European Commission proposes the EU Artificial Intelligence Act

European law of this kind begins with a proposal. The Commission drafts a text, the European Parliament and the Council of the European Union then negotiate their own positions, and if the three institutions agree on a common version it becomes a regulation binding in every member state. What appeared on 21 April 2021 was the opening text of that process rather than a law.

The Commission had set out the approach a year earlier in a white paper. Its central move was to regulate uses rather than the technology itself. A system fell into a tier according to what it was used for and what could go wrong. At the top sat a short list of uses judged unacceptable and banned outright. Below that came high-risk uses, allowed but carrying mandatory requirements on the provider before the system reached the market. Below that came systems owing only transparency, chatbots and generators of synthetic media among them, whose users had to be told they were dealing with a machine. Everything else was left alone.

Enforcement fell largely to national authorities. The proposal added a European Artificial Intelligence Board to coordinate how the rules were applied across the union.

The text was drafted before general-purpose systems reached the public, and the category it did not yet have, a model trained broadly and adapted afterward to many tasks, was the one being named elsewhere that year. Stanford’s report on foundation models followed four months later, and the term entered the negotiations over this proposal later still.

Tiered risk, as a way of writing rules for software whose uses cannot all be listed in advance, was argued over in other jurisdictions afterward.