Blog

Gemini splits chat history from model training, retention stays when history is off

Google separates Gemini's chat history and AI training settings. What changes, which retention and processing remain, and what professionals should check now.

· By

A smartphone on a wooden desk shows two separate toggle switches, a hand touching one, beside a confidential paper folder and documents.
Gemini separates chat history from model training, but with history off Google still keeps chats for roughly two months.Image: IamVera.ai — original editorial illustration

Since 16 November 2026, Google offers two separate choices in Gemini: you can keep your chat history while turning off the use of those chats to train generative AI models. However, turning off training stops no retention or processing: with history off, Google keeps chats for roughly two months for service, feedback and safety, per its own documentation.

What exactly did Google change about the Gemini settings on 16 November 2026?

On 16 November 2026, Google announced that the setting for chat history and the permission to use chats for training generative AI models now stand separately from each other. Where these choices previously sat in a single toggle, a user can now keep their history, choose the retention period themselves and delete manually, and separately decide whether the conversations may contribute to model training.

The official announcement from Google states that future chats with history turned off are still kept for roughly two months, in order to respond to requests, process feedback and protect Google, its users and the public. The roll-out is worldwide but phased; the training choice sits in a separate place in the settings. That last point is confirmed among others by the technical reports from Ars Technica, Android Authority and 9to5Google.

Why does turning off training not mean Gemini stops keeping or processing your chats?

Because the new toggle concerns only one use of the data: whether your conversations may train generative models. It is not a delete button and not a ban on all processing. That distinction is the core point of this change.

Google's privacy documentation explains that chats with history turned off are not used for model training, unless the user gives feedback. Google can meanwhile still process those chats for other purposes around service improvement and safety. In our assessment, the biggest misunderstanding lies here: giving feedback on an answer can trigger additional handling that the training opt-out does not necessarily cover. The sources describe a feedback exception, but do not fully work out its precise scope; that therefore remains a point to verify yourself. In addition, business accounts and school or company environments may fall under admin settings that determine the available choices.

  • History retention: whether and how long conversations remain visible and retrievable.
  • Training permission: whether chats may train generative models.
  • Feedback path: giving feedback may entail processing outside the opt-out.
  • Service and safety processing: retention up to roughly two months with history off, even without training.
  • Deletion: the choice intended to remove history, subject to Google's documented exceptions and retention practices.

What does this mean for directors, lawyers and CISOs who use Gemini with sensitive information?

Our analysis: because turning off training does not equal non-retention, lawyers, doctors and directors who paste confidential documents into Gemini face the risk that this content remains with Google for weeks; in our analysis, this makes it important for the responsible organisation to map the actual retention and processing purposes of the accounts used before approving that application. Because giving feedback may open an additional processing path, employees who give an answer a thumbs-down may trigger additional processing that is not fully covered by the opt-out; this is why the working-instruction policy should state that feedback on sensitive conversations is turned off or prohibited where the documentation allows it. Because work or school accounts may be subject to administrator settings, the effective result can depend on the administrator as well as the individual account settings; this is why the CISO must check the effective setting at the admin level rather than relying on what an employee sees in the app. And because a product label is no proof of the underlying processing, a privacy assurance resting on a toggle can fall short for an access or accountability request; this is why the controller records the account status and chosen settings at the moment of use, and draws up separate policy for data that simply must not go to a consumer AI service.

This aligns with a broader line in the topic hub on AI privacy and GDPR: a technical choice is only a privacy guarantee once the whole processing chain is known and recorded. See also the earlier example of oversight of training AI models with user data, where the separation between keeping and training was likewise central.

Which settings should I concretely check before relying on the privacy choice?

Check per account the effective state of the history and training setting, the retention period, the feedback configuration and any admin settings. Do not rely on the default value or on the app display alone, because the roll-out is phased and the actual state may differ per account.

  1. Establish whether history is on or off and which retention period applies; choose the shortest suitable period.
  2. Find the separate training toggle and check whether training is actually off, independently of the history choice.
  3. Check what consequences giving feedback has for the processing of chats with history off, and limit or avoid feedback on sensitive conversations where the documentation and account settings allow it.
  4. For work or school accounts, check the admin settings, because those determine what the user can actually choose.
  5. Bear in mind that with history off, retention of roughly two months still applies for service, feedback and safety.
  6. Record the account status and settings at the moment of use and determine which data never goes to a consumer AI service.

For the legal weighing it is relevant that temporary retention and pseudonymous processing are distinct concepts; for the legal treatment of pseudonymised data under the GDPR, consult applicable legal guidance and assess whether the data remains attributable. See the rules for pseudonymisation under the GDPR. And as we described earlier about why a privacy label does not replace a confidentiality boundary, the lesson here is the same: a toggle limits one use of the data, but does not replace investigation of the full chain.

Sources and references

  1. Gemini app updates: New chat history and training controlsGoogle · 2026-11-16
  2. Gemini Apps Privacy HubGoogle Support · 2026-11-17
  3. Google will let you opt out of AI training in Gemini without losing your historyArs Technica · 2026-11-17
  4. Google now lets Gemini users opt out of AI training while keeping their chat historyAndroid Authority · 2026-11-17
  5. New Gemini setting lets you keep chat history, but turn off AI training9to5Google · 2026-11-16

Sources: The article relies on Google's announcement and the Gemini Apps Privacy Hub, with confirming coverage from Ars Technica, Android Authority and 9to5Google.

← All articles in this topic ← All articles