Leave it off unless you actually live inside Gmail and want Gemini to pull flight times, receipts, and order numbers on command. Gemini's Personal Intelligence is off by default, opt-in per app, and easy to disconnect, so there is no penalty for waiting. Here is the line most explainers skip: Google does not train on your inbox itself. But your prompts and Gemini's answers can train the model, and a subset of those chats get read by actual humans.
On March 17, 2026, Google made Personal Intelligence free for every US user on a personal Google account. Before that it sat behind the paid Gemini tier. Now the toggle is one tap away for people who never asked for it, which is why it deserves a straight answer rather than a shrug. The feature is genuinely useful. It also changes what a single tap can expose, and most coverage rushed past that part to demo the flashy lookups.
What Personal Intelligence actually connects to
Personal Intelligence links Gemini to four data sources: Gmail, Google Photos, YouTube history, and your Search history. With Gmail connected, Gemini can find a flight confirmation, summarize a long thread, or pull an order number out of a receipt. With Photos connected, it can surface images from a specific trip or read details out of a picture, like a license plate or a product label. At rollout the feature was limited to US users on personal Google accounts, in English, and is not offered to Workspace business, enterprise, or education accounts.
Each source changes what Gemini can answer in a specific way. Gmail turns your mailbox into a searchable layer, so a question like find my last car rental receipt resolves against real messages instead of your memory. Photos lets the model read inside images, which is why it can answer questions about a photo's contents and not just its filename. YouTube and Search history feed taste and intent, so recommendations and follow-ups lean on what you have already watched and looked up. None of this is magic. It is retrieval over accounts you already own, with the model acting as the index.
The connections show up across three surfaces: the Gemini app, AI Mode in Google Search, and Gemini inside Chrome. Register that breadth before deciding. Granting access in one place can make personalized answers appear in others, so the decision is not contained to a single app. Connect Gmail expecting it to matter only inside the Gemini app, and you may be surprised when a personalized answer turns up in Search or while browsing. Map the access to every surface it touches before you decide it is harmless.
Off by default, and opt-in per app
Nothing turns on until you flip it. Personal Intelligence ships off, and the permissions are granular: you can connect Gmail and leave Photos disconnected, or the reverse. Google groups these under a Connected Apps dashboard in Gemini's settings, where each source has its own switch. If you want the email lookups without handing over your photo library, that combination is allowed. The practical move is to connect the narrowest source that does the job and leave the rest off, then add more only when you hit a real limit.
Google didn't flip this on for you. The danger is you flip it on once, forget the scope, and stop noticing what you're sending.
The training-data nuance everyone skipped
Gemini does not train its models directly on your Gmail inbox or your Photos library. What can be used for training are your prompts and Gemini's responses to them. Google also confirms that a subset of chats are reviewed by human reviewers to improve its services. So the inbox itself is not the training corpus, but the conversation you have about your inbox can be.
If you ask Gemini to summarize a sensitive email and quote part of it back, that quoted text now lives inside a prompt and a response. Those are the artifacts eligible for training and the occasional human review, not the raw message sitting in your inbox. The distinction is real, but it is thinner than the headline suggests. Your inbox isn't the training data. Your questions about it are.
This is why the wording of your prompts matters as much as the toggle. A request like check whether my last invoice was paid keeps the sensitive detail on Google's side, in your inbox, and out of the conversation. A request like read me the full text of that legal notice and explain it pulls the sensitive content directly into a prompt and a response, which is exactly the surface that can be sampled. Same feature, very different exposure, decided entirely by how you phrase the ask.
Want personalization without the training and human review? Turn off the Keep Activity setting at myaccount.google.com under Gemini Apps. Google states that with it off, future chats are not used to train its models, unless you separately send feedback. Personalization and training are two different switches, so check both.
Should you turn on Gemini Gmail access? A straight verdict
Turn it on if you mostly use Gmail and Photos for logistics and want faster lookups; leave it off if your inbox holds legal, medical, or financial material. The reason that line works is the one constraint underneath the whole feature: convenience and exposure scale together here. The more you let Gemini reach into your accounts, the more useful it gets and the more of your material can pass through prompts that are eligible for review.
Connect it if you genuinely run your life out of Gmail and Photos and want a faster way to retrieve things you already own. Leave it off if you would not be comfortable with a sample of your queries being reviewed, or if your inbox carries sensitive material you would rather not narrate to an assistant. The feature is useful and the controls are real. A middle path also exists: connect it, keep the genuinely private threads out of your questions, and turn off Keep Activity so the convenience does not come with the training and review.
| Situation | Turn it ON | Leave it OFF |
|---|---|---|
| Primary use | Fast lookup of your own flights, receipts, orders, trip photos | You rarely need Gemini to read personal accounts |
| Comfort with review | Fine with prompts and responses possibly training the model | Do not want any query sampled for human review |
| Inbox sensitivity | Mostly ordinary mail and logistics | Legal, medical, financial, or confidential threads |
| Control habits | Will manage Connected Apps and Keep Activity deliberately | Prefer not to track another set of toggles |
| Reversibility need | Comfortable enabling now, disconnecting later | Prefer zero data flow until you decide otherwise |
How to turn off Gemini Gmail access, delete memories, and disconnect
All three controls live inside Gemini's settings, and they do different jobs. Disconnecting an app stops Gemini from reading that source going forward. Deleting a saved memory removes a specific stored fact. Turning off Keep Activity stops new chats from being logged and used for training. They are not interchangeable, and turning off one does nothing to the others, so decide which problem you are actually solving before you start tapping.
- Disconnect a source: open Gemini settings, go to Connected Apps, and turn off Gmail, Photos, YouTube, or Search individually.
- Delete saved memories: manage saved info in Gemini settings and remove items one by one or clear them in bulk.
- Stop training and logging: at myaccount.google.com, open Gemini Apps activity and turn off Keep Activity.
- Important caveat: deleting a saved memory may not fully erase it while the original conversation is still saved, so delete that conversation too.
The order matters if you are cleaning up after the fact. Disconnect first so no new data flows in. Then turn off Keep Activity so future chats stop being logged. Then work backward through what already exists: delete the saved memories you do not want kept, and delete the underlying conversations that hold the detail. Doing only the first step leaves a history behind. Doing only the last step lets new data refill it.
That last caveat is the one people miss. Memory deletion and conversation deletion are separate actions. If you only delete the memory, the underlying chat can still hold the detail. Clear both when you want something gone.
Why the EU, UK, and Japan have a different default under GDPR
The popular shorthand that the EU banned Gemini's photo scanning is not quite right. What happened is that Google disabled Personal Intelligence by default in the EU, UK, and Japan as a preemptive compliance move under GDPR and comparable local laws. Photo analysis can touch data that GDPR treats as a special category, which carries a higher consent bar, so Google chose to make the feature opt-in rather than risk deploying it the way it did in the US.
The signal in that decision is useful even if you live in the US. When a company sets a stricter default for one region purely on regulatory risk, it is flagging where the sensitive edges are. Here those edges are the photo analysis and the personal-data combination, the exact parts a US user enables with a single tap. Same feature, looser default, and the judgment lands back on you. Borrow the stricter region's posture: treat the feature as opt-in for yourself, even though your account does not force you to.
The trade you are really making
Personalization of this kind asks you to merge two things that used to stay apart: your private archive and a cloud model trained on everyone's interactions. Gemini's design keeps your inbox out of the training set, which is meaningful. But the queries that make the feature worth using are the same queries that can be sampled and reviewed. There is no version of this where you get deep personalization and zero data flow.
That tension is the reason a separate category of tool exists: an external memory layer that you control, rather than a personalization feature bolted onto a search-and-ads company. The question is not only what an assistant can do with your data. It is who holds the data and on what terms, and whether the default leans toward keeping it yours or toward feeding the product.
Where MemX fits
MemX is a consumer memory app that sits over your own documents, photos, and notes across Android, iOS, and WhatsApp, so you can ask questions of your material without routing your private archive into a general model's training pipeline. It is private by architecture: per-user isolation, customer-managed encryption keys, encryption at rest, and an on-device first pass. That is a different posture from a free personalization toggle, not a claim of end-to-end encryption or zero-knowledge. The point is structural: a memory layer should answer to you first.
If you do enable Gemini's Personal Intelligence, treat it as a retrieval convenience and keep the genuinely sensitive material out of the conversation. If you want a memory layer whose default assumption is that the data stays yours, that is a separate design choice worth making deliberately rather than tapping into by accident.
01Is Gemini Personal Intelligence on by default?
No. In the US it ships off and you opt in per app. You can connect Gmail without connecting Photos, or any other combination. In the EU, UK, and Japan, Google sets it off by default as a GDPR compliance measure.
02Does Gemini read my Gmail emails?
Only if you connect Gmail under Connected Apps. Once connected, Gemini can read messages to answer your questions, like finding a receipt or summarizing a thread. It stays off until you opt in, and you can disconnect it at any time in Gemini settings.
03How do I stop Gemini from reading my Gmail?
Open Gemini settings, go to Connected Apps, and turn off Gmail. This stops Gemini from reading new email. To also stop training, turn off Keep Activity in your Gemini Apps activity at myaccount.google.com. They are separate switches.
04Can I delete what Gemini remembers about me?
Yes. You can delete saved memories in Gemini settings, individually or in bulk. One caveat: a deleted memory may persist while the original conversation is still saved, so delete that conversation too for a complete removal.
05Did the EU ban Gemini photo scanning?
Not exactly. Google preemptively disabled Personal Intelligence by default in the EU, UK, and Japan to comply with GDPR and local laws, since photo analysis can touch special-category data. It was Google's own restriction, not a recorded enforcement action.
