Google adds separate opt‑out for Gemini AI training
Google now lets Gemini users disable AI training while keeping chat history, adding separate toggles in activity settings.
Source: Ars TechnicaGoogle has updated its Gemini AI chatbot to give users a dedicated switch for stopping their conversation data from being used to train future models, while leaving chat‑history retention untouched. The move, reported by Ars Technica on November 17, 2026, separates two previously bundled settings and marks a notable shift toward finer‑grained privacy control for everyday AI users.
What exactly did Google change in Gemini?
Previously, the only way to keep Gemini from learning from a user’s inputs was to turn off the entire chat‑history feature. Doing so erased the local record of past conversations, making it impossible to resume a thread later. According to the Ars Technica article, the new rollout introduces two independent toggles inside the Gemini activity settings: one for Chat History and another for Model Training. The history toggle continues to let users set a retention limit, while the training toggle is an opt‑out that prevents future interactions from being added to Google’s training corpus.
How does the new opt‑out work in practice?
The training toggle lives under the Activity section of the Gemini settings page. Users can also reach it via their Google Account at Data and Privacy > My Activity > Other activity > Gemini Apps Activity. Once the switch is turned off, Gemini will still respond to queries and retain the conversation locally (subject to the chosen history limit), but Google will no longer feed that data into the models it trains. The change applies only to interactions that occur after the toggle is set; any data already used in training remains part of the model’s knowledge base.
What data does Google still keep, and for how long?
Google’s policy, as described in the source, states that any data submitted to Gemini before opting out may stay in the training set for up to three years. This includes user inputs, Gemini’s replies, uploaded files, tool‑call logs, location information, and data from linked apps. Even after opting out, Google may retain the data for other purposes—such as responding to user requests, processing feedback, or safeguarding the platform—for up to three years. If a user disables history entirely, the platform currently holds that data for 72 hours for the user and 24 hours for Google; the upcoming update will extend Google’s retention to two months while still giving the user a short‑term window to retrieve the data.
When will the separate toggles be available to everyone?
The article notes that the feature is rolling out globally but may take a few weeks to appear for all accounts. Because the change is delivered as a server‑side update, users do not need to install new software; they simply need to refresh the Gemini app or web interface and look for the new Model Training switch in the Activity settings.
What does this mean for creators, developers, and researchers?
For creators who rely on Gemini for brainstorming, content generation, or rapid prototyping, the ability to keep a conversation history while opting out of training means they can build longer, iterative workflows without sacrificing privacy. Developers integrating Gemini into apps can now offer end‑users a clearer consent model: retain the dialogue for stateful interactions while guaranteeing that none of the exchanged data will improve Google’s future models.
Researchers studying AI alignment or data‑privacy impacts can use the change as a natural experiment. By comparing model performance on data streams that were opt‑out versus those that remained in the training pipeline, they can quantify the marginal contribution of everyday user interactions to model quality.
How can you explore this change with Grafy?
Grafy’s Chat app lets you converse with a variety of large‑language models side by side, including models that emulate Gemini‑style capabilities. While Grafy does not host Google’s Gemini directly, you can test the same privacy workflow by using the Chat interface to start a session, enable or disable history in the Grafy settings, and observe how the platform respects your preferences. For a deeper dive into the policy language, you can upload the Ars Technica article (or Google’s privacy documentation) into Reader and ask the AI to summarize the key points or compare them with other providers’ practices. Links: Grafy Chat and Grafy Reader.
Broader implications for AI privacy
Google’s decision reflects growing user demand for granular data‑control options. By decoupling training consent from conversation continuity, the company acknowledges that privacy and utility can coexist. However, the retention windows—up to three years for training data and up to two months for non‑training storage—still give Google substantial leverage over user‑generated content. Stakeholders should weigh the convenience of a persistent chat history against the long‑term exposure of personal or proprietary information, especially in professional or research contexts.
Frequently asked questions
Can I stop Google from using my past Gemini chats for training?
No. The opt‑out only applies to future interactions. Any data submitted before the toggle was enabled remains part of the training set for up to three years, as explained by Ars Technica.
Will disabling chat history delete my data from Google’s servers?
Disabling history prevents new data from being saved locally, but Google retains existing history for 72 hours (24 hours for the user) and will keep it for up to two months after the upcoming update, even if history is turned off.
How do I know if the training toggle is active on my account?
Open the Gemini app, go to Settings > Activity, and look for the Model Training switch. It should be set to Off to opt out of training. You can also verify the status in your Google Account under Data and Privacy > My Activity > Other activity > Gemini Apps Activity.
Source: Ars Technica — Google makes it easier to use Gemini without training AI on your data


