13/12/2023
December 13, 2023
The EU has reached a historic agreement on AI regulation, balancing innovation and oversight. The new law, focusing on high-risk AI, includes strict rules and sanctions, while allowing exceptions for law enforcement.
After extensive debate spanning several days, European institutions have reached a consensus on a regulation framework for artificial intelligence (AI). This pivotal regulation aims to strike a delicate balance between stringent regulation and fostering innovation, introducing specific restrictions and potential sanctions.
The European Union (EU), spearheaded by Commissioner Thierry Breton, has achieved a "historic moment" by finalizing a comprehensive AI regulation. As the first major global power to regulate this technology, the EU has navigated complex debates between advocates of flexible regulation to promote innovation and those pushing for a strict regulatory framework.
Despite the challenges, the EU's AI Act is notably pro-business. However, Minister Jean-Noël Barrot emphasizes the need for continued discussions to refine certain aspects. The regulation represents a significant step in a four-year process, underscoring the EU's commitment to evolving and improving its approach to AI governance.
The regulation primarily focuses on AI systems categorized as "high risk," which could impact health, safety, fundamental rights, and democracy. Such systems will require a thorough impact analysis before market launch, alongside mandatory transparency and explainability of AI models.
The regulation also addresses the rapidly evolving landscape of generative AI, exemplified by OpenAI's ChatGPT. It reflects a compromise between the need for self-regulation, as advocated by AI leaders like Mistral AI and Alep Alpha, and the European Parliament's call for stricter controls on predominantly American AI giants. The final agreement tailors obligations based on the "systemic risks" posed by generative AI models.
A contentious aspect of the regulation is its list of prohibited AI systems, including those using sensitive biometric categorization (e.g., political opinions, religious beliefs, race) and non-targeted extraction of facial images for facial recognition databases. Exceptions to these prohibitions are granted in specific circumstances, particularly for law enforcement, under strict judicial oversight.
Member states are now tasked with aligning their national laws with this regulation, set to be implemented in 2026. A dedicated control structure attached to the European Commission will oversee compliance, with the authority to impose fines up to 35 million euros or 7% of global turnover in case of violations.