Meta cancels AI training with data under GDPR for the time being

The US platform operator Meta has stopped training its AI models with data from the EU until further notice. The company is thus complying with a request from the Irish data protection authority - which initially wanted to approve the project, but now sees a need for further discussion following protests.

Meta, operator of Facebook, Instagram, Threads and Whatsapp, announced on Friday that it would not be training its major language models with data from Facebook and Instagram for the time being. "We are disappointed by the request from the Irish Data Protection Commission (DPC Ireland), as the lead regulator for the European Supervisory Authorities, to postpone the training of our LLMs with the shared content of our adult users from Instagram and Facebook," the company said in an update to a description of the project published four days ago. Meta stated that it had already informed the supervisory authorities of its plans in March.

Facebook points to Google and OpenAI

The Zuckerberg empire thought it was already well on the way to being allowed to work with the large amount of user data from the EU and the European Economic Area, the scope of the GDPR. 260.7 million EU citizens use Facebook every month, the company stated in its transparency report on the DSA in April. 264.3 would access Instagram every month, 44 million of them from Germany. Meta now wants to use this vast amount of content to feed its AI offerings: Llama, for example, as well as the Meta AI Assistant. The company argues that only with EU data is it possible to adapt to local conditions, and that Google and OpenAI have also already used EU data for this purpose.

Meta sees itself as a model student: data originating from profiles of minors is consistently sorted out, and direct messages between users are not used for training. More than two billion notifications have been sent to users since 22 May to inform them of the plans and to enable them to object to the use of their own data using an objection form. However, it also depends on the usage options whether the EU can actually use AI or whether it only wants to watch AI innovations in the rest of the world as a spectator while the rest of the world benefits from them. Meta wants Europeans to be part of these opportunities.

Meta argues that data processing for AI training does not require separate consent for this form of data processing. It is permissible as a "legitimate interest" within the meaning of the General Data Protection Regulation. Data protection activists such as the Austrian organization NOYB take a different view - and had therefore also filed complaints against Meta's plans with other data protection supervisory authorities outside Ireland. The organization founded by Max Schrems also criticized the opt-out form for being inadequate weeks ago.

Data for AI training is becoming scarce

The Irish data protection authority has welcomed Meta's announcement that it will discontinue its use for now. "The decision was made after intensive discussions between DPC and Meta," the authority announced. It will continue the discussion with the company together with other European supervisory authorities. The authority did not explain on Friday what exactly triggered the change of heart at DPC.

Whether Meta is actually concerned with the economic and social well-being of the EU is difficult to verify. An entirely different reason for the plans for large-scale data evaluation for LLM training could lie in a problem that all major AI model developers are currently facing: in just a few years, the training data for the huge models could become scarce. What was still a fairly abstract problem on the horizon in 2022 has become more acute, according to estimates by scientists published in April 2024 weiter.

(nie)