ChatGPT, Claude, Mistral Vibe and the GDPR
ChatGPT, Claude and Mistral Vibe are powerful tools. ChatGPT and Claude come from American providers, Mistral Vibe from France. For all three, the same applies: whoever uploads a client file sends along everything in it. PrivaSearch sits in between: the AI only gets checked passages in which names, BSNs (Dutch citizen service numbers) and addresses have already been replaced, and the originals and the key stay in the EU.
What happens when you upload
An employee drags a client file into the chat and asks about the notice period. The whole document goes along, with name, BSN, address, health information and bank details, to a provider whose processing usually takes place outside the EU.
Personal data may only leave the EEA with a valid basis, such as the EU-US Data Privacy Framework or standard contractual clauses, plus an assessment of the risks. When someone "just uploads" a file, nobody has recorded that.
Free and personal accounts fall under consumer terms. There, the provider is itself responsible for what it does with the input; your organisation has no contract that limits it.
With ChatGPT Free, Plus and Pro, training on conversations is on by default unless the user turns it off. With the consumer versions of Claude, the user chooses; whoever says yes has conversations kept for up to five years. With Mistral Vibe Free, Pro and Education, conversations are used for training until the user turns that off. You don't know what a colleague has set on a private account.
No entry in the record of processing activities, no DPIA and no trace of what was shared. If a client or the supervisory authority asks about it, you have no answer.
Plans compared
Business plans cover much more than the consumer versions. One thing doesn't change: the complete client file goes to the provider.
| Topic | ConsumerChatGPT Free, Plus, Pro · Claude Free, Pro, Max · Mistral Vibe Free, Pro, Education | BusinessChatGPT Business · Claude Team · Mistral Vibe Team | EnterpriseChatGPT Enterprise · Claude Enterprise · Mistral Vibe Enterprise |
|---|---|---|---|
| Data processing agreement | No. The provider is itself responsible. | Yes. | Yes. |
| Training on your input | Depends on the user's setting. | Not by default. | Not by default. |
| Storage in the EU | Not with ChatGPT and Claude. With Mistral Vibe, yes by default. | Not by default with ChatGPT and Claude. With Mistral Vibe, yes by default. | Possible with ChatGPT Enterprise. With Claude only through third-party cloud platforms. With Mistral Vibe by default. |
| What goes along | The complete client file. | The complete client file. | The complete client file. |
| Contract | Personal account. | Per user, monthly or yearly. | Custom contract, usually with a minimum number of users. |
| Suitable for client files with personal data | No. | Only with a transfer basis, a DPIA and strict agreements. | Possible, but price and contract are made for large organisations. |
As of September 2026 for OpenAI and Anthropic and October 2026 for Mistral, based on their public terms and help pages. These change regularly: check them before you make a choice. Even with EU storage, OpenAI and Anthropic remain American companies; the US CLOUD Act can oblige such a company to hand over data, even if it is stored in a European data centre. Mistral is a French company, but there too, uploading sends the complete client file to the provider.
What about a European provider?
A European AI service largely removes the concern about transfers to the US. The rest of the GDPR still applies. You still need to be able to explain why the AI needed your client's name, BSN and salary to answer your question. Usually it didn't, and then they shouldn't go along (Art. 5(1)(c)). PrivaSearch takes them out before the AI sees the client file.
How PrivaSearch solves it
The problem isn't the AI, but that the raw client file goes along. PrivaSearch takes the personal data out before any AI is involved, and has a person check that this worked.
The MCP connection, step by step
MCP (Model Context Protocol) is the standard that lets an AI assistant use outside tools. For your assistant, PrivaSearch is such a tool: it can search, but not browse.
search_project_documents for a project and a search query,
for example "notice period".
read_document_source the assistant fetches one passage of at most 1,200
characters, again after the same permission check.
{
"source": "Document 3F9A1C",
"page": 4,
"excerpt": "[PERSOON_A4F2] may terminate
the lease with a notice period
of three months."
}Why this matters under the GDPR
What remains your job
Who it's for
Large organisations sign a custom contract with EU storage and their own lawyers. A law firm of twelve people, an accountancy firm, a healthcare practice or a small municipality usually can't. PrivaSearch is made for them: one price per organisation, set-up in days and the paperwork included. See how the pricing works.
Frequently asked questions
No product takes care of the GDPR for an organisation; you do that with measures, agreements and accountability. PrivaSearch provides the technical measures (pseudonymisation, mandatory review, permission checks, logging, retention periods, storage in the EU) and the documents to demonstrate them. The assessment of your situation, and the DPIA, remain yours and your data protection officer's.
Pseudonymised passages go to the provider of the AI assistant. Names, BSNs, addresses and other detected data have already been replaced by then, and a colleague has checked that. For you, it legally remains personal data, because you hold the key; for the recipient it cannot be traced back. That's why we recommend a business account with a data processing agreement, and you record the transfer in your register.
With anonymising, the way back is gone for good. With pseudonymising, the link still exists, but separately and secured. PrivaSearch pseudonymises on purpose: that way you can still see in your own environment who a passage is about, while the AI can't. For you, pseudonymised data still falls under the GDPR.
No. The link between pseudonym and person is kept in a separate vault with a key per organisation, and there is no MCP function for it. Revealing is only possible in the web environment of PrivaSearch, by a project member with write access, and every reveal goes into the activity log. In PrivaSearch, every project member can open the original of a document; that is logged too.
That can happen, because no model detects everything. That's why the review is mandatory: a colleague sees the original next to the pseudonymised text, highlights what was missed and only then approves. Until that approval, a document is not searchable, not even for a connected assistant.
Then that name goes along with the question to the provider of the assistant, because that happens outside PrivaSearch. If the assistant then searches in PrivaSearch, the name in the search query is replaced. The built-in AI training teaches employees exactly this: ask about the role or the pseudonym, not about the person.
It is a big step forward: there is a data processing agreement and your input is not used for training by default. But the complete client file still goes to the provider, there is no review of what's in it, and the plan doesn't control who may see the document. PrivaSearch and such a business account actually complement each other: the connection works best with a business account.
The transfer outside the EU is then less of an issue; the rest remains. The complete client file goes to the provider, on a personal account conversations are used for training until the user turns that off, nobody checks in advance what's in it, and nothing is recorded in your register. ‘In the EU’ says where your data is stored, not what you may put in: data minimisation applies to every provider.
Yes. Searching, opening sources and viewing passages can all be done in PrivaSearch itself. Semantic search then uses a European AI provider, and only on reviewed, pseudonymised text. The MCP connection is an extra, not a requirement.
Claude (in the browser, in the desktop app and in Claude Code), ChatGPT Business and Mistral Vibe, which was called Le Chat until May 2026. The connection uses the open MCP standard with OAuth 2.1, and access tokens expire after eight hours. You connect at your organisation's own address and approve the connection after you have signed in; there are no separate access codes. You can revoke it at any time on your settings page.
The AP warns that entering personal data into an AI chatbot can lead to a data breach, for example if it happens against the organisation's agreements. The AP received several notifications of this. Its advice: make clear agreements about what is and isn't allowed, and make sure employees know them.
In EDPS/SRB (4 September 2025), the Court of Justice ruled that pseudonymised data is not personal data in every case and for everyone. If a recipient cannot reasonably identify anyone, it may not be personal data for that recipient. For whoever holds the key, it does remain personal data, with all the obligations that come with it.
That depends on what you process and how much. If you work with special category data, such as health data, or on a large scale, a DPIA is often mandatory. PrivaSearch produces a DPIA starter document that already carries out the test of whether a DPIA is required, based on your own set-up. Have it reviewed by a lawyer or your data protection officer.
Originals, the pseudonym vault, the search index and detection run in data centres in the EU. Which parties are involved in which link is listed in the sub-processors overview.
We show it with fictitious client files from your sector: upload, detection, review, search and the connection with your AI assistant. Afterwards you'll know exactly what your team can use safely.
Book a demo →