Gemini Trains on Your Chats Unless Keep Activity Is Off

Last updated: November 4, 2026

Short answer to "does Gemini train on my chats": yes, by default, on a personal Google account, unless Keep Activity is off. A Tell HN thread hit 500+ points today because the poster learned this after paying for a subscription. Google's own documentation says the same thing the thread says, and it is not hidden - it is just not where most people look.

This is the Gemini twin of the story we covered earlier today about Codex training toggles. Different vendor, same lesson: the setting that matters is a toggle you set, not the plan you pay for.

What Google documents#

Everything below is from Google's Gemini Apps Privacy Hub, read in full today. It is a vendor statement, so it shows intent and policy, not verified behavior.

• Keep Activity on. Google says it uses your activity "to provide, develop, and improve its services (including training generative AI models), as well as to protect Google, its users, and the public with the help of human reviewers." Activity includes chats, uploads, feedback, product usage and location info.
• Human review. "A subset of chats are reviewed by human reviewers (including Google's trained service providers)." Reviewed data is disconnected from your account and saved for 3 years. The page says chats reviewed by humans "are not deleted when you delete your activity" and are retained for up to three years.
• Auto-delete. Activity is auto-deleted after 18 months by default; you can pick 3 or 36 months, or never.
• Keep Activity off. Future chats "won't be used to train our AI models, unless you choose to send Google feedback." They are still stored for 72 hours so Gemini can respond and for safety.
• Temporary Chats. Google says they "are not used to train Google's AI models," with the same 72-hour retention.
• Audio and Live video. Not used to improve Google services by default; that is a separate opt-in in Gemini Apps Activity.
• Work and school accounts. The hub notes these "may be subject to different data handling terms" under the Workspace privacy hub. This post is about personal accounts.

The part that surprised the poster is the coupling: the same switch that controls training also controls whether you get a saved history. The thread's blunt summary is that if you want your previous chats to stay visible to you, you cannot also opt out of training and review.

There is a second coupling worth knowing about as a builder. Google says tasks that use a remote browser or remote computer to execute code are "only available when Keep Activity is on," and that Workspace connected apps are unavailable with it off. The agentic features and the privacy setting are tied together.

The angle: pricing a privacy setting as a product tier#

Who wins and who loses here is mostly about positioning. Google's consumer bundle sells Gemini alongside storage and Workspace-style perks, so "paid" in the buyer's head means "I am the customer." The documentation says paid or not, personal-account chats feed improvement unless you change the toggle. That is a mismatch between expectation and policy, and it is exactly the expectation gap that tends to produce a backlash cycle every few months across vendors.

Winners: anyone selling a clean split between a consumer product and a zero-retention business product, and local models (see our best local coding LLMs). Losers: Google's own trust with the people most likely to evangelize its tools - developers - at the moment its models are competitive.

For builders there is a practical split worth drawing:

1. Consumer Gemini app, personal account. Treat as "may train and may be read" unless Keep Activity is off.
2. Gemini API. Different terms, and the free-versus-paid distinction does matter there. Our directory-backed comparison shows Google's unpaid API quota can be used to improve products outside the EEA, Switzerland and the UK, while paid tiers say otherwise. Re-read the terms for your tier; do not infer one surface from the other.
3. Work account. Governed by Workspace terms; confirm with your admin rather than guessing.

If you are choosing a Google plan for coding specifically, Google AI Pro vs the Gemini API covers which surface to use for which job.

What to do Monday (five minutes)#

1. Open Gemini Apps Activity and check whether Keep Activity is on.
2. If you paste proprietary code or client material into the app, either turn Keep Activity off or start those sessions as Temporary Chats.
3. If you keep Keep Activity on, shorten auto-delete to 3 months and delete old sessions that contained secrets. Remember human-reviewed copies are on a separate 3-year clock.
4. Move real work to a surface with contractual terms: the paid API tier, a business account, or a local model.
5. Never rely on a toggle to un-send a secret. Rotate any credential that went into a chat.

The same hygiene applies to Anthropic and OpenAI consumer products; see Fable 5 data retention for enterprise for how the Claude side is documented.

What people are actually saying#

The Tell HN thread (500+ points, 130+ comments at time of writing) is split in a useful way.

• "Expected, and documented." Several commenters say this is standard for free and consumer products and that the poster missed text Google does publish. One notes the training exclusion is a distraction, because Google's terms also grant a license to use content for operating the service and creating new features, so the training toggle is not the only way content is used.
• "The coupling is the problem." Others object that history and training are bundled: the only way to keep chats visible to yourself is to accept review and training, which one commenter calls an all-or-nothing pattern. A few say the human-review part is why they turned history off.
• The counter-case. A commenter points out that Gemini Enterprise shows a per-session note that chats are not used to improve models, so the "paying means you are the product" framing is wrong for business accounts. Another argues contractual zero-data-retention policies from large cloud vendors are believable because lying would invite lawsuits, and that local models sidestep the question entirely.
• A caution on trust. One commenter recalls a past case of a toggle being ignored by a different vendor's CLI, arguing that backlash risk keeps vendors honest, while another says companies do what benefits them when customers cannot verify. We cannot settle that from outside; it is why we label everything above as documented policy rather than verified behavior.

Continue Reading#

• Codex Training Setting Re-Enabled? Two Switches to Check - the OpenAI version of this check
• Which AI APIs Train on Your Data? - 18 providers checked against their own terms
• Fable 5 Data Retention for Enterprise - Anthropic's retention story
• Google AI Pro vs Gemini API for Developers - which Google surface fits which job
• Best Local Coding LLMs (2026) - the no-vendor-in-the-loop option

Sources#

• Google - Gemini Apps Privacy Hub
• Google - Manage Gemini Apps Activity and Keep Activity
• Tell HN - Gemini trains on chat history, even for subscribers (personal accounts)
• Google Terms of Service - permission to use content