GDPR and AI models: legitimate interest, anonymization, and the right to erasure
We already covered what GDPR says about AI in general. The question that generates the most uncertainty is more specific: what legal basis lets you train a model on personal data, when does that model fall outside GDPR, and what does the right to erasure mean once it's trained? The EDPB answered all three in Opinion 28/2024.
The legal basis: why legitimate interest dominates the debate
For most AI model developers, consent is unworkable at the scale of a training set — asking individual consent from millions of people whose data appears on the web isn't operationally realistic. The European Data Protection Board's (EDPB) Opinion 28/2024 confirms that legitimate interest (GDPR Art. 6(1)(f)) can be a valid basis, but requires passing a three-step test: identifying a genuine legitimate interest, demonstrating the processing is necessary for that purpose, and weighing that interest against the data subject's rights and reasonable expectations — backed by mitigation measures such as data minimization, reinforced transparency, and ease of exercising rights.
When a model actually falls outside GDPR
Anonymization is the other point where most companies get it wrong. The EDPB sets a high bar: a model only falls outside GDPR's scope if it is impossible, using reasonably likely means, to identify a person from it — neither through direct inference from its parameters, nor through memorization of training-set fragments the model can reproduce verbatim. Pseudonymization (replacing names with identifiers) isn't enough, and many models marketed as "anonymized" wouldn't survive a real audit against this standard, especially large ones trained on unfiltered web data.
The right to erasure doesn't always mean retraining from scratch
Here's the nuance that brings the most relief to technical teams: when a data subject exercises their right to erasure over data used to train a model, the EDPB doesn't automatically require retraining the entire model. It recognizes valid alternative measures — when full retraining would be disproportionate in cost or time — such as output filters that prevent the model from reproducing that data, selective unlearning techniques (machine unlearning), or controlling access to identifiable data sources. The condition is that these measures be equally effective at guaranteeing the right, not a shortcut for inaction.
What to document starting now
Whether you train your own model or contract an AI vendor, three pieces of documentation hold up compliance under an inspection: the legitimate interest assessment (LIA) justifying the legal basis, a technical description of what anonymization was applied and why it's sufficient, and a documented procedure for handling erasure requests that explains which alternative measure will be applied and why it's equivalent to full retraining.
Document your AI models' data processing
Record of Processing Activities (RoPA) template and the Basic GDPR Pack, to document your legal basis and the measures applied.