When Keep Activity is enabled, chats from personal Gemini accounts may be used to improve Google services, including training generative AI models. Paid subscribers do not receive a separate opt-out from model training and can only disable chat history or switch to Temporary Chat.
Where the controversy comes from
A Tell HN discussion appeared on the Hacker News front page this week with a specific question: Do paying Google Gemini users still have their conversations used for model training by default?
Google’s answer is essentially yes.
According to Google Gemini Apps Help, individual Google AI subscribers who enable Keep Activity allow their saved chats to be used to improve Google services, including training generative AI models.
In other words, paid subscriptions such as Gemini Pro and Ultra do not automatically include the right to opt out of model training. Free and paid users face the same limitation: there is no standalone opt-out control.
Paying does not mean opting out
The most common misunderstanding is treating “free” and “paid” as two separate data-handling regimes.
Google’s documentation does not promise that paid users’ conversations are isolated from those of free users. The determining factor is Keep Activity, not the subscription tier. As long as activity is retained, saved conversations may be used to improve the services, and generative AI model training is included.
That leads to an uncomfortable conclusion:
Paying for Gemini buys model capabilities and usage limits, not default data isolation.
This does not mean Google is secretly processing conversations. Rather, its product design does not offer a dedicated control for training. Users would reasonably expect a clearly labeled setting such as “Improve the model” or “Use for training” that they could enable or disable themselves.
Gemini personal accounts currently lack that level of clarity. The relevant discussion was submitted through Tell HN to the Hacker News front page. Commenters focused on one question: Why are the feature settings and authorization for model training bundled together?
There are only two ways to opt out
According to Google’s support documentation, individual Google AI subscribers who do not want their chats used for model training have two options:
- Disable chat history by turning off Keep Activity.
- Use Temporary Chat for temporary conversations.
Neither option is a standalone “opt out of training” control.
Turning off Keep Activity prevents chats from being saved. The tradeoff is losing the ability to retain context across conversations. Personal prompts, project background, and follow-up discussions may no longer be available as historical context.
Temporary Chat treats a conversation as a temporary space and does not add it to the account’s Keep Activity history. It works well for one-off tasks, but it also sacrifices long-term context and the ability to continue asking questions based on an existing conversation.
The available choices can be summarized as:
User goal: opt out of model training only
Google provides:
- Disable chat history
- Use Temporary Chat
There is no opt_out_training=true parameter. Users who want to minimize data collection must also disable chat history.
For most people, this adds friction to the experience. For engineers who routinely have Gemini process code, daily reports, and project materials, it changes the workflow directly.
Human review makes the risks more concrete
Training use is only part of the controversy. Google’s official documentation also says that Gemini app activity may be reviewed by human reviewers. Google states that the relevant records are disassociated from users’ accounts after human review.
This clarification addresses one concern: reviewers should not retain a long-term link showing that a particular piece of content belongs to a specific account.
However, “disassociated after review” does not mean the data was never accessible to internal staff. The review itself has already taken place. More cautious enterprise customers will continue to ask:
- Which data is covered by human review?
- Can reviewers see complete conversations or only filtered samples?
- At what point does disassociation occur?
- Are review logs, screenshots, or caches retained?
- Do the rules remain consistent across regions, subscription tiers, and enterprise workspaces?
The current documentation does not provide enough information to answer all of these questions.
“The connection is severed” describes the state after review. It does not establish that the review process carried no risk of access. Those are not the same thing.
Do not treat the chat box as an enterprise safe
Google’s help documentation explicitly warns users not to enter confidential information into Gemini conversations. This should be treated as a hard rule, not a soft reminder hidden in the corner of a dialog.
In particular, users should not submit:
- API keys, access tokens, and database passwords
- Unpublished customer code
- Contracts, merger materials, and financial forecasts
- Internal incident records and security vulnerability details
- Exported datasets containing personal information
- Production logs, complete request bodies, and user data
A practical approach is to use layers of protection rather than assuming every conversation is inherently safe:
- By default, use only synthetic data, public documentation, and redacted samples.
- Before pasting code, remove credentials, domains, customer names, and internal identifiers.
- Replace real secrets with placeholders, such as writing a token as
YOUR_API_KEY. - When real data is required, prefer Gemini Enterprise, Vertex AI, or another service with contractually defined controls over enterprise data.
- Do not treat a Pro subscription as a guarantee that data will not be used for training.
No elaborate scripting is necessary. Many data leaks do not come from sophisticated attacks; they begin with a routine copy and paste.
Editorial perspective
The real significance of this controversy is not another change to Gemini’s privacy language. It is that paying users still do not receive clear data controls. Google groups “retain history,” “human review,” and “use for training” under Keep Activity, forcing users to disable the entire feature to reduce risk. That may be understandable for a consumer product, but it is not precise enough for subscribers who work with code and business materials.
An independent training opt-out will gradually become a hard requirement for enterprise procurement. A statement that data “will not be used for training” is not enough; customers must also be able to verify the scope of human review, retention periods, and third-party processors.
What has changed is that chat models are no longer merely search boxes. They have become everyday software interfaces. The debate is no longer about individual questions and answers, but about whether an account can accumulate context. A more reasonable design would separate training consent from product features, much as it does for advertising tracking, with a dedicated toggle, a clear default state, an explanation of its scope, and a way to withdraw consent. Gemini’s current problem is that it hides data controls behind a feature setting.