> "Noyb's copycat actions

What?

> are part of an attempt by a vocal minority of activist groups to delay AI innovation in the EU [...]," Meta told us.

Handing over the data of EU citizens to be exploited by US companies is not "innovation in the EU".

> [noyb:] But Meta is even limiting this (statutory) right by saying that it only applies if people opt-out before the training has started. Meta will likely also be unable to comply with other GDPR rights (like the right to be forgotten, the right to have incorrect data rectified or to give users access to their data in an AI system). Furthermore, Meta provides AI models (such as Llama) as open-source software for anyone to download and use. This means that Meta can hardly call back or update a model once it is published.

Somewhat worried about the implication for open source models in general if this argument is fully accepted. For instance if your open-source model has trained on a news article (maybe even with a license from the publisher) and a subject in the article later exercises their right to be forgotten, supposedly even deleting and re-training the model would be insufficient?

Feels like a step backwards in terms of privacy if only centralized models were to be permitted under the GDPR, opposed to ones you can run on your own hardware.

> "Most other AI providers (like OpenAI or French Mistral) have zero access to social media data and still outcompete Meta's AI systems," noyb said,

I don't think this is true - both are using web-scraped data including many social media sites. OpenAI has a partnership with reddit, even. To my understanding Meta is likewise only using public posts (but they would have the advantage of being able to bypass their own bot-prevention/rate-limits).