IA · 18 November 2026 · 4 min read

Google splits Gemini privacy controls: keep chat history without feeding AI models, but server retention hits 60 days

In brief: Google has overhauled how privacy and activity logs operate within Gemini, uncoupling the chat history toggle from AI model training permissions for the first time. Previously, users had to allow Google to train its models on their prompts if they wanted to keep their past conversations visible. At the same time, the company extended the temporary retention window for chats generated with history turned off from 72 hours to approximately two months.

by Team Mocchi's

Google splits Gemini privacy controls: keep chat history without feeding AI models, but server retention hits 60 days

For years, conversational AI tools have pushed users into an inconvenient trade-off: anyone wishing to keep an indexed history of past prompts and research had to accept that their queries, uploaded attachments, and generated code could be ingested to train future frontier models. Google is now restructuring that balance across its Gemini mobile and web apps, offering independent toggles while quietly extending how long temporary sessions live on its servers.

Decoupling chat history from model training

Until this rollout, the Gemini activity hub operated as an all-or-nothing proposition. The single "Keep Activity" setting tied the convenience of a visible chat sidebar directly to granting Google permission to use conversation transcripts to train generative systems. Turning off AI training required disabling history altogether, making it impossible to resume earlier working sessions.

Under the new global rollout, the two preferences are fully separated. Users can maintain a searchable chat archive while explicitly turning off the model training toggle. For accounts that previously disabled activity tracking, both toggles will remain switched off by default. Those who had history enabled will retain their conversation log while gaining the option to opt out of AI training.

As reported by Ars Technica, this opt-out applies strictly to future inputs. Any files, prompts, geolocation details, and tool invocations submitted before the toggle was updated and already ingested into review pipelines cannot be retroactively pulled from trained models, and may remain stored on Google's backend for up to three years.

Two-month server retention even with history off

Alongside the welcome flexibility in UI settings, Google introduced a less prominent change regarding unindexed conversations. Previously, when users interacted with Gemini while history was turned off, Google held that data on its servers for a 72-hour operational window before purging it.

The updated Gemini privacy documentation extends this temporary retention period to roughly two months (60 days). As TechCrunch points out, this is double the 30-day window maintained by OpenAI for temporary ChatGPT sessions and by Anthropic for incognito Claude chats.

Google explained that holding unlogged conversations longer is necessary to monitor system errors, handle user feedback, and prevent abuse as models become increasingly autonomous. While Google emphasizes that temporary chats will not be fed into training corpora, the data remains accessible to automated trust, safety, and compliance pipelines for substantially longer than before.

Mocchi's take

Splitting the interface history from model training addresses a long-standing grievance for teams seeking everyday utility without contributing intellectual property to proprietary models. However, expanding the retention window to two full months for supposedly off-the-record prompts serves as a stark reminder for Italian businesses: consumer-facing AI interfaces are never true zero-footprint environments. For organizations handling trade secrets, regulated personal data, or proprietary codebases, moving workloads to enterprise agreements—such as Google Cloud Vertex AI or corporate Workspace tiers with contractual non-retention clauses—remains the only rigorous operational path.

Further reading

All articles on the Mocchi's blog