Using Claude, ChatGPT or Mistral on a client file: what leaves, what stays

Lawyers, notaries and doctors all ask the same question before pasting a file into an AI: where does it go, who keeps it, and for how long? The answer depends on the provider, on the plan, and on settings that have changed several times since 2025. Here is the state of play, from each provider's own pages.

1. The summary table

Sources: OpenAI (data controls, retention, business plans, residency); Anthropic (training, retention, organisations, DPA); Mistral (training, location); Microsoft (Copilot).

2. Five points that are easy to miss

Mistral now trains by default on its individual and Team plans

According to Mistral's help page (updated 31 August 2026), on the Free, Pro, Student and Team plans, input and output data are used by default to train its models unless you opt out. Only Enterprise is excluded by default. The setting is in the privacy settings, and a Team admin can switch it off for the whole organisation (how to). A French provider and EU hosting do not, on their own, tell you how your conversations are used.

Claude's consumer plans: a setting decides training and retention

Since Anthropic's 2025 consumer terms update, conversations on the Free, Pro and Max plans are used to train models when the "Help improve Claude" setting is on, and may then be kept for up to five years in de-identified form; otherwise, a deleted conversation leaves the servers within 30 days (retention). Team, Enterprise and API data are not used for training.

Claude Cowork reads your folders, and sends what it reads

Cowork works on folders on your computer. According to its architecture overview, in cloud sessions the agent's work, including any local files it opens through the desktop app, is processed on Anthropic's servers. Even in a local session, it is the Claude model, hosted by Anthropic, that reads the files' contents. Anthropic itself advises against granting access to local files with sensitive information (use Cowork safely).

Deletion is not always the end: legal holds and flagged content

In the New York Times lawsuit against OpenAI, a federal judge ordered on 13 May 2025 the preservation of all ChatGPT output logs, deleted conversations included. An order of 9 October 2025 ended that duty for new data, but logs already preserved are still held, as is data from accounts the newspaper flags (Engadget). The case is ongoing. Separately, Anthropic says it keeps conversations that violate its usage policy for up to two years, on every plan (retention). A deletion setting does not cover every case.

Copilot's guarantees apply to work accounts

For Microsoft 365 Copilot, prompts, responses and data accessed through Microsoft Graph are not used to train foundation models, and the service falls under the EU Data Boundary, except models supplied by Anthropic, which are off by default for EU tenants (Microsoft Learn). These commitments apply to work accounts, not to consumer Copilot.

3. What this means for a professional bound by secrecy

A business plan with no training and a DPA is better than a consumer account. It does not change the basic fact: the file is still transmitted to a third party that is not bound by your secrecy, under policies it can change. That is why the French Bar's 2026 ethics guide asks lawyers to systematically pseudonymise or anonymise data before any use of generative AI, and why the French health authority's guide on generative AI in healthcare (2025) asks professionals to check that no information allowing direct or indirect identification, or covered by medical secrecy, is shared with a tool whose confidentiality is not assured (our translation).

4. The workflow that keeps the file at home

1. Pseudonymise locally. Hexagone AI detects and replaces names, companies, addresses, amounts and references on your own Mac or PC. The same person keeps the same pseudonym across the folder.
2. Review the protected copy. Every detection appears in the preview pane, and you can edit anything before exporting.
3. Point your AI at the protected folder. Hexagone AI can maintain a protected mirror of a folder that follows every change. Give Claude Cowork, ChatGPT or Mistral that mirror, never the original.
4. Re-identify the answer locally. The mapping never leaves your computer; the real names are put back on your machine.

Whatever the provider does with what it receives, what it receives no longer identifies your client or your patient.