Meta’s EU AI training plan remains stalled as GDPR and AI Act tighten data rules
Daniel Simon ·
Meta’s push to train its generative AI systems on Europeans’ Facebook and Instagram content remains on hold after Ireland’s Data Protection Commission asked the company to pause the plan in June 2024. The pause followed coordinated complaints from privacy group NOYB that argued Meta’s proposed “legitimate interests” legal basis for model training violated the GDPR’s consent standard.
Meta disclosed the delay in a June 2024 company statement and said it would engage with the Irish regulator, its lead EU supervisory authority. The DPC confirmed it had requested Meta to postpone training on EU and EEA user data while it assessed the company’s legal basis and user redress mechanisms.
What the filings say
In formal complaints filed on June 6, 2024 across multiple EU countries, NOYB argued Meta could not rely on legitimate interests to repurpose years of user content for AI model development without granular, opt‑in consent. The filings cited the European Data Protection Board’s 2023 binding decisions that forced Meta to move behavioral advertising to a consent basis, contending the same logic applies to AI training that changes the purpose of processing.
The EU AI Act adds a second layer. General purpose AI providers must publish a sufficiently detailed summary of the training data sources and must respect EU text and data mining opt outs under copyright law. That will require model documentation that goes beyond high level “publicly available data” descriptions common in current model cards.
Meta has also told investors in securities filings that regulatory changes in the EU could restrict the way it develops and deploys AI features. The company reported rising capital expenditures tied to AI infrastructure, which heightens the cost of delays if European data cannot be used or if models must be retrained to exclude it.
The stakes for users and models
For European users, the GDPR question is whether Meta can switch the purpose of processing from social networking to model training without explicit consent and with meaningful opt out that does not degrade core service. If regulators require opt in, uptake rates would likely be low based on prior consent regimes, which would limit access to European languages and cultural context in future models.
For Meta’s AI roadmap, the constraint is material. The company has released Llama models under permissive licenses that compete with closed systems. Without European data, Meta may face coverage gaps in smaller EU languages and local domains that degrade output quality and safety evaluations. Competitors that secure licensed corpora or obtain consent at source could gain an advantage in Europe.
Next steps turn on two timelines. Under the GDPR, the Irish DPC can coordinate with other EU authorities and escalate to an EDPB binding decision. Noncompliance can trigger fines tied to global turnover and orders to stop processing. Under the AI Act, general purpose AI obligations phase in over the coming compliance period, with penalties for missing transparency and copyright requirements also measured as a share of global turnover.
Meta has said it will continue to engage with regulators on a path to train on European data. The company will need to align GDPR purpose limitation with AI Act transparency at the same time that privacy groups seek a firm consent standard. The outcome will shape how quickly Llama and its successors can incorporate European content without breaching EU law.