Private LLM for GDPR: the articles that apply
Which GDPR articles apply when you train or run a language model on personal data, and where a model in your own environment helps.
The GDPR articles that apply
| Article | What it says (paraphrased) | LLM question |
|---|---|---|
| Art. 5 | Purpose limitation, data minimisation, storage limitation, integrity and confidentiality | Was the data collected for a purpose compatible with training? Are you using only what the task needs? |
| Art. 6 | Every processing needs a lawful basis | Which basis covers training — often legitimate interests, which needs a balancing test |
| Art. 9 | Health, biometric and other special-category data is prohibited unless an exception applies | Remove or exclude special categories unless you have one |
| Art. 22 | Limits on decisions based solely on automated processing with legal or similarly significant effects | Keep a person in the loop for decisions about individuals |
| Art. 25 | Data protection by design and by default | Minimise and pseudonymise before training |
| Art. 28 | Processors only under a contract with set terms | Every AI vendor handling personal data needs a DPA |
| Art. 35 | DPIA for processing likely to result in high risk, notably with new technologies | Most LLM projects on personal data will need one |
| Arts. 44–49 | Transfers outside the EEA need adequacy or safeguards | Where do prompts, logs and support access physically go? |
Is the trained model personal data?
The European Data Protection Board says it depends. In Opinion 28/2024 it said a model trained on personal data can't be assumed to be anonymous: it is anonymous only if personal data can't be extracted from it or obtained through queries, with very low likelihood, and that has to be assessed case by case. In practice, if you fine-tune on identifiable records, treat the weights as personal data — store them, control access and delete them accordingly — unless you can show otherwise.
See how this would work on your data.
Book a callHow a private deployment changes the work
| Question | Shared hosted API | Model in your environment |
|---|---|---|
| Processors in the chain | Model vendor plus its sub-processors | Your existing cloud provider (already under a DPA) plus IRIS8 if we have access |
| Transfers | Check vendor regions, logging and support locations | Pin to your EU region; no third-party logging |
| Model as personal data | Usually not applicable — you don't hold weights | You hold and control the weights; deletion and retraining are possible |
| DPIA evidence | Rely on vendor documentation | Your own logs, data lineage and test results |
It doesn't remove obligations — you still need a lawful basis, a DPIA where required and a way to handle access and erasure requests. It makes the facts you put in the DPIA ones you control.
What we'd do first
- 1Before any training: map which fields contain personal data, drop or pseudonymise what the task doesn't need, flag special categories, and give your DPO the data lineage for the DPIA.
- 2Then a narrow task — such as routing or summarising support tickets — trained in your EU region in an 8-week pilot.
- 3The EU AI Act adds separate duties for some uses, such as credit scoring; see our banking page.
Frequently asked questions
Is any LLM "GDPR-certified"?
GDPR compliance depends on how the controller processes data. Art. 42 certification schemes exist, but none on its own means an LLM satisfies GDPR (GDPR Art. 42).
Do we need consent to train on customer data?
Not necessarily. Art. 6 lists six bases; legitimate interests is often discussed for training but needs a balancing test (GDPR Art. 6; EDPB Opinion 28/2024).
How do we handle an erasure request if data trained the model?
Keep lineage of which records went into which model version, so you can retrain without them or show the model is anonymous.
Do we need a DPIA?
Likely, for new technology on personal data at scale (Art. 35). Your DPO decides.
Where does the data go?
Nowhere outside your chosen environment.