AI confidentiality: what data can you entrust to ChatGPT, Claude or Gemini?

An architect loads a competition brief into a general-purpose AI to save time writing the submission. Three weeks later, they realise they forgot to check one setting: by default, their conversations are used to train the model. The brief, covered by a non-disclosure agreement, has potentially fed a system they no longer control. This scenario repeats every week in practices, and AI confidentiality remains the topic few professionals take the time to address seriously.

This article distils, for professional use, what you need to know before sending a file or a prompt to an AI tool. It builds on our *Annexe Confidentialité et privacy des outils IA* (Educasium, Pablo Nastar, May 2026), compiled tool by tool. Vendors' policies move fast: the principles below remain valid, but the precise settings must be checked with each vendor at the time of use.

Table of contents

1. Three texts frame your use of AI
2. The question to ask before each tool
3. The change few users noticed (Claude, October 2025)
4. Tool by tool: the summary
5. What you can entrust, and what you must never put in
6. The three-step decision grid
7. The limits of this article
8. Frequently asked questions

Three texts frame your use of AI

Before choosing a tool, you need to know the framework. Three texts stack up, and the third is the least well known.

The RGPD, the European baseline

The RGPD (EU Regulation 2016/679) governs any personal data: name, contact details, identifiable photo, plan with a precise address. Such data may only be processed for a clear purpose, on a legal basis, in minimal quantity, with a right to erasure. Practical consequence: you cannot freely load a client's named data into a general-purpose AI without precaution. And if the AI is hosted outside the EU, the vendor must provide equivalent safeguards.

The AI Act, the AI-specific framework

The AI Act (EU Regulation 2024/1689) is the first European text specific to artificial intelligence, entering into force in stages between 2025 and 2027. For common professional uses, the main points to remember are the transparency obligation when content is AI-generated, and the documentation required from providers of systemic-risk models such as ChatGPT, Claude or Gemini.

The Cloud Act, the text people forget

The Cloud Act (US law of 2018) allows US authorities, under certain conditions, to access data held by American companies, even when stored in Europe. This point directly concerns Google, Microsoft, Anthropic and OpenAI. For highly sensitive data (strategic competition, audit, litigation), it pushes you towards a European alternative such as Le Chat by Mistral AI.

The question to ask before each tool

Rather than a list of tips, a verification grid that applies to any tool, new or old.

Where are the data stored? EU, United States, elsewhere? Hosting determines the applicable legal regime.

Is the vendor subject to the Cloud Act? An American company, or one controlled by an American company, is. That changes everything for sensitive data.

Are conversations used to train the model? And is there an opt-out? This is the central question, and the answer has changed for several tools in 2025.

How long are data retained? And can they be deleted? Deletion is never instantaneous with the large vendors.

Does the professional version bring real guarantees? A Data Processing Agreement (DPA) that complies with the RGPD is the element that turns a consumer tool into a defensible enterprise tool.

The simple rule that follows: free versions carry maximum risk. Professional versions, provided the right settings are activated and the DPA is signed, carry reduced risk.

The change few users noticed (Claude, October 2025)

Claude, published by Anthropic, long held a reputation as a tool more cautious than others with data. That positioning has changed, and most users have not noticed.

Before October 2025, Anthropic did not use user conversations to train its models, except with explicit feedback. Since 8 October 2025, the policy is reversed on the Free, Pro, Max and Claude Code plans: a toggle "You can help improve Claude" appears. If you accepted it, or ignored it without unchecking it, your conversations can feed the training of future models, with retention of up to five years. If you deactivate it, retention drops to 30 days and there is no training.

The Team, Enterprise and API plans are not concerned by training by default and benefit from contractual guarantees. The immediate action, if you use Claude Pro in a professional context, is therefore to open the privacy settings and check this toggle. The same vigilance applies to ChatGPT, where the opt-out must also be activated manually on Plus versions.

Tool by tool: the summary

The table below condenses the key points per tool, for professional use (state as of May 2026, to be verified with the vendor).

Reading the table: for sensitive data, the ideal remains Mistral Le Chat (EU sovereignty) or NotebookLM (closed corpus, the AI only cites your documents). For daily practice, Claude Pro or ChatGPT Plus with opt-out activated, and Midjourney Pro for images.

What you can entrust, and what you must never put in

The following classification is the most useful part day to day. It distinguishes three levels.

What you can entrust without particular precaution

Generic methodology questions, public documents (PLU available online, open standards, press articles), your anonymised notes, drafting sketches, spelling or grammar review requests. No identifiable data, no secret.

What you can entrust with a professional version and opt-out activated

Meeting notes (with participants' consent), anonymised internal technical studies, your technical submissions in progress, plans and diagrams without client identifiers (address removed, project name anonymised). Provided you are on a version where data is not used for training.

What you must never entrust to a general-purpose AI

Identifying personal data (names, contact details, bank details, social security numbers, health data), administrative documents signed by a client, confidential competition briefs under non-disclosure agreement, cadastral plans that could be published without explicit consent, legal documents in litigation. For these cases, the European alternative or stepping away from AI is required.

One particular case deserves to be named: recording a meeting to generate minutes. You must inform all participants that the meeting is being recorded and that the transcription will be processed by an AI. This is a legal obligation. Failing that, it is a fault.

The three-step decision grid

Before sending anything, ask yourself two questions. If this content leaked tomorrow into a public blog, would it be a problem for my client or for me? And am I on a version where my data is not used for training?

Choose the tool according to sensitivity. Low sensitivity: any tool with opt-out activated. Medium sensitivity: Pro or Team version with a DPA that complies with the RGPD. High sensitivity: European version (Mistral) or no AI at all.

Document your choices. If questioned by a client, an insurer or an authority, you must be able to explain how you used AI, with what precautions, and why you believe you met your obligations.

The limits of this article

This article is not legal advice. For sensitive uses (health data, litigation files, security-critical contracts), consulting a legal adviser or a data protection officer remains necessary. Vendors' policies evolve quickly, and what was true in May 2026 may no longer be so by the time you read these lines. Always check the settings and the terms of service of the tool at the time of use.

Besides, AI is never mandatory. For certain high-stakes tasks, the right decision remains not to use AI, or to limit yourself to a closed corpus like NotebookLM. The productivity gained does not always justify the risk taken on data that does not belong to you.

Frequently asked questions

Can ChatGPT be used with confidential company data?

Yes, provided you use a Team or Enterprise version that disables training on your data, and that you have signed the DPA offered by OpenAI. With the free version or Plus without opt-out activated, the risk is real and discouraged for any client, contractual or strategic data. The practical rule is to systematically anonymise sensitive elements before submitting them, even on a professional version.

What data should never be sent to an AI?

Identifying personal data (names, contact details, health data, bank details), signed administrative documents, confidential briefs under non-disclosure agreement, legal documents in litigation and cadastral plans that could be published without the client's explicit consent. For these categories, no general-purpose AI offers sufficient guarantee.

Does the Cloud Act apply to ChatGPT, Claude and Gemini?

Yes, these three tools are published by American companies (OpenAI, Anthropic, Google) subject to the Cloud Act, which allows US authorities to access certain data even when stored in Europe. For highly sensitive data, a European alternative such as Le Chat by Mistral AI, not subject to the Cloud Act, is more defensible.

Does Claude use my conversations to train itself?

Since 8 October 2025, on the Free, Pro, Max and Claude Code plans, yes, by default, unless you deactivate the toggle "You can help improve Claude" in the privacy settings. The Team, Enterprise and API plans are not concerned by training and benefit from contractual guarantees. If you use Claude professionally, checking this setting is the first action to take.

Does an AI training course cover confidentiality and the RGPD?

A serious professional AI training course necessarily covers these topics, because mastery of the tools comes through mastery of the settings and of the legal risks. At Educasium, confidentiality and the RGPD are addressed within the programmes, based on our tool-by-tool annex. It is also what distinguishes a cobbled-together use from one that is defensible in a business context.

Going further

AI confidentiality does not depend on the tool, but on the version chosen, the settings activated and the use you make of it. Choosing the right tool for the right case is a skill, and it can be acquired. To choose between tools based on your uses, see our comparison ChatGPT ou Claude en entreprise. To structure your requests and reduce information leaks in prompts, see our formation au prompt engineering.

Qualiopi-certified programmes, eligible for OPCO and FIFPL funding. To train your team on good practices and the right settings, contact Educasium and specify your sensitive use cases.