ChatGPT and GDPR: can you put client data into it?

Who is responsible for the data sent to ChatGPT?

You are. When a law firm, a notary office or a company pastes into ChatGPT a letter, a contract or a spreadsheet containing the names of clients, employees or counterparties, it decides on this processing of personal data and its purpose: it is the controller within the meaning of the GDPR. The resulting obligations remain its own, whatever the tool.

The CNIL (the French data protection authority) says so plainly in its Q&A on generative AI: it is the user organization “that will be held liable” if its staff misuse AI. And as a user, “it is therefore up to you to question the provider about its system’s GDPR compliance” (our translation).

Data about a client company (its name, legal form, contact details) is not personal data (Recital 14 of the GDPR). But its directors, partners, employees and contacts are individuals. And for professions bound by secrecy, the client’s name remains covered by professional secrecy: see AI and professional secrecy.

Is ChatGPT GDPR-compliant?

The useful question is not whether ChatGPT is “compliant” in general, but whether your use of it is: you are the one accountable. European data protection authorities are watching the service closely. In Italy, the Garante per la protezione dei dati personali fined OpenAI 15 million euros (decision no. 755 of November 2, 2024, announced on December 20, 2024), in particular for training ChatGPT without an adequate legal basis and for failing to notify a data breach. OpenAI challenged the decision: the Court of Rome upheld its challenge in judgment no. 4153/2026, published on March 18, 2026, and the decision was removed from the Italian authority's website.

These proceedings concern OpenAI's own processing. They do not settle your own compliance: if an associate pastes a client's contract into a personal account, the firm is the one that must justify that transfer. Hence the three questions that follow: which plan, which contract, which data.

Which ChatGPT plan is compatible with the GDPR?

What OpenAI publishes for each family of plans, as of October 12, 2026. Terms change often: check them on the pages cited in the sources before any decision.

On training, OpenAI writes that, by default, it does not use data from ChatGPT Enterprise, Business, Edu and its API platform, “including inputs and outputs”, to train or improve its models. On individual plans, a deleted conversation is erased within 30 days, with exceptions (a conversation already de-identified, a security or legal obligation).

One exception to know: even after turning off training, a thumbs up or thumbs down on an answer allows OpenAI to use the whole conversation concerned to train its models.

The data processing agreement (DPA, Article 28)

When a provider processes personal data on your behalf, Article 28 of the GDPR requires two things: that it provide “sufficient guarantees”, and that a contract set out what it may do with this data. At OpenAI, this contract is the Data Processing Addendum (DPA), offered for ChatGPT Business, ChatGPT Enterprise and the API; the customer requests it through a form. In this framework, the customer is the controller and OpenAI the processor.

OpenAI offers its DPA only for ChatGPT Business, ChatGPT Enterprise and the API: Free, Go, Plus and Pro accounts fall under its general privacy policy. Using a personal account for client data therefore means handing over that data without the contract required by Article 28. It is also the typical case of uncontrolled use: an employee who opens ChatGPT on their personal account to save ten minutes.

A DPA is not a blank check. Read it in full: retention periods, sub-processors, and any clauses on “de-identified” or aggregated derived data. The April 2026 guide of the CNB (Conseil national des barreaux, the national bar council) also points out that a no-reuse commitment is generally “neither provable nor verifiable” (our translation).

Hosting and transfers outside the EU

OpenAI is an American company. To meet European requirements, it offers data residency at rest in Europe (EEA and Switzerland): conversations, files and custom GPT data for eligible ChatGPT Enterprise and Edu customers, and API projects created in the Europe region. An option for processing in Europe also exists for some customers.

Three limits:

• Individual plans are not covered.
• The scope is partial: account data, billing data, usage statistics and logs without content stay outside residency.
• Sovereignty is not settled for all that. The CNB guide uses as a criterion hosting by a French or European company not subject to extraterritorial laws, which as things stand excludes American companies.

What does the CNIL say about ChatGPT and personal data?

In its Q&A on using a generative AI system, the CNIL gives advice that applies directly (our translation):

• With a consumer service, “never share confidential information such as personal data”; users “should only submit information they are authorized to share”.
• To provide client or employee data, it seems “generally more appropriate and more secure to favor deploying ‘on-premises’ solutions”.
• When using external infrastructure, secure it with a data processing agreement.
• Depending on the deployment mode, prohibit entering certain confidential or personal data, govern use through an internal policy and train users.

In other words: the GDPR does not prohibit ChatGPT, but it requires deciding upstream which data may go into it. The simplest answer to defend is: no personal data in the clear.

What is the DPO's role with ChatGPT?

When a firm or a company has appointed a data protection officer, the use of ChatGPT goes through them. Article 39 of the GDPR entrusts them at least with “informing and advising” the controller and employees, monitoring compliance with the regulation and internal rules, “including […] awareness-raising and training of staff”, and advising on the impact assessment. In practice, five workstreams:

Pseudonymization on the computer simplifies each of these workstreams: if only copies without names go to the AI, the record, the impact assessment and the information given to data subjects concern a much lower risk. It is also the easiest measure to demonstrate during an inspection. For in-house legal teams and DPOs: document anonymization software for in-house counsel and DPOs.

Minimization, pseudonymization, anonymization

Article 5 of the GDPR requires data to be “adequate, relevant and limited to what is necessary in relation to the purposes for which they are processed”. To rephrase a letter, summarize a contract or review legal briefs, the AI needs no name at all. Sending the names anyway goes beyond what is necessary.

To work with the AI and then put the real names back into its answer, pseudonymization is the right fit: stable placeholders, such as [PERSONNE_1], and a mapping table that never leaves your computer. The details: anonymization vs. pseudonymization under the GDPR. To choose a tool: comparison of anonymization software.

How to use ChatGPT with client data: the method

1. Choose the plan. For regular professional use, a plan with a DPA (Business, Enterprise or API) rather than personal accounts. Sign the DPA and read it.
2. Write down the rule. A use policy: which accounts, which tasks, which data never in the clear. Train the team.
3. Pseudonymize on the computer. Before sending, never inside the AI: names, organizations, addresses, email addresses, phone numbers, IBANs, private dates, ID numbers. With AnonymFiles, they are detected automatically and replaced with stable placeholders across all documents.
4. Review the copy. Next to the original, looking for indirect identifiers: an amount, a reference, a project name. Nothing is exported without your approval.
5. Send, then check the answer. The AI can make mistakes: figures, references and quotations must be checked.
6. Re-inject on the computer. The app puts the real names back into the answer, as text, Word or PDF. The mapping table never leaves the machine.

On the network side, AnonymFiles works with no account and no telemetry; the license is checked on the computer, with no network call. On a Mac, the detection models are included: no connection at all. On Windows, they are downloaded once, at the first processing run, without any document text. See what the app does and does not do on the network and how to check the offline operation yourself.