Is your data used to train ChatGPT, Claude or Gemini?

You discuss a client problem, a draft contract, a product idea with ChatGPT, Claude or Gemini. Depending on your settings, that conversation may serve to train the next version of the model. It is neither a secret nor an anomaly: it is a default choice at several providers, one you can turn off in a few clicks once you know where to look.

What actually happens by default

On the main consumer assistants, your conversations may be used for training unless you have explicitly disabled the option. This is not specific to one player: a Stanford University study published in October 2025 analysed the privacy policies of six major companies in the sector (Amazon, Anthropic, Google, Meta, Microsoft, OpenAI — nearly 90% of the US chatbot market) and found that each processes conversation data for training by default, absent explicit user action to decline.

Concretely, your text is removed from the direct link with your identity (“de-identified”) before entering a training pipeline, but it is not ruled out that excerpts are reviewed by internal teams for quality and safety.

What each platform says, in plain terms

Platform	By default	How to disable	If you disable
ChatGPT (OpenAI)	Used if “Improve the model for everyone” is on	Settings > Data controls > turn off the option	New conversations are no longer used for training
Claude (Anthropic)	Used by default on Free, Pro and Max unless declined	Privacy settings > model improvement	Data kept 30 days for a safety check, not used for training
Gemini (Google)	Activity kept unless disabled	“My Activity” in Gemini Apps settings	History not kept for product improvement

This table reflects the consumer settings. Enterprise plans (ChatGPT Business/Enterprise, Claude for Work, Gemini via Workspace) generally have distinct contractual guarantees, often with training excluded by default — always check the specific terms of your subscription rather than those of the free account.

Why it matters, concretely

This is not a theoretical question as soon as you work with information that is not yours alone: client data, non-public internal figures, legal drafts. A poorly checked setting can push this information beyond the strict scope of your workstation, even indirectly and de-identified. It is the same caution reflex detailed in our confidentiality habits at work with AI: the risk is not that AI “steals” your data, but that it travels further than you expected by default.

The GDPR also frames these uses in Europe — see our breakdown on AI and the GDPR according to the CNIL for the regulatory part.

How to check your own setting, right now

1. Open the privacy settings of each tool you use regularly.
2. Look for a heading like “model improvement,” “Improve the model” or “activity.”
3. Disable it if you handle sensitive professional information.
4. For a one-off, very sensitive exchange, prefer a temporary or incognito mode rather than changing the global setting: at OpenAI, Temporary Chat is never used for training, even though it is still kept up to 30 days for safety. The same principle applies to the files you upload (Excel, PDF), governed by the same setting as the text of conversations.

This check takes less than two minutes per tool. It is worth redoing after an update to the terms of use, because these default settings have already changed in the past without the user necessarily being warned in a very visible way.

Key takeaway

None of the three platforms hides the information, but none puts the setting in front of you at first launch either. The simple rule: before pasting a piece of data you would not want reused elsewhere, a two-minute detour through the privacy settings changes everything.