Does Google train Gemini on your data?

Training default

Your content is used for training unless you opt out.

Opt-out

No standalone training toggle (unlike ChatGPT and Claude): the only opt-out is turning off 'Keep Activity' at myactivity.google.com/product/gemini — which also stops saving your chats. Future chats then don't appear in your Activity and are kept only 72 hours. Temporary Chats and signed-out use avoid training per-chat.

Retention

Saved activity auto-deletes after 18 months by default (options: 3 or 36 months, or auto-delete off). — With Keep Activity off (or in Temporary Chats), chats are still retained with the account for 72 hours. Human-reviewed chats are kept up to 3 years, disconnected from the account, and are NOT deleted when you delete your activity. Audio and Gemini Live video/screenshares aren't used for improvement by default (separate opt-in setting; Live transcripts still follow Keep Activity).

Human review

A subset of chats is reviewed by human reviewers (including Google's trained service providers); reviewed chats are retained up to 3 years, disconnected from the account. Google's own warning: don't enter confidential information you wouldn't want a reviewer to see or Google to use to improve its services.

Feedback carve-out

Yes — even with Keep Activity off, submitting feedback sends the feedback plus up to the last 24 hours of your chats (including uploads and Connected Apps data); reviewed feedback is retained up to 3 years, disconnected from your account.

Keep Activity is on by default for users 18+. Sub-surfaces (Gemini in Chrome, the Google app, Messages) inherit this policy. EEA/UK/Switzerland: GDPR bases differ (legitimate interests for model improvement; consent for some features).

Training default

Your content is used for training unless you opt out.

Opt-out

Same as the free tier — no standalone training toggle: turning off 'Keep Activity' is the only opt-out, and it also stops saving your chats (72-hour retention only).

Retention

Same as the free tier: 18-month auto-delete default (options: 3/36 months or off). — 72-hour retention with Keep Activity off; human-reviewed chats kept up to 3 years, disconnected from the account.

Human review

Same shared Gemini Apps Privacy Notice as the free tier: a subset of chats is human-reviewed and retained up to 3 years, disconnected from the account.

Feedback carve-out

Same as the free tier — feedback ships up to the last 24 hours of chats into a 3-year, account-disconnected review pipeline, even with Keep Activity off.

Renamed from Google One AI Premium. No Google page states paid consumer tiers differ from free on training or review — the shared Privacy Notice mentions paid plans only as a collected-data category ('Subscription information'), with no carve-out; differences are quota/feature only (e.g. expanded context, higher limits). Treat 'identical to free' as inferred from that absence.

Confidence: inferred — see notes; this cell rests on inference from primary sources, not an explicit provider statement.

Training default

Your content is used for training unless you opt out.

Opt-out

Same as the free tier — no standalone training toggle: turning off 'Keep Activity' is the only opt-out, and it also stops saving your chats (72-hour retention only).

Retention

Same as the free tier: 18-month auto-delete default (options: 3/36 months or off). — 72-hour retention with Keep Activity off; human-reviewed chats kept up to 3 years, disconnected from the account.

Human review

Same shared Gemini Apps Privacy Notice as the free tier: a subset of chats is human-reviewed and retained up to 3 years, disconnected from the account.

Feedback carve-out

Same as the free tier — feedback ships up to the last 24 hours of chats into a 3-year, account-disconnected review pipeline, even with Keep Activity off.

Formerly the 'Gemini Advanced' branding (retired). Pro buys quota and features (1M-token context, ~4x limits, Gemini in Workspace apps) — no Google statement offers paying consumers reduced training or review. Inferred identical to free from the single shared notice.

Confidence: inferred — see notes; this cell rests on inference from primary sources, not an explicit provider statement.

Training default

Your content is used for training unless you opt out.

Opt-out

Same as the free tier — no standalone training toggle: turning off 'Keep Activity' is the only opt-out, and it also stops saving your chats (72-hour retention only).

Retention

Same as the free tier: 18-month auto-delete default (options: 3/36 months or off). — 72-hour retention with Keep Activity off; human-reviewed chats kept up to 3 years, disconnected from the account.

Human review

Same shared Gemini Apps Privacy Notice as the free tier: a subset of chats is human-reviewed and retained up to 3 years, disconnected from the account.

Feedback carve-out

Same as the free tier — feedback ships up to the last 24 hours of chats into a 3-year, account-disconnected review pipeline, even with Keep Activity off.

Top consumer tier (sold in 5x/20x variants): up to 20x Pro's usage limits plus early features (Deep Think, Gemini Spark) — quota and features only; data treatment inferred identical to free from the single shared notice.

Confidence: inferred — see notes; this cell rests on inference from primary sources, not an explicit provider statement.

Training default

Your content is not used for training unless you explicitly opt in.

Opt-out

Nothing to opt out of — no-training is the contractual default.

Retention

Admin-determined: Gemini in Workspace prompts/responses 90 days to indefinite; the Gemini app under Workspace up to 36 months (default 18). — With conversation history off, new chats are kept up to 72 hours. Feedback is retained up to 18 months and not used to train Workspace generative AI models. EU/US data-region controls reach organizational-unit level; Client-Side Encryption can block Gemini access to content entirely.

Human review

Content is not human-reviewed or used for generative AI model training outside the customer's domain without permission.

Feedback carve-out

Voluntary in-product feedback is not used to train the generative AI models supporting Workspace; retained up to 18 months.

The controlling commitment is the Workspace Service Specific Terms §12.11 'Training restriction'; prompts are Customer Data under the Cloud Data Processing Addendum.

Training default

Your content is used for training unless you opt out.

Opt-out

No in-product opt-out — the path out is the paid tier (link an active Cloud billing account). EEA/Switzerland/UK: the paid no-training terms apply automatically, even to unpaid quota.

Retention

Unclear — no fixed retention window is published for unpaid services; submitted content and generated responses are used to provide, improve and develop Google products and ML technologies. — Content uploaded to model tuning is used only for tuning (per the API terms); tuning is no longer offered in the Gemini API/AI Studio since May 2025. An account with active Cloud billing (or a Workspace enterprise account) makes AI Studio count as a Paid Service.

Human review

Human reviewers may read, annotate and process unpaid API input and output (disconnected from account, API key and Cloud project first). Google's own instruction: 'Do not submit sensitive, confidential, or personal information to the Unpaid Services.'

Feedback carve-out

Not separately specified — unpaid content is already used for product improvement by default.

The training default is the Gemini API Additional Terms' unpaid-services clause (effective March 23, 2026). Unpaid Services also can't serve API clients to EEA/CH/UK users.

Training default

Your content is not used for training unless you explicitly opt in.

Opt-out

Nothing to opt out of — no-training is the contractual default.

Retention

Prompts and responses are logged for 55 days solely for abuse monitoring (Prohibited Use Policy enforcement), then deleted. — Abuse logs aren't used to train models other than policy-enforcement classifiers and may be stored or cached transiently in any country where Google maintains facilities. Grounding with Google Search or Maps stores prompts, context and output for 30 days. Interactions/Live API keep conversation state unless explicitly disabled (store=false; omit SessionResumptionConfig).

Human review

No product-improvement review of paid-tier content; logged abuse-monitoring data is processed for policy enforcement only.

Feedback carve-out

Not separately specified in the API terms.

Processed under the Data Processing Addendum for Products Where Google is a Data Processor. Zero data retention is per-project on an approved request: user content and identifiable metadata (IP addresses, account IDs) are then cleared prior to logging.

Training default

Your content is not used for training unless you explicitly opt in.

Opt-out

Nothing to opt out of — no-training is the contractual default.

Retention

No default at-rest retention of prompts/responses; Gemini models cache inputs/outputs in-memory only (24-hour TTL, project-isolated, disableable per project) — which per Google's current docs does not violate zero data retention. — Abuse monitoring: flagged prompts are stored up to 90 days in the customer's selected region — Google Cloud Master Agreement customers are exempt from this logging by default; others can request an exemption. Grounding with Google Search/Maps stores data (Maps storage can't be disabled — use Web Grounding for Enterprise for strict ZDR). Request-response logging is off by default (writes to BigQuery if enabled). Some advanced/partner models require extra logging under an Advanced AI Safety Addendum.

Human review

Only authorized personnel assessing abuse-flagged prompts; no training or product-improvement use of customer data.

Feedback carve-out

Not separately specified — customer data is governed by the Cloud Data Processing Addendum.

Google's docs now brand this 'Gemini Enterprise Agent Platform' (Vertex AI renamed; same data-governance terms and URLs). Region pinning keeps ML processing in the selected US/EU multi-region. Note a 2026 doc change: the 24-hour in-memory cache is no longer described as breaking ZDR — older 'disable caching for ZDR' guidance is obsolete.