Tutorials
Stop Google Training AI on You: 7 Settings to Fix Now
Google quietly expanded which of your data can train Gemini. Walk through the 7 exact toggles that pull your account back out of the training pool.
July 12, 2026
You probably didn't get an email about it. But a recent tweak to Google's privacy controls means more of what you type, search, and say into Assistant can now feed Gemini training. If you want out, the settings exist. They're just buried three menus deep.
This tutorial walks you through every toggle that actually matters. No fluff, no scare tactics, just the exact clicks. According to TechCrunch's July 6 write-up, the change quietly broadened the scope of data Google can use for AI improvement, and most users are still opted in by default.
And yes, this includes Workspace users on personal Gmail accounts. Especially them.
What You'll Learn
By the end of this guide, you'll have:
- Turned off the specific activity settings Google uses to train Gemini
- Locked down Gemini Apps Activity (the big one people miss)
- Adjusted YouTube and Maps history so they stop feeding model training
- Cleared past data that's already in the pool
- Set an auto-delete window so you don't have to think about this again
Whole thing takes about 12 minutes. Grab coffee.
Why This Matters Right Now
Google's privacy help center confirms that human reviewers may read Gemini conversations, and those samples can be retained for up to three years even after you delete your chat. That's a wild retention window. Most people assume "delete" means gone.
Photo by Domenico Loia on Unsplash
It doesn't.
The scope also grew in early 2026 to include voice inputs, uploaded images, and shared documents when Gemini is invoked inside Docs or Gmail. Based on Google's own Gemini Apps Privacy Hub, if you've ever asked Gemini to "summarize this doc," a snippet of that doc may have been sampled for review.
So. Let's fix it.
Prerequisites
You need:
- A Google account (personal or Workspace)
- Access to a browser signed in as yourself (not a shared machine)
- About 12 minutes
- Optionally: your phone, to verify the same toggles on the Android/iOS Google app
Nothing else. No tools, no extensions, no scripts.
Step 1: Open Your Activity Controls
Go to myactivity.google.com/activitycontrols. Sign in if you're not already.
You'll see three big cards:
- Web & App Activity
- Timeline (formerly Location History)
- YouTube History
These are the three main data streams Google pulls from. Each one has its own AI-training implication. Don't just kill them all yet, because some genuinely improve product experience. We'll go through them one by one.
Step 2: Fix Web & App Activity
Click Web & App Activity. You'll see a toggle at the top and a checkbox underneath labeled something like "Include audio and video recordings" and "Include Chrome history."
Here's what to do (well, what you should do):
- If you want Google Search and Assistant to still personalize results: leave the main toggle ON, but uncheck Include audio and video recordings. That single checkbox is where voice snippets get sampled for training.
- If you don't care about personalization: turn the whole thing OFF.
- Scroll to Auto-delete and set it to 3 months. Not 18, not 36. Three.
The 3-month window is the shortest Google offers. Anything older gets nuked automatically going forward. Pretty solid default.
Step 3: Kill Gemini Apps Activity (The Big One)
This is the setting most articles skip. Head to myactivity.google.com/product/gemini.
You'll see Gemini Apps Activity. This is separate from Web & App Activity. Repeat: separate. Turning off one doesn't turn off the other.
When this is ON, your Gemini conversations, uploads, and shared docs are eligible for human review and model training. Turn it OFF.
Then click Delete at the top and choose All time. This wipes existing conversations from your account view. (Not gonna lie, the 3-year human-review retention still applies to samples already pulled, but new stuff won't be added.)
One caveat: turning this off breaks Gemini's ability to remember prior chats. If you rely on that, weigh the tradeoff. Personally, I'd take the privacy.
Step 4: Handle YouTube History
Back to activity controls. Click YouTube History.
Photo by JESHOOTS.COM on Unsplash
Google uses your watch and search history for recommendations, which is fine, but also for training video understanding models. The setting bundles both.
Your move:
- Turn OFF Include YouTube searches if you don't want your search terms sampled.
- Keep Include the YouTube videos you watch on IF you rely on recommendations. Otherwise off.
- Set auto-delete to 3 months.
Recommendations degrade if you disable this entirely. Just a heads up.
Step 5: Trim Timeline (Location) Data
Click Timeline. Google renamed Location History to Timeline in 2024, and in 2026 the data moved from cloud to on-device by default. Good change, honestly.
But check two things:
- Location History should say "Off" or "On this device only."
- Web & App Activity's location tracking is a separate stream. Verify at myaccount.google.com/data-and-privacy under Location Sharing.
If you use Google Maps for commute predictions, on-device only is the sweet spot.
Step 6: Lock Down Assistant and Voice Data
Google Assistant (and its Gemini-powered successor on Pixel) records voice snippets when you say "Hey Google." These snippets historically were reviewed by contractors, and while Google's transparency report says the practice is opt-in now, the default varies by account age.
Go to myactivity.google.com/product/voice-and-audio and:
- Turn off Voice & Audio Activity entirely, OR
- Leave it on but uncheck the contractor review consent buried under "Manage Activity."
The second option keeps voice recognition personal to your device but stops human review. Best of both worlds if you actually use voice.
Step 7: Do the Privacy Checkup
Google bundles all of this into a wizard at myaccount.google.com/privacycheckup. Run through it. It'll flag anything you missed and also show you third-party apps with account access, which is a whole other rabbit hole.
While you're there, revoke access for any app you don't recognize or haven't used in six months. Old OAuth grants are a quiet security risk.
Common Pitfalls
A few things that trip people up:
Workspace admins can override you. If your account is managed by a company or school, some of these toggles may be greyed out or reset nightly by admin policy. Nothing you can do from the user side. Talk to IT.
Incognito doesn't mean private. Search in Incognito still hits Google's servers under an anonymized session ID. It's not a training opt-out, it's just a local history skip.
Android Auto and Nest devices have their own settings. The web toggles don't touch them. Check each device's companion app.
"Pause" isn't "Off." Some settings only offer a pause option. Pausing stops new collection, but existing data stays unless you also click Delete.
And one more: turning off Web & App Activity will make Google Search feel dumber for a few weeks. That's normal. It's rebuilding personalization from a blank slate.
Verify It Actually Stuck
Head back to myactivity.google.com after 24 hours. Filter by Today. You should see either:
- No entries (if you turned everything off), or
- Only entries from products you kept enabled
If you see Gemini or Assistant entries after disabling them, double-check the toggle. Sometimes Google's UI shows "Off" while the underlying flag hasn't propagated to all data centers. A hard sign-out and sign-in usually fixes it.
Also check takeout.google.com. Export your Gemini Apps data. If the export is empty or errors out, your delete worked.
What This Doesn't Cover
Opting out of training doesn't mean Google stops using your data for product functionality. Gmail still scans for spam. Maps still logs your searches. Your queries still hit their servers. This is only about training AI models on your data for improvement.
If you want true data minimization, you're looking at DuckDuckGo, Kagi, or Proton search, not a Google settings tweak. Different problem.
Also worth mentioning: rivals aren't automatically better. Every AI-integrated product has some version of this fight. Anthropic's Claude does not train on API data. For consumer Claude.ai chats, Anthropic updated its policy in 2025 so users are prompted to opt in or out and can toggle it later under Privacy & Personalization — still a clearer, per-surface choice than Google's buried toggles. Microsoft Copilot's Enterprise tier explicitly excludes training. Read the fine print wherever you go.
Next Steps
After you've done the above:
- Set a calendar reminder for 6 months to redo the Privacy Checkup. Google shifts defaults during major product launches, and settings occasionally reset silently.
- Consider a secondary account for AI experimentation. Keep your primary account clean.
- If you're in the EU or UK, you have additional GDPR-based deletion rights. File a subject access request via Google's data protection form.
- Audit the same categories on any other AI-connected account: OpenAI, Anthropic, xAI. Each has similar opt-outs, similarly buried.
And if you use Chrome, check chrome://settings/ai. That's where experimental on-device model training toggles live. Turn off anything labeled "Improve AI features by sending data to Google."
That's the whole tour. Twelve minutes of clicking versus years of your queries in a training corpus. Reasonable trade.
