Guidelines · 2019
Ethics Guidelines for Trustworthy AI
Guidelines from the European Commission's High-Level Expert Group on Artificial Intelligence
Guidelines defining seven requirements for trustworthy artificial intelligence, the conceptual basis for the preparatory work on the European AI regulation.
I was one of the group’s fifty-two members, contributing to the economic and security-related strands, sitting on a voluntary basis alongside my duties at ANSSI (the French national cybersecurity agency). This text is not mine to claim; it is the product of a negotiation between industry, researchers and civil-society representatives, and it bears the marks of that negotiation.
The seven requirements set out in April 2019 became the conceptual foundation for the obligations the European Union’s AI Act places on high-risk systems. On that count, the document achieved what a non-binding text can achieve: it supplied the vocabulary and the analytical structure for a regulation that came later. Few expert reports can claim as much.
Two criticisms deserve to be addressed rather than sidestepped. The first came from within the group itself: several members publicly accused the process of sidelining red lines (the uses that should have been prohibited outright) in favour of an operational consensus, and called the exercise ethics-washing. The charge holds: the final version favours workable principles over prohibitions, which was the price of unanimity.
The second is a matter of timing. The April 2019 text says nothing at all about what now dominates the debate: general-purpose models, compute scale, emergent capabilities. This is not a failure of foresight so much as a mismatch of calendars, but the European regulation had to be reopened in 2023 to bring these objects within its scope, which is a fair measure of how far an ethical framework tied to a given state of the art can travel in time.
The twin document on investment, published two months later, has aged worse still.