The latest discussions on the governance of foundation models and general purpose AI often ask “How can we enforce rules that affect cutting-edge foreign providers of such an opaque technology?”

Today, we’re publishing a blueprint for the European AI Office – a centralised institution that could be established through the European Union (EU) AI Act – responsible for overseeing and supporting the implementation and enforcement of the regulation across the EU, notably on transnational issues like general purpose AI.

This work follows our “Giving Agency to the EU AI Act” memo, which analysed different institutional models on their suitability in enforcing the AI Act at the EU level. In particular, we compared a board, which would have relatively limited authority, to an agency, which would have much more enforcement power. In this blueprint, we presume that, as reflected by the European Parliament’s June 2023 negotiating position, the AI Act would establish an AI Office (which, in terms of authority, sits somewhere between an agency and a board).

With this in mind, we sought to answer the question: What mechanisms would enable the European AI Office to function effectively, efficiently, coherently and legitimately? We identify mechanisms that would enable the AI Office to satisfy these objectives. We conducted desk research and expert interviews to analyse various design features of relevant institutions and the proposed European AI Office. Based on a review of historical examples and solicitation of expert opinion, we developed recommendations for the AI Office, spanning legal, structural, financial, functional, and behavioural mechanisms.

Our recommendations are summarised in the figure below: