong to the source, not to you

Sources are shared across Neuplace: everyone who follows the same newsletter follows the same source. So a source is classified once, and that result is reused for everyone who follows it. The consent screen says this too.

Because of that, we made one more choice: if you haven't turned Auto Tagging on, you don't see tags at all — not even tags that exist because someone else turned it on. Opting in is how you get the feature; not opting in means it isn't there.

What turning it off does, and doesn't do

Turning Auto Tagging off stops any new requests on your behalf and hides tags from your view. It does not reach back and un-send a request that already happened, and it does not remove a tag already stored on a shared source. We'd rather say that plainly than imply otherwise.

A request describes a publisher, never a reader.