European Parliament negotiators reached a consensus late last Friday on language for the long-debated AI Act — the first major artificial intelligence legislation in the European Union. The draft law emerges amidst rising concerns among EU member states about the impact that regulations will have on innovation. As a result, the tentative deal could still unravel due to infighting.
The AI Act’s journey began about two and a half years ago when the first version of the law was made public in 2021. With the rise of foundation models — versatile AI systems capable of a wide range of applications — there has been a shift in the EU’s legislative approach towards focusing on the most advanced technologies. However, this pivot has sparked intense debate, with critics arguing that it could stifle innovation.
One critic of the draft legislation is France’s Emmanuel Macron, who has expressed concerns that the AI Act could kill innovation on the continent. The language of the Act still requires approval from EU member countries and the European Parliament, a process that could take months. With opposition from heads of state like Macron, there remains considerable uncertainty surrounding the fate of the AI Act.
Under the current version of the law, AI applications will be categorized by risk, with high-risk AI systems subjected to the most stringent rules. These requirements include pre-market testing, mandated risk-mitigation systems, standards for training datasets, and human oversight.
Companies will be responsible for self-assessing whether their models meet the compliance threshold for general-purpose foundation models. Initially, these models will be classified based on their computing power requirements, but this classification may change as other metrics and benchmarks become available.
The AI Act also introduces various transparency measures, such as requiring companies to notify users when they are interacting with AI systems like chatbots or emotion recognition systems. It also mandates the labeling of deepfakes and other AI-generated content, along with other disclosure and documentation requirements. EU copyright law must also be followed.
While the Act exempts more mundane AI applications from strict rules, it bans AI uses deemed to pose unacceptable risks. These include biometric systems that identify individuals based on sensitive biological data, scraping facial images from online databases, emotion recognition systems at work or in schools, social scoring systems, AI systems that manipulate human behavior, and most forms of predictive policing.
One of the areas of contention revolves around the potential use of AI in the fight against terrorism. To address these concerns, the draft law allows countries significant flexibility to use AI in a national security context. However, questions remain about the operational and compliance burdens placed on companies by the law. Similar concerns arose with the EU’s data privacy law, the General Data Protection Regulation, which some argue has led to a competitiveness gap between the American and European tech sectors.
The draft law establishes a new European AI Office for enforcement and a stringent penalty system with fines ranging from 1.5% to 7% of a company’s global sales turnover. The AI Act is expected to come into force by 2026, allowing businesses a two-year window to come into compliance. However, given the rapidly evolving AI landscape, there are concerns that the law may be outdated by the time it takes effect.
Supporters of the AI Act are hoping it will set a global standard for AI regulation. However, critics warn that if the law stifles innovation as expected, it could further sideline the EU and any countries that follow its lead.
As the debate continues, the future of AI regulation in the European Union remains uncertain. The outcome of the ongoing negotiations and the ultimate impact of the AI Act on innovation will shape the EU’s role in the global AI landscape.
