My first thoughts on the EDPB opinion regarding AI models and personal data

Full analysis likely to come in the coming days, but some first thoughts on the EDPB’s “AI models & personal data” Opinion. In essence, the EDPB doesn’t exclude anything entirely but leaves it up to national supervisory authorities to make a call in cases they are handling – and while it doesn’t prohibit relying on legitimate interest, it sets the bar pretty high in certain respects.

Here is a slightly exaggerated take on what the EDPB seems to be saying (a more in-depth write-up will perhaps be less snarky):

On whether there is processing of personal data in an AI model:
– Public figures make a whole AI model non-anonymous, per the EDPB (the “any data subject” in �43 is significant)
– Inference attacks & more can make a whole AI model non-anonymous (see ��40-41, which overlooks the CJEU’s “lawful”/”legal” requirement from Breyer)
– The difference in perspectives between user and AI model provider is irrelevant. For instance: does your AI model lead to the creation of synthetic data (the actual effect of Generative AI), and does a query by a user relate to a person seen by the user as a natural person? Sorry, but your AI model is involved in the processing of personal data.
– Nowhere does the EDPB seem to look at whether something is actually personal data for the AI model provider. It always presumes that it is, and only looks at whether anonymisation has taken place and is sufficient. If insufficient, “the SA would be in a position to consider that the controller has failed to meet its accountability obligations under Article 5(2) GDPR”.

On legitimate interest as a legal ground:
– Necessity is a major stumbling block. I mean, read and re-read �73:
“the intended volume of personal data involved in the AI model needs to be assessed in light of less intrusive alternatives that may reasonably be available to achieve just as effectively the purpose of the legitimate interest pursued. If the pursuit of the purpose is also possible through an AI model that does not entail processing of personal data, then processing personal data should be considered as not necessary. This is particularly relevant for the development of AI models. When assessing whether the condition of necessity is met, SAs should pay particular attention to the amount of personal data processed and whether it is proportionate to pursue the legitimate interest at stake, also in light of the data minimisation principle

– On the balancing test: significant focus on “a risk for the privacy of data subjects whose data might appear in the deployment phase of the AI system” (�79), but no consideration given here to whether that information was public to start with (only later re reasonable expectations of data subjects – �94).

I appreciate the fact that the EDPB has given some thought to what they consider useful measures to mitigate the risk and to facilitate the rights of data subjects. Some aren’t necessarily realistic, but it’s something, I guess.

GDPR data protection