Paying for privacy is the new subscription tier Google forgot to mention.
The Summary
- Google uses Gemini chat history to train AI models for all personal account subscribers, with no opt-out unless users disable chat history entirely or use temporary sessions
- Human reviewers read a subset of chats, which are retained for up to three years even after deletion, disconnected from accounts but still used for training
- The only way to stop training on your conversations is to turn off "Keep Activity," which auto-deletes all chats after 72 hours
The Signal
Google just confirmed what many suspected but few wanted to admit: your subscription buys you access, not privacy. The Gemini app activity policy states plainly that saved chat history "is used to improve Google services for everyone, including by training generative AI models, if your Keep Activity setting is on." This applies to personal Google AI subscriptions. The payment gets you faster responses and higher usage limits. Your conversations still feed the machine.
The human review component adds another layer. A subset of chats are reviewed by human reviewers, and these reviewed chats are kept separate from the deletion mechanism users expect. When you delete your Gemini activity, those human-reviewed conversations stick around for up to three years. Google disconnects them from your account identifier, but the content remains in their training pipeline.
"Please don't enter confidential information that you wouldn't want a reviewer to see or Google to use to improve our services, including machine-learning technologies."
Here's what makes this worth your attention:
- The opt-out mechanism destroys the product's core value (persistent context across conversations)
- Enterprise Gemini users get real data protection, personal subscribers do not
- The disclosure exists but requires reading support documentation most users never see
This isn't about Google being uniquely bad. It's about the economic model of foundation model companies colliding with user expectations around paid services. OpenAI offers opt-out for ChatGPT Plus subscribers. Anthropic doesn't train Claude on conversations by default. Google chose differently, likely because their search and ads infrastructure has always run on user data as fuel. The AI product inherits the same fundamental assumption: more data makes better models, and users will accept this trade even when they're paying customers.
The real signal is structural. Training foundation models requires massive amounts of human-AI interaction data. The companies building these models face a choice: pay for that data, restrict training to synthetic or licensed sources, or treat user conversations as fair game. Google picked option three for personal accounts while offering option one (real privacy) to enterprise customers who can negotiate contracts. This creates a two-tier system where privacy becomes a business feature, not a consumer right.
The Implication
If you use Gemini for anything you wouldn't want in a training dataset, switch to temporary chats or turn off Keep Activity now. Accept the 72-hour deletion window as the price of not being training data. Better yet, if your work involves confidential information, push for enterprise accounts with actual data protection agreements.
For everyone else, this is a preview of how AI companies will segment privacy. Consumer tools will default to data collection. Business tools will offer protection at a premium. The gap between what personal users get and what enterprise customers negotiate will keep widening until regulators force convergence or competitors build business models that don't require harvesting subscriber conversations.
Watch what happens to Google's enterprise Gemini adoption rates. If businesses avoid the consumer product specifically because of this training policy, that's the market pricing in the risk of letting AI companies read everything.