What happened
If you pay for Google Gemini, you'd probably assume your conversations stay between you and the model. They don't. A "Tell HN" post that racked up 447 points and 121 comments in roughly seven hours laid out the fine print: on personal accounts, your saved Google Gemini chat history can be used to train the company's AI models, and the only ways out are switching chat history off or using temporary chat sessions.
The poster, jameslk, found this out too late. He'd assumed a paid subscription meant he wasn't the product. It doesn't.
What "training on your chats" actually means
Think of an AI model as an extremely well-read intern. Every time you hand it another stack of paperwork, it gets a little sharper at spotting patterns. Your Gemini conversations are part of that stack — unless you've turned the pile off. Google's own help page is blunt about it: app activity "is used to improve Google services for everyone, including by training generative AI models, if your Keep Activity setting is on." That toggle is the whole ballgame.
Human reviewers can read a subset of chats
Here's the part that makes people sit up straight. Google states plainly that "a subset of chats are reviewed by human reviewers." The company says those transcripts are disconnected from your account before anyone reads them, but it also warns you not to enter confidential information you wouldn't want a reviewer to see. Both sentences live on the same support page.
And deleting doesn't undo it. Chats that have been reviewed "are not deleted when you delete your activity. Instead, they are retained for up to three years."
Why subscribers are the ones complaining
Paying customers assumed the transaction worked like a gym membership: money in, privacy out. Instead, a subscription buys you a better model, not a quieter data pipeline. The HN thread filled up with people saying some version of "I thought paying meant I wasn't the product." One commenter pointed out that human review isn't unique to Google — every frontier model provider monitors interactions for problematic behavior. Another argued that the Terms of Service carve-outs go further than people think, covering tool calls made while the model runs on your machine, not just the text you type. Treat that second one as a claim from a commenter, not a Google statement, but it's the kind of detail worth checking in whatever tool you use.
What it means for you
Strip away the jargon and it comes down to this: what you type into Gemini may be read by a machine that's learning, and occasionally by a person, for up to three years. Here's how that lands in real life.
At home
You ask Gemini to help plan a tight month's budget, describe a symptom before a doctor's appointment, or draft an awkward text to a family member. That's exactly the kind of ordinary, revealing material training data is made of. None of it is catastrophic on its own — but it's your life, sitting in a retention window you didn't choose.
At work
You paste a client email or an internal document and ask for a summary. That's confidential information, which Google explicitly tells you not to enter. If summarizing documents is your main use, switch those sessions to temporary chats and keep your saved history for the boring stuff.
Running a business
Pricing sheets, supplier lists, an unreleased product name, the first draft of a launch email. A single pasted snippet can carry more commercial detail than a whole meeting. Small teams are the most exposed here, because there's no IT department telling anyone which tab to close.
Studying
Students hand AI tools the most personal writing of their lives — essay drafts, study notes, reflections on why a subject isn't clicking. It feels like a private tutor. The setting that decides whether it stays private is buried two menus deep.
Creative work
Unpublished chapters, half-finished scripts, a joke you're still workshopping. Writers assume drafts are the safest thing they own. Reviewed chats can outlive your deletion by up to three years, which is longer than most drafts survive anyway.
Earning income
Freelancers and solo consultants often use one account for everything: client contracts in the morning, dinner ideas at night. If you want to play with AI without feeding a giant model your archive, mykreatool.com collects free AI tools you can test on throwaway material first.
How to try it right now
The fix takes about two minutes. Start with the free option, because it costs nothing and requires no subscription.
1. Use Temporary chat. Open Gemini and start a temporary chat session instead of a normal one. Nothing from that session is saved, so there's nothing sitting around to be trained on. This works on the free tier too. It's the simplest habit change: temporary chats for anything sensitive, normal chats for "give me a pasta recipe."
2. Turn off Keep Activity. In Gemini's settings, find the Keep Activity toggle and switch it off. Google's explanation of what that setting controls is here: Gemini Apps activity.
3. Know the trade-off before you flip it. Turning Keep Activity off means you'll lose chat sessions after 72 hours. Your history stops being a notebook and becomes a whiteboard.
4. Delete what's already saved. Clear your existing Gemini activity, then read the retention caveat: chats already reviewed by humans aren't removed by that deletion and can be kept for up to three years. Details are on Google's Gemini Apps help page.
5. Change what you paste, not just the setting. Google's own guidance is not to enter confidential information you wouldn't want a reviewer to see. Treat that as the rule for every AI tool, not just this one.
If you use a work or school Google account, check with whoever administers it — the controls for those accounts can be managed centrally, and you may not be the one holding the switch.
Upsides and what changes
The honest upside: you get a cleaner mental model. Once Keep Activity is off, you stop guessing whether a conversation is being filed away, and you stop pasting things you'd rather not see again.
There's a real trade-off too. Your chat history is genuinely useful — it's how Gemini remembers that you're vegetarian, that your company sells to dentists, or that you hate bullet points. Kill the history and you'll be re-explaining context every 72 hours. That's the price of the switch, and it's worth deciding deliberately rather than by accident.
One more thing changes: your instinct. Once you know which setting controls training, you'll start looking for the equivalent in every AI tool you use. That habit is worth more than any single toggle.
Limitations
A few things this fix doesn't do. It doesn't stop human review of chats you've already had, and it doesn't pull back anything already used for training — Google's policy describes what happens going forward, not a recall button. Turning off Keep Activity also doesn't make Gemini a private vault; it makes it a tool that forgets. And none of this is unique to Google. Commenters in the thread pointed out that every major AI provider monitors conversations for policy violations, and independent reporting on human review teams at other labs backs that up. There's also a real gap between what a settings page promises and what happens in aggregate: transcripts can be stripped of your account details and still describe you precisely enough to be recognizable to anyone who knows your situation. If a conversation truly can't leave the room, it shouldn't go into a chatbot at all.
Conclusion
Paying for Gemini doesn't buy you privacy — it buys you a better model. The training switch is Keep Activity, it's buried in settings, and turning it off costs you chat history after 72 hours. That's a fair trade for most people, but only if you make it on purpose. So here's your one action for today: open Gemini, find the Keep Activity toggle, and decide. Then start using temporary chats for anything you wouldn't say out loud in a meeting.
Comments 0