The European Union is equipping itself with legal arsenal to regulate the development of artificial intelligence, and in particular to contain or even ban certain systems deemed risky or dangerous.
It was well worth a tweet from Thierry Breton. On the social network X, the European commissioner responsible for digital saluted this Tuesday, May 21, the adoption of the AI Act. “ The twenty-seven member states have just unanimously adopted the AI Act, which is about to become official law » on the Old Continent.
The AI Act, as its name suggests, is a European regulation on artificial intelligence (AI). It is the result of several months of negotiations to establish a legal framework for this technology, which has been booming since the end of 2022. It was on this date that the American company OpenAI released ChatGPT, a chatbot which immediately caused a sensation .
The AI Act will take full effect in 2026
This unanimous approval allows the text to overcome the final obstacle that remained on its path and for which there could have remained a final uncertainty. Of course, there are still a few steps left on the path to the AI Act, but procedural ones. In particular, it must be published in the Official Journal; twenty days later, it will come into force.
Then, it will be necessary to wait two years for the text to apply in its entirety – for very specific provisions, exceptions are to be expected. This is the case for uses of AI falling into the category of unacceptable risks. In this case, the bans must apply from six months after entry into force.
This is the same pace that we observed for the General Data Protection Regulation (GDPR): the text came into force on May 24, 2016, but it is from May 25, 2018 that it is become fully applicable. At the latest, according to the AI Act, certain measures will be required by 2030.
The logic of the AI Act is to propose regulation according to the degree of dangerousness of the systems. Four risk thresholds have been designed, with more or less severe rules depending on the levels (minimal/zero, limited, high and unacceptable). This last level will be banned in the EU – like social scoring or predictive policing.
“ AI systems presenting only limited risk would be subject to very light transparency obligations, while high-risk AI systems would be permitted, but subject to a series of requirements and obligations to access the European Union market », summarizes the European Council in a press release.
Same logic for general purpose AI models. Those that do not pose systemic risks will be subject to a relatively lenient framework, albeit with certain expectations in terms of transparency. On the other hand, those with a stronger impact will have to meet stricter requirements. Financial penalties are provided for in the event of an infraction.
Subscribe to Numerama on Google News so you don’t miss any news!