Data Governance · 3 min read
Gemini will separate chat retention from model training control
Google will replace Gemini’s Keep Activity setting with two controls within weeks, letting users keep chat history while excluding text and attachments from model training.
Key takeaways
- Google will roll out two replacement settings for Gemini’s Keep Activity control within weeks.
- The new controls separate chat retention from model-training consent.
- The training setting will cover chats, all text and attachments.
- ChatGPT, Microsoft Copilot and Claude already allow training opt-out without deleting chat history.
- Ziff Davis filed its copyright lawsuit against OpenAI in April 2025.

By Priya Raman
Google will replace Gemini’s single “Keep Activity” control with two independent settings within the coming weeks, ending a requirement that users delete chat history to reject model training.
The present arrangement creates two incompatible choices: keep conversations, including their retention period and use in training, or remove the history and opt out. Google’s revision separates those decisions, although the announcement specifies a rollout window rather than an exact launch date.
Under the planned design, the first setting will let users keep or delete past chats and choose how long Gemini retains their history. The second will determine whether chats—including all text and attachments—feed model training. That scope extends opt-out beyond typed exchanges to attached files.
Google said in an email to Gemini users: “With this change, you will be able to keep your history while independently controlling generative AI model training.” The announcement gives users separate control over two data uses that Keep Activity currently combines.
What changes for current Gemini users?
Today, Keep Activity governs whether Gemini saves chats, for how long it stores them, and whether those chats help train the AI. Users cannot disable training while retaining history under that control. Once the replacement arrives, they can reject training without sacrificing access to previous conversations.
ZDNET’s report said the old restriction made Google an outlier among major AI providers. The same report argued that Google’s past actions had produced a privacy-unfriendly reputation, while welcoming the change as a step in the right direction.
How do rival assistants handle training consent?
ChatGPT exposes “Improve the model for everyone,” which ZDNET called double-speak for model training, and users can disable it. Microsoft Copilot provides “Training on conversation activity” and “Training on voice conversations”; disabling both opts out of training.
Claude’s “Help improve our AI models” option is off by default, although users can inspect and change it. Across these three products, opting out does not require deleting conversation history, and most AI services provide a training setting users can disable manually.
ZDNET characterized Google as the only major provider that had imposed the history-deletion trade-off. The planned Gemini design would close that functional gap, assuming the rollout delivers the promised separation. The source gives no firm deployment date, enterprise administration details, or test results for the two controls.
What should data teams do?
Do not treat the announcement as equivalent to an implemented control. Teams should inventory where staff use Gemini, identify whether histories contain internal text or attachments, and document the permitted retention period and training choice.
During the coming-weeks rollout, users should verify that Keep Activity has been replaced by the two controls. They should confirm that a retained history remains accessible with training disabled, and separately check the stated deletion and retention-duration options.
Because the source describes user settings rather than organizational controls, data teams should not assume an administrator can enforce either choice centrally. Until enterprise behavior is documented, internal policy should state whether staff may use Gemini and retain its chat text and attachments.
Data owners should also separate retention duration from training consent in internal policies. One decision governs how long a record exists; the other governs whether that record may improve Gemini. Combining them, as Keep Activity does, can force a choice between usable history and training exclusion.
A separate disclosure in the report notes that Ziff Davis, ZDNET’s parent company, filed an April 2025 lawsuit against OpenAI. Ziff Davis alleged copyright infringement in OpenAI’s training and operation of its AI systems. The litigation concerns OpenAI, not Google’s Gemini controls, but it keeps training provenance on the wider governance agenda.
The practical verdict is narrow but useful. Within weeks, Gemini is expected to match major competitors on a basic control: reject training without deleting history. Data teams should validate the live implementation before treating the two settings as independent, because the source gives a rollout window rather than a firm date or detailed test matrix.
ZDNet Big Data (Source)
More from Priya Raman
Show full bio
Correspondent covering regulation, policy and their impact on Analytics Digest is a professional daily at Analytics Digest.
13 articles