A data protection officer is asked by the board to find "a European alternative to ChatGPT". The shortlist she finds names Vibe from Mistral, Lumo from Proton and Euria from Infomaniak. All three call themselves European. Reading the small print, the word turns out to mean something different for each.
The same week, a team lead asks whether ChatGPT is European by now. Doesn't OpenAI keep data in Europe these days?
Both questions hinge on the same word, and the answer depends on which part of "European" you have in mind.
#What a European AI chatbot settles for you, and what it doesn't
For an AI chatbot, European can refer to three things: where the provider is established, where your conversations and files are stored, and where the language model comes from. They do not always line up, and each answers a different question.
Establishment decides which law applies to the provider, and so which government can demand the data. Storage decides where the data physically sits. The model decides who processes your prompt, if the model does not run at the provider itself.
None of the three settles whether the tool trains on your input, which plan your organisation needs, or what an employee types into the box. For those, you look at the settings and terms of each tool, exactly as you would for an American one.
#What European means for Vibe, Lumo, Euria and ChatGPT
The sources are linked in the sections below, checked 11 November 2026.
#Where the provider is based decides which law applies
The US CLOUD Act requires a provider under US jurisdiction to disclose data in its possession, custody or control when the US government asks, "regardless of whether such communication, record, or other information is located within or outside of the United States" (checked 11 November 2026). The location of the server is irrelevant to that obligation. The company's establishment is not.
That is why Mistral is a stronger choice on jurisdiction than an American provider with a European data centre. Mistral is a French company, and according to Mistral's help centre (checked 11 November 2026) data is hosted in the EU by default. There are exceptions: if you deliberately use the US API endpoint, your data is stored in the United States, and depending on the feature, data may be transferred temporarily to sub-processors outside the EU. What that means in practice is covered in Mistral Vibe is the European option. It still cannot stop the paste..
Why the jurisdiction question carries so much weight, and how it collides with Article 48 of the GDPR, is the subject of Data sovereignty starts in the prompt field.
#Data in Europe with an American provider
The team lead is right that ChatGPT can keep data in Europe. According to OpenAI (checked 11 November 2026), a new ChatGPT Enterprise or ChatGPT Edu workspace can choose to store customer content at rest in Europe: conversations, prompts, uploaded files and custom GPTs. Since January 2026, eligible Enterprise and Edu customers can also have model inference run in Europe. OpenAI only mentions new workspaces here, and ChatGPT Business is not on the list.
For anyone who does not want data stored outside Europe, that is a real improvement. It does not turn ChatGPT into a European tool. OpenAI is still an American company, and the CLOUD Act looks at who controls the data, not where it sits.
Under the GDPR, transfers to such a provider are covered if it participates in the EU-US Data Privacy Framework. That gives the transfer a legal basis. It does not remove the possibility of a US order.
#Switzerland is in Europe, but not in the EU
On Lumo's privacy page (checked 11 November 2026), Proton writes that Lumo is "built and based in Europe". That is accurate: Proton is a Swiss company. So is Infomaniak, which makes Euria and says in its Euria guide (checked 11 November 2026) that processing takes place exclusively in Switzerland, on its own servers.
Switzerland is not an EU member, and the GDPR is not Swiss law. For an organisation in the EU, that matters less than it sounds. The European Commission has recognised Switzerland as providing adequate protection (checked 11 November 2026), so you can transfer personal data to a Swiss provider without an additional transfer mechanism. And the CLOUD Act's disclosure obligation does not reach a Swiss provider with no US parent.
If your own policy says "EU providers only", Lumo and Euria fall outside it, strictly read. If what the policy means is "no transfers without adequate protection", they do not. It is worth writing that distinction into the policy before someone requests a Swiss tool.
#The model can come from somewhere else entirely
A European chatbot does not always run on a European model. Proton says Lumo currently uses models including Qwen 3.5 and GLM 5.3, and Qwen is developed by Alibaba Cloud (checked 11 November 2026). Proton adds that the models run exclusively on servers it controls. Your prompt does not go to the model's developer: the model is downloaded software that Proton runs itself.
That distinction is the one that counts. If the model runs at the provider, what matters for your data is mainly the provider's own establishment. If a European tool forwards your prompt to another company's model, for example through an American provider's API, that company is a sub-processor and your prompt travels into its jurisdiction. A European front end on an American model is European for storage and not for processing.
You will not see the difference on the tool's homepage. You see it in the sub-processor list and the data processing agreement.
#What European leaves open
Training has nothing to do with establishment. Mistral's Vibe uses inputs and outputs for training by default (checked 11 November 2026); only Enterprise customers are opted out by default. Proton says Lumo does not use conversations to train models, and Infomaniak says Euria does not use generated content for training. Three European tools give three different answers, and with Vibe the answer also depends on the plan. How the large American tools compare is in ChatGPT Business vs free and personal AI accounts.
The GDPR applies whoever the provider is. Pasting a customer file into a chatbot is processing personal data, which needs a lawful basis, data minimisation and, for business use, a data processing agreement. A European provider makes the transfer question easier. It does not make a passport number in a prompt any less sensitive.
Then there is the employee. A tool approved because it is European feels safe, and that is exactly why people paste into it more freely. With Vibe, the jurisdiction can be right while the content of the prompt is not.
- Where is the provider established, and does it have a US parent or US presence? That decides whether the CLOUD Act applies to it.
- Where is the data stored, and which exceptions does the provider list? A different API endpoint, or sub-processors outside the EU.
- Which model does the tool run on, and who runs that model? The provider itself, or a company in another country.
- Does the tool train on input, on the plan you use? And who can turn that off: the employee or the admin.
- Is there a data processing agreement for business use? Without one, customer data does not belong in it.
The answers belong in a decision, not in an impression. In the BeeSensible AI register, every AI tool in the catalogue shows the vendor country, the hosting region, the model provider and whether the tool trains on input. Every tool starts on No decision yet; the organisation chooses Allowed, Limited use or Not allowed itself, and with Limited use it can record a condition such as "Enterprise only". When an employee opens the tool, the browser notice shows that condition. What happens next in the text box is a separate moment: a name or IBAN in a prompt is marked before the message is sent, in Mistral's Vibe just as in ChatGPT. Turning those decisions into a list employees can actually use is covered in How do you build a list of approved AI tools?, and for privacy teams setting this up there is AI governance for privacy officers.
European is a good argument in such a decision, and on jurisdiction often the strongest one. It is one line in the file, next to training, the plan and what may go into a prompt.
