← Tous les articles

GDPR and AI: what data you can send to an AI model, and where it goes

Under the GDPR you can send personal data to an AI model when four things are true: you have a lawful basis for that use, you have a data processing agreement with the provider, you know where the data is processed and kept, and you send only what the task needs. In practice that rules out pasting customer data into consumer chatbots, and it makes the choice of provider and region a legal decision, not a technical detail. As of September 2026, for example, Anthropic's own API offers no EU-only processing, while OpenAI and Microsoft Azure offer European processing under conditions. And the copies you make along the way, embeddings, logs and memory, are personal data too.

Can a business put personal data into ChatGPT or Claude?

A business can put personal data into ChatGPT, Claude or similar models only through business channels with a data processing agreement, never through a personal account. The consumer apps have their own terms and training settings. The business routes differ: through OpenAI's API, data is not used for training by default and abuse-monitoring logs are kept for 30 days by default (OpenAI, your data). Anthropic's data processing addendum names the customer as controller and Anthropic as processor. Either way, the lawful basis, the purpose and the retention are your responsibility as the controller, and your customers' privacy notice has to say which provider processes their data and why.

Where is data processed by the major AI providers?

Where data is processed depends on the provider and the route, and it changes often, so check it on the day you sign. As documented in September 2026:

  • Anthropic, first-party API: inference runs globally or, at a small surcharge on recent models, in the US. There is no EU-only option (Anthropic, data residency).
  • OpenAI: offers storage and processing in Europe (the EEA and Switzerland) for organisations approved for modified abuse monitoring or zero data retention. Some features, including hosted vector stores and conversation storage, keep data until it is deleted and fall outside zero retention.
  • Microsoft Azure: an EU "data zone" deployment processes data only within EU member states, Microsoft acts as the processor, and OpenAI does not receive the data (Microsoft, data privacy).
  • AWS Bedrock: retention is configured per region, and with cross-region inference enabled, retained inputs and outputs are stored wherever the request was actually processed.

If a client requires processing inside the EU, that requirement decides the route. Promising EU-only processing on a provider that does not offer it is a contract problem waiting to happen.

What should never go into an AI prompt?

What should never go into an AI prompt is anything the task does not need. Data minimisation is a GDPR principle, and it is also good engineering, because less data means less to leak and fewer tokens to pay for. In particular, keep out of prompts unless the task truly requires them: special categories such as health data, tax and social security numbers (ΑΦΜ and ΑΜΚΑ), bank account numbers, identity card numbers, and data about third parties you have no basis to process. A support agent that answers "where is my order" needs the order number, not the customer's tax number. Writing these rules down and training staff on them is also part of the AI literacy measures the AI Act expects, as explained in AI literacy under Article 4.

How do you remove personal data before it reaches the model?

You remove personal data before it reaches the model with validators for structured identifiers, a name detector for free text, and pseudonymisation that can be reversed only by authorised people. For Greek data this needs Greek-specific work:

  • Tax numbers (ΑΦΜ) have a check digit, which makes them detectable, but a popular validation library accepts the all-zero placeholder as valid, so reject it explicitly.
  • Social security numbers (ΑΜΚΑ) are eleven digits with a check digit and a date of birth in the first six, so validate both.
  • Greek IBANs are 27 characters with a mod-97 check and are almost never a false positive.
  • Names and addresses need a Greek-capable entity detector. The best free Greek model scores around 77% F1 on names and carries a non-commercial licence, so a commercial product needs another route and a labelled Greek test set.

Replace detected values with placeholders, keep the mapping outside the AI system, and restore the real values only for the people allowed to see them.

Are embeddings and AI logs personal data?

Embeddings and AI logs are personal data when the text they came from was. Research has shown that text can be reconstructed from its embedding: in one study, about 92% of 32-token inputs were recovered exactly from their vectors (Morris et al., 2023). A vector index is therefore a copy of your documents, not an anonymous set of numbers. The same holds for prompt and response logs, agent memory, evaluation sets and caches. When someone exercises the right to erasure, deleting the original record is not enough: every derived copy has to go. The cheap way to make that possible is to attach a subject key to every derived item at the moment it is created. How to design privacy into a build from day one is in GDPR without the drama.

When does an AI project need a DPIA?

An AI project needs a data protection impact assessment when the processing is likely to result in a high risk to people, which covers most AI uses that profile people, make decisions with legal or similar effects, monitor systematically, process special categories, or index large amounts of text about identifiable people (GDPR Article 35). Even when you conclude a DPIA is not required, keep the note that says why. An incomplete assessment is itself a finding in an audit.

When should AI run on your own machines instead?

AI should run on your own machines when the data is too sensitive to leave them or no cloud route meets the client's rules. Small models make this more practical than most people assume. The intent classifier I trained for watch it think is a 5.28 MB model that runs entirely in the browser, so nothing typed leaves the page. My 72-million-parameter assistant runs on a phone's processor. Small local models cannot match the large hosted ones on hard tasks, so the usual design splits the work: sensitive steps such as detection and classification run locally, and only minimised or pseudonymised text goes to a large model.

Who enforces the GDPR and the AI Act in Greece?

In Greece the Hellenic Data Protection Authority enforces the GDPR, implemented by Law 4624/2019, and is now also the main supervisory authority for the AI Act under Law 5321/2026 (HDPA legal framework). A complaint about an AI feature can therefore raise both laws at the same desk. The AI Act side is explained in the EU AI Act for Greek businesses. Reviewing documents against both laws, with every citation resolved to its source, is one of the systems under AI for legal and compliance.

Frequently asked questions

Is it legal to use ChatGPT with customer data?

It can be, through a business plan or the API with a data processing agreement, a lawful basis, a stated purpose and minimised data. Pasting customer data into a personal ChatGPT account is not a compliant setup.

Do I have to tell customers that AI processes their data?

Yes. The privacy notice should name the AI feature, its purpose, the provider and where processing happens, and how long each kind of data is kept. If customers talk to an AI system, the AI Act also requires telling them so at the start.

Can I use AI on employee data?

With care. Employee data needs a lawful basis that works in an employment relationship, where consent is rarely valid, and AI used to evaluate employees is high-risk under the AI Act from December 2027.

If you want an AI feature that passes a data protection review the first time, tell me what data it would touch and who the client is. The route, the redaction and the records are decided before any code is written.