Google Won’t Train on Your Inbox. Just the Summary.

Google’s Search help documentation makes a promise and then immediately explains its own limits, and the second half is the part nobody reads.

“Search services do not train generative AI models directly on your Gmail inbox, Drive, Calendar or other Google Workspace apps or on imagery or audio from your Google Photos library.”

— Google Search Help

Then, further down the same page: when you interact with Search, summaries, excerpts, generated media, and inferences drawn from your relevant emails, files, and photos may be used to answer your prompts, and Google trains its generative models on those summaries, excerpts, generated media, and inferences. Google’s own documentation says both things, and both are accurate. The first sentence is the headline. The second sentence is the policy.

Read those together and the promise gets narrower fast. Google is not vacuuming your mailbox wholesale into a training corpus, and anyone claiming otherwise is wrong. What it is doing is letting Search read the emails and files it decides are relevant to your question, produce a compressed version of what’s in them, and then train on the compressed version. The original stays out. Everything the original meant goes in.

That distinction holds up legally and collapses practically, because a summary is not a lesser thing than the document. It’s the document with the filler removed. A three-paragraph summary of your custody arrangement, your diagnosis, your salary negotiation, or your severance agreement contains the entire substance of the email it came from and none of the boilerplate. If your concern about training data was ever “I don’t want a model learning the private facts of my life,” the promise not to ingest the file itself does not address that concern. It relocates it.

Google’s position isn’t unreasonable and it deserves to be stated fairly. If you deliberately connect Gmail and Drive to Search so it can answer questions about your own material, the product cannot work without processing that material, and it cannot improve without learning from how the processing went. Google published the distinction rather than hiding it, offers controls, and tells users plainly not to connect content they wouldn’t want used this way. That’s more disclosure than most of the industry manages. A company that wanted to be sneaky here would simply not have written the second paragraph.

The trouble is that meaningful consent requires understanding what you agreed to, and this is a genuinely hard thing to hold in your head. Connecting an app feels like a convenience toggle, the same mental category as linking a calendar or enabling notifications. It is actually a decision about what a model gets to learn from, mediated by a relevance judgment you don’t see and can’t audit. Nobody knows which emails Search decided were relevant to a given question, what it extracted, or what inference it drew and kept. The subject matter isn’t disclosed to you. Only the fact that the category exists.

There’s a control-panel detail that makes this worse in practice. Disconnecting an app and deleting your Search Services History are separate actions. Turn off the connection and the derived material doesn’t automatically go with it. You have to go delete the history as its own step. That’s a defensible design if you think of history as a distinct user-owned record, and it’s a trap if you think, as most people will, that unplugging the source unplugs everything downstream of it. Depending on your settings, some of that material can also be reviewed by human beings, which is standard practice across the industry and still not what anyone pictures when they connect their Drive.

None of this is happening in isolation. Google has spent the last year rebuilding Search into something that acts on your behalf rather than handing you links, and an agent that acts on your behalf needs to know about you. Personal context is the whole product. The connected-apps feature isn’t a bolt-on privacy compromise — it’s the mechanism that makes agentic Search work at all, which means the derivative-training question is going to get bigger, not smaller, as more of the interface depends on it.

So the honest version of the choice looks like this. Connecting Gmail to Search buys you a genuinely better assistant, one that can answer questions about your actual life instead of the general internet. The price is that a compressed record of the parts of your life Search found relevant becomes training material, sitting in a category you can delete but never inspect. That might be a fine trade. Plenty of people will make it happily and shouldn’t feel foolish for it. But it should be made with the second sentence in view, not the first. Google, having written both, knows exactly which one is going to end up in the headline.

Sources: Google Search Help · Engadget · Gemini Apps Help