Turn off model training? Check what Claude and Gemini still retain

Google's supplied Gemini activity illustration, 13 August 2025; sample settings. Source: Google. Google and Gemini are trademarks of Google LLC. PING is not affiliated with or endorsed by Google.

Before pasting a client document into a chatbot, check more than whether your chats train its models. Training, storage, feedback and safety review are different uses. Claude and Gemini offer controls, but turning one off is not a promise that nothing is retained or seen.

• Check your own account's training controls; do not assume a default.
• A training opt-out is not zero retention or a confidentiality agreement.
• Feedback and safety review have separate conditions.

Start with the right kind of account

This guide concerns personal Claude and Gemini accounts, not an employer's approved service. Anthropic's consumer training explanation covers Free, Pro and Max, including Claude Code used through those plans. Commercial products have different terms. Google's Gemini privacy hub likewise points work and school users to separate information.

Claude: a future-training choice

Anthropic documents the control under Settings, Privacy, with the label Help Improve our AI models. Inspect the setting on your own signed-in account. PING has not checked your current choice and is not asserting which position new accounts use.

Turning it off excludes new chats and coding sessions from future model training; previously stored conversations also stop being used in future training runs. It does not remove material from training already underway or undo a model already trained. Anthropic also preserves safety-related uses for flagged conversations. Treat that as a limited control, not a retrospective erase switch.

Anthropic says Incognito chats are not used to improve Claude, even with model improvement enabled. However, submitting thumbs-up or thumbs-down feedback can include the whole related conversation and involve retention for up to five years. Check what you are submitting, not just the small comment box.

Gemini: distinguish activity from retention

Google says future chats are not used to improve its models when Keep Activity is off and you do not submit feedback. Temporary chats are also excluded from model training. Its privacy hub nevertheless describes 72-hour retention of these chats for responses and safety, including possible human safety review. 'Not used for training' therefore does not mean 'immediately deleted'.

Feedback is an important exception: related content can be retained for up to three years, disconnected from the account. Previously human-reviewed conversations are also not removed by deleting activity and can remain for that period. Google directs users to Gemini Apps Activity; some connected apps are unavailable when activity is off.

Make a useful pre-upload routine

Before using either service, identify the account type, open the relevant control and read the explanation alongside it. Record the check date without copying private conversations. If the interface or policy differs, stop and consult current official guidance rather than following an old screenshot. Recheck after an account or plan change.

For a London freelancer working with a Nairobi client, the practical question is permission to disclose the material, not which city appears in the account. Do not upload confidential client files, subscriber lists or unpublished reporting simply because training is disabled. Ask the accountable organisation which tools and data uses it permits.

The payoff is a deliberate choice before sharing, not an absolute privacy claim. Use only material you are authorised to provide, consider whether you need to provide it at all, and remember that feedback changes the data-use question. More practical coverage is at PING's People desk.