Key takeaways
- No LLM is GDPR compliant on its own. Your use of it is, which depends on your lawful basis, a processor contract, where inference and logs run and what personal data you send.
- Open-weight models let you run inference on servers you control. You still owe GDPR duties as the controller, and you must read the licence.
- The AI Act's general-purpose model duties (Articles 53 and 55) fall on the model's provider and have applied since 2 August 2025. Your own duties depend on your use case.
- EDPB Opinion 28/2024 says that if you deploy a model that retains personal data, you should assess whether it was developed unlawfully. Ask providers what they can show.
What makes an LLM GDPR compliant?
No language model is GDPR compliant on its own. Compliance describes how you use it: your lawful basis (Article 6, plus Article 9 for health data), a processor contract under Article 28 if a provider handles your data, a lawful route for any transfer outside the EU (Articles 44 to 46) and security measures under Article 32. A vendor's “GDPR compliant” label does not replace those checks.
At ESADEX we build LLM systems to meet GDPR and the EU AI Act. We start from your data and your deployment options, then choose the model, and we record the reasoning in your DPIA.
Which deployment fits your data?
| Deployment | Who processes your data | What to check | Often a fit for |
|---|---|---|---|
| API from a provider's EU region | The provider, in the EU region | The operator's parent company and its legal exposure, transfer terms, retention of prompts and logs, and whether your inputs train the model | Lower-sensitivity text, with a signed processor contract |
| EU-based provider hosting open-weight models | An EU company under EU law | Where inference, logs and backups run, sub-processors and support access | Most business uses that involve personal data |
| Open-weight model on your own servers or EU cloud tenancy | You | Your own security, logging and updates, and the model licence | Health, legal, HR and other sensitive data |
| Consumer chat tool | The vendor, under consumer terms | Whether the terms include a processor contract and exclude training on your inputs | Not for personal data unless business terms cover it |
Our guide to EU-hosted AI platforms explains how to test a hosting claim in writing. Our guide to GDPR-compliant RAG covers the retrieval layer.
Open-weight or closed model?
Open-weight means you can download the model and run it yourself. Your prompts then stay on infrastructure you control, which removes a third-party processor from the picture. It does not make the system compliant. You become the operator, so you own security, access control, logging and updates. Read the licence before you commit, because open-weight licences differ on commercial use and other conditions.
A closed model reached through an API is often simpler to start with. The trade-off is that the provider processes your prompts, so you need the processor contract, a transfer route and written terms on retention and training.
Which EU rules apply to a language model?
| Rule | Who it mainly binds | What it means for you |
|---|---|---|
| GDPR | You as controller, and any provider as your processor | Lawful basis, contract, transfers, security, and a DPIA when the risk is high. |
| AI Act, general-purpose AI model rules (Articles 51 to 55) | The company that places the model on the market | Technical documentation, a copyright policy and a training data summary. These have applied since 2 August 2025, and the Commission can fine from 2 August 2026. Models already on the market before 2 August 2025 have until 2 August 2027. |
| AI Act, Article 50 transparency | Providers and deployers of chatbots and generated content | Tell people they are dealing with AI and label certain generated content. See our Article 50 guide for the details. |
| AI Act, high-risk rules | Providers and deployers of high-risk systems | These depend on your use case, not on the model. Annex III obligations apply from 2 December 2027 under Regulation (EU) 2026/1744. |
Using a model through an API does not normally make you its provider. According to the Commission, fine-tuning does so only in exceptional cases, when your changes use more than one third of the original model's training compute. Open-source model providers get relief from some documentation duties, but not from the copyright policy or the training data summary, and not if the model carries systemic risk. Building a product on a model can still make you the provider of an AI system, which is a separate role.
What did the EDPB say about training data?
In Opinion 28/2024, adopted on 17 December 2024, the European Data Protection Board said that supervisory authorities decide case by case whether a model is anonymous. Both the likelihood of extracting personal data directly from the model and the likelihood of obtaining it through queries must be insignificant, taking into account all the means reasonably likely to be used. If personal data stays in a model and you deploy it as a separate controller, the Board expects you to assess whether the model was developed through unlawful processing, looking at the source of the data and any infringement finding by an authority or court. Ask your provider for its answers and keep them in your DPIA.
Six questions to ask before you choose
- Where do inference, logs, backups and support access sit?
- Which company contracts with you, and who owns it?
- Does the provider keep your prompts or train on them, and will it confirm in writing that it does not?
- Which sub-processors handle your data, and how will you hear about changes?
- Can you delete prompts, outputs and logs when someone exercises the right to erasure under Article 17?
- Does the licence allow your use, and can the provider show its AI Act documentation for the model?
Tell us your use case, the type of data involved and where the model must run. We reply within 4 business hours and can prepare a scope and architecture document with a fixed estimate in about five days.
Discuss my LLM projectFAQ
Can I use ChatGPT or another US model with personal data under GDPR?
Sometimes. You need a business contract with a processor agreement, a lawful transfer route such as the EU-US Data Privacy Framework (if the provider is certified and the decision remains in force) or standard contractual clauses, and settings that stop training on your inputs. Check each point in writing and assess the sensitivity of the data first.
Is an open-source LLM automatically GDPR compliant?
No. Running it on your own servers keeps data under your control, but you still need a lawful basis, security measures, retention rules and, where the risk is high, a DPIA.
Does it matter if the LLM is hosted in Sweden, Germany or another EU country?
For GDPR, the operator matters more than the member state, because the GDPR applies across the EU and the EEA. Choose a country for latency, customer expectations or sector rules, then check the operator, its sub-processors and its support access.
Which AI Act rules apply to a company that only uses an LLM?
Your duties depend on your use case. A chatbot needs Article 50 transparency. A system used for hiring or credit scoring can be high-risk. The model-level duties in Articles 53 and 55 sit with the model's provider.
Sources
- Regulation (EU) 2016/679 (GDPR), Articles 6, 9, 17, 28, 32, 35 and 44 to 46, Official Journal of the EU
- Regulation (EU) 2024/1689 (AI Act), Articles 25, 50 and 51 to 55, Official Journal of the EU
- European Commission, Guidelines for providers of general-purpose AI models, FAQ
- European Data Protection Board, Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models, 17 December 2024
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal of the EU
- Commission Implementing Decision (EU) 2023/1795 on the EU-US Data Privacy Framework, Official Journal of the EU
About the author
Artem Synytskyi · Co-founder · AI
Artem runs AI at ESADEX, building LLM, RAG and speech-to-text systems for regulated environments.
