The Council's position puts general purpose models inside the Artificial Intelligence Act

2022-12-06EU Policy

Council of the European Union, general approach on the Proposal for a Regulation laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), Council document 14954/22, adopted 6 December 2022; press release "Artificial Intelligence Act: Council calls for promoting safe AI that respects fundamental rights" (6 December 2022)

The Council of the European Union adopted its general approach on the proposed Artificial Intelligence Act, the common position from which it will negotiate with the European Parliament once Parliament adopts its own. Against the Commission's April 2021 proposal, the Council narrows the definition of an artificial intelligence system to systems developed through machine learning approaches and logic- and knowledge-based approaches, to separate them from simpler software. It extends the ban on social scoring to private actors, adds a horizontal layer to the high-risk classification so that systems unlikely to cause serious fundamental rights violations are not captured, and adjusts the requirements for high-risk systems on data quality and technical documentation to be more technically feasible. For general purpose artificial intelligence systems, the Council's text provides that certain requirements for high-risk systems would apply in relevant cases, with an implementing act to specify how they are to be applied after consultation and impact assessment. National security, defence and military purposes are excluded, as are systems used solely for research and development and non-professional use. Regulatory sandboxes may permit testing in real-world conditions, fines for small and medium-sized enterprises and start-ups are capped more proportionately, and users of emotion recognition systems must inform the people exposed to them. Ivan Bartoš, the Czech deputy prime minister for digitalisation, said the text balances innovation against fundamental rights.

What it changesA provider of a general purpose model now has to plan for a regime in which its technical documentation, training data description and risk management records are regulatory filings rather than internal notes. Those documents and a patent specification draw on the same engineering facts, and a trade secret program draws its boundaries around the same facts, so they should be written together from the start: decide which parts of the training pipeline, data curation and evaluation method will be claimed in a patent application, which will be disclosed only to the extent the regulation requires, and which must stay secret, and write each document to that decision rather than letting the compliance team describe in an annex what the patent team planned to keep. The definition matters as well: a system built on machine learning falls inside the Act on the Council's text, and the obligations attach to the provider of the model whether or not it holds patents on it. Nothing is final until the trilogue ends.