Chat history and model training are separate choices.
AI / The story

Gemini separates chat history from AI training, with a retention tradeoff

Google is rolling out independent controls for saved chats and training use. Turning history off will still leave a service retention period.

GeekRadar3 min read

Google is separating two choices that should be easy to make independently: keeping a Gemini conversation and allowing that conversation to help train generative AI models. Its November 16 announcement introduces controls that let people retain chat history while opting out of that training use.

There is a second change with a different implication. Google says that when the new History setting is off, future chats will still be saved for around two months for purposes including responding, processing feedback and protecting users and the service. “History off” therefore does not mean “nothing retained.”

The announcement says the changes will roll out globally over the coming weeks. They are not a promise that every account has the new controls today. The practical improvement is that returning to an earlier discussion no longer has to be treated as the same decision as contributing its contents to model training.

Consider an ordinary long-running task: comparing several approaches to a project, then returning a week later to the assumptions. The saved thread is useful because it preserves the reasoning and corrections. That usefulness does not, by itself, express consent to another use of the text. Separating the settings gives each decision a clearer purpose.

It also makes the interface easier to evaluate. A user can ask what is retained for their own continuity, what is used to train models and what is kept for service operations. These questions overlap in the data involved, but the answers need not be identical.

Retention still needs its own reading

The updated Gemini Apps Privacy Hub describes different retention rules for different circumstances. It lists an 18-month default for saved activity, with other deletion settings, and separate handling for material already reviewed by humans. Such reviewed material can be retained for up to three years and is not removed simply by deleting the account’s activity.

These distinctions matter because an interface often presents privacy as a small number of switches while the underlying service has several purposes for keeping data. One switch can change future behavior without undoing every earlier processing step. A training opt-out should therefore be read according to its documented scope, rather than expanded into a claim about retroactive deletion.

This does not mean the controls have no value. It means their value is specific. Keeping a useful history while declining training is a meaningful choice; understanding the remaining retention prevents that choice from being mistaken for something broader.

Memory is another layer

The change also sits alongside Gemini’s editable memory. A saved conversation, a summary used to personalize future answers and permission to train a model serve different functions. Deleting or disabling one should not be assumed to clear the others without checking their own controls.

For a user reviewing an account, a sensible sequence is to inspect history and its deletion period, review the training preference, and then inspect personalization or memory settings. That is a way to understand the configuration, not a guarantee of complete erasure. Work and school accounts can have different arrangements, so this consumer-app announcement should not be substituted for an organization’s policies.

The important improvement is that a convenient feature no longer needs to carry an unrelated choice along with it. The remaining challenge is legibility: people should be able to keep the conversations they need while understanding which copies and uses remain after each switch changes.