top of page

Op-ed: 'A law for foundation models: the EU AI Act can improve regulation for fairer competition'

Regulate 'Systemic Foundation Models' only: my OECD piece argues that the AI Act should apply a DSA-minded approach, limiting Article 28b to highly capable & systemically relevant foundation models only. It would already offer sufficient protection, while being better for competition.


Last Thursday, the OECD published my long piece on foundation models and the AI Act. Written (to large parts) already in May, my job allowed me to finalize it only now. The article has three objectives: (a) illustrating the mindset behind the new Article 28, 28a and 28b, (b) explaining the European Parliament's holistic approach along the AI value chain, and (c) elaborating how to improve it further. A short summary:

  • Generative AI is hyped and there are many misleading statements. However, the underlying foundation models (i.e. GPT-4, PaLM, LLaMA) have become indeed so powerful that a regulatory intervention seems inevitable. The EU should focus on the already widely recognized harms, while preventing even more market concentration by a few dominant tech companies.

  • The proposal of the European Commission is - at a conceptual level - not capable to deal with foundation models. As undefined and mouldable digital plasticine, they do not fit in a classical product safety approach that is based on a single intended purpose and limited risk classes.

  • The EP filled this regulatory gap by introducing a five-level-approach that addresses the challenges posed by foundation models. Besides information sharing obligations in Article 28, certain minimum requirements for developers are introduced, with risk identification, testing and evaluation, documentation as the most important ones.

  • Wanting to promote competitiveness in the digital market & boost innovation by all kind of companies, I propose to limit the scope of Article 28b to 'systemic foundation models' (SFM), meaning a small number of highly capable as well as systemically relevant models, and outline how criteria to determine SFMs could look like.


There are even some thoughts on AGI ... curious? Read the full text here.


 

Please find below the AI value chain diagram in a higher quality in case you want to download it separately:


bottom of page