QWE AI Academy Two camps. Camp A pastes client notes, health stuff, and half-finished code into Gemini Advanced and assumes the subscription bought silence. Camp B flips Keep Activity off, watches history and Memory vanish, then wonders why the paid assistant feels dumber than the free tier they left.
Neither holds up for daily work. Paying does not create a separate training tier on personal Google accounts. Turning the whole thing off also strips the continuity you bought. You need modes, not one heroic toggle.
Funny thing about paying for “smarter over time”: the feature that makes Gemini remember your projects is the same pipe Google documents for model improvement. You are not imagining the tradeoff. The product is built that way.
What Google actually says (and what paid does not buy)
Training is on by default for adult personal accounts when Keep Activity is on – free or paid, same rule. No subscriber-only escape hatch in the consumer app. That is why the November Tell HN thread stung: people had been acting like the invoice was the privacy policy.
Activity can be used “to provide, develop, and improve its services (including training generative AI models),” per the Gemini Apps Privacy Hub (wording as of late 2025; this may have changed). A subset is human-reviewed. Reviewed copies can sit up to three years, disconnected from your account, even after you delete the row in My Activity.
Workspace, education, and paid API / Vertex sit under different contracts. Personal paid does not inherit those terms. If you need contractual non-training, that is a different product line – not a Gemini Advanced checkbox.
Why “just turn Keep Activity off” is a partial fix
Off stops future chats from feeding training (feedback is the exception – more below). Chats still hang around up to 72 hours for service and safety. So you do not get instant amnesia at the infrastructure layer.
What you do lose: sidebar history, pinning, search, Memory / past-chat personalization, and most Connected Apps. Gmail, Drive, Docs-style hooks go dark. Device assistance, Notebook, Messages, Phone, and WhatsApp-style leftovers can remain; the Workspace-shaped ones do not. Official manage-activity help is blunt about that – see Manage & delete your Gemini Apps activity.
Memory needs Keep Activity. Opt out of training and you also opt out of the long-thread continuity that makes the paid plan feel worth it. That is the real cost, not the toggle label.
Mode-switch workflow (what to run when)
One account. Two behaviors. Do not mix them in a single thread.
Sensitive or one-off → Temporary Chat. Web: next to New chat. App: same control, dashed/temporary marker. Confirm the indicator before the first keystroke. Temporary chats are not used to train models, do not land in Recent or Activity, and are kept up to 72 hours for service/safety – again the Privacy Hub, not a blog rumor.
Health, payroll, client names, secrets in code, anything you would not paste into a shared Doc: Temporary. You give up Gems and most Connected Apps inside that session. Good. That is the point of the mode.
Gate question before every prompt: would I paste this into a public Google Doc? No → Temporary Chat. Settings will not save you if the habit is wrong.
Recurring non-sensitive work → Keep Activity on, blast radius small. In Activity settings (gemini.google.com → Settings & help → Activity, or myactivity.google.com/product/gemini): set auto-delete to 3 months (default is 18; 36 or never are also options). Delete specific threads when a project ends. Export keepers with copy-paste or Takeout before the window closes.
Deleting your Activity list does not recall material already sampled for human review. Three-year retention on those copies still applies.
Going fully private default? Turn Keep Activity off – pick turn off and delete if you want a clean slate. Future chats stay out of training unless you send feedback.
Feedback trap (easy to miss): with Keep Activity off, submitting feedback still hands Google the feedback plus the last 24 hours of chats and related content for improvement. That bundle can be retained up to three years, disconnected. Off is not “feedback-safe.”
Freelance pattern that survives client work
Maya – paid personal plan, research gigs. Old habit: full briefs into normal Gemini for outlines. After she read the hub text herself, she split traffic.
Named clients, dollar figures, unpublished findings → Temporary only. Useful paragraphs get copied into her notes the same sitting. Her voice guide and non-sensitive desk research → normal chats, Keep Activity on, 3-month auto-delete. Connected Apps only on days she accepts the trade. For work that needs contractual non-training, she uses a separate Workspace-class account rather than hoping personal billing covers it.
Result she cares about: client strings never sit in long-lived personal Activity. Her own style threads still compound. Not elegant. Effective.
What settings cannot promise
Google still processes prompts to answer you. 72-hour service/safety retention applies on Temporary and on Keep-off chats. Human reviewers see “a subset.” No public figure for how large that subset is, and nothing stating paid personal accounts are sampled less than free ones. If that rate matters to your threat model, the docs do not answer it.
How much opacity are you fine with on a tool that drafts your week? That question sits outside any toggle.
Practical floor: do not put material in Gemini you are unwilling to have a contractor read under some retention path. Settings change exposure. They do not invent a zero-knowledge consumer chat.
FAQ
Does Gemini Advanced / Google AI Pro stop training on my chats?
No. Personal free and paid use the same Keep Activity rules. The subscription is not a privacy tier for the consumer Gemini app.
I turned Keep Activity off – why is the product worse, and can Google still see chats?
Imagine you shut it off Monday, then Monday night you tap feedback on a bad answer. History, Memory personalization, and most Connected Apps (Workspace included) need Keep Activity on – so the assistant feels stateless by design. Chats can still be held up to 72 hours with the account for replies, safety, and feedback handling. That feedback path can pull the prior 24 hours into review and hold it up to three years. Temporary Chat is the cleaner per-session move when you want training off without relying on a global off switch plus perfect feedback discipline.
Is Temporary Chat actually excluded from training?
Google’s Privacy Hub says Temporary chats are not used to train models or for human-review improvement of AI, are not saved to Recent or Activity, and still get short service retention (up to 72 hours). People treat “temporary” like “invisible to Google.” Wrong layer. It is excluded from the training/improvement path described for normal Keep Activity chats; it is not a claim of zero server-side processing. Check the temporary marker at the top before you type. If the marker is missing, you are in a normal chat and the Keep Activity rules apply.
Open Gemini, read the Activity line once, and put the next sensitive prompt in Temporary Chat before you paste anything.