GDPR and AI training data: what you can and can't do

Training and fine-tuning models on real data is where AI meets GDPR head-on. The rules are workable, but the details matter - and getting them wrong is expensive. Here is a practical map.

Personal data doesn't stop being personal in a dataset

If your training data contains information about identifiable people, GDPR applies to that processing - including the act of training. 'It is just for the model' is not an exemption.

You need a lawful basis

Like any processing, training on personal data needs a lawful basis - usually legitimate interest or consent. Legitimate interest requires a balancing test you can document; consent must be specific and freely given. Decide and record this before you train, not after.

Data minimization and purpose limitation

Collect only what you need, and use it only for the purpose you stated. Repurposing a dataset gathered for one thing to train a model for another is a classic GDPR trip-wire.

The hard parts: erasure and transparency

GDPR gives people the right to erasure and to information about processing. Models that have memorized training data make deleting an individual's data genuinely hard. Plan for it: keep provenance, prefer fine-tuning you can re-run, and avoid training on data you may have to unwind.

The safest training data is data you can account for - where it came from, why you hold it, and how to remove it.

Keep it in-jurisdiction

Transferring personal data outside the EU adds a layer of complexity (transfer mechanisms, adequacy, the CLOUD Act question). Training and storing data on EU infrastructure you control sidesteps most of it - the data never leaves, so there is no transfer to justify.

A short checklist

• Document a lawful basis before training.
• Minimize and scope the dataset to its stated purpose.
• Track provenance so you can honor erasure and transparency.
• Keep training and storage on EU-based infrastructure.

GDPR doesn't forbid training on personal data - it demands that you can account for it. The teams that keep data in-jurisdiction and document as they go spend far less time defending it. (Overview, not legal advice.)