“Historic! With the political deal on the AI [Artificial Intelligence] Act sealed tonight, the EU [European Union] becomes the first continent to set clear rules for the use of AI” said Thierry Breton, the European Commissioner for the Internal Market, on December 8, 2023. He welcomed the European agreement reached between the three European Union bodies (Council, Parliament and Commission), which paves the way for the formal adoption of an EU Artificial Intelligence Act in the coming weeks, and application two years after the decision.[1]
In practice, this is an important step towards the spread of trusted artificial intelligence in the Union, one of the cornerstones of Europe’s digital transition agenda. The process was launched in April 2021 with a proposal from the European Commission articulated in three parts: a text of general guidelines, a plan coordinated between the Member States and a proposal for an AI Act.
The different types of AI risk
- Unacceptable risks (with regulated exceptions): these are AI systems that threaten individuals, such as cognitive-behavioural manipulation, social scoring or biometric identification and categorisation.
- The high risks concern AI systems that negatively affect safety or fundamental rights, either because they are involved in products covered by European legislation, such as toys, aviation, cars, medical devices and lifts; or because they are involved in the following specific areas: critical infrastructure, education and vocational training, employment, essential private services, law enforcement, migration and, finally, assistance in legal interpretation and application of the law. ChatGPT-3 level generative AI systems are included in this list.
- ChatGPT-4 level generative AI systems are classified as likely to present a systemic risk.
- Other AI systems are considered to be of limited risk.
Source: “EU AI Act: first regulation on artificial intelligence”, European Parliament Topics, December 19, 2023.
This proposal was based on a risk-based approach according to use. Thus, uses considered contrary to European values, such as social rating or behavioural manipulation, would be banned. High-risk AI systems would be subject to specific requirements, notably on the quality of training data, documentation, accuracy and robustness. AI systems that are not high-risk would only be subject to limited transparency requirements. These provisions would be supplemented by standards, and their implementation would be the responsibility of each Member State. The convergence of views apparent in the various policy documents[2] may have given the impression of a consensual exercise, despite the technical difficulties and the expected opposition from the Tech lobbies.
This was without taking into account the emergence of generative AI in 2022, in particular ChatGPT, which gave rise to real opposition between the European governing bodies. It was inevitable that they would be covered by the AI Act, but this has revealed differences between the EU. From the outset, the European Parliament adopted a position of great vigilance, even if this meant slowing down the spread of these systems and penalising the EU in international competitio...
