AI giants can now be fined up to €35m, or 7 per cent of total worldwide annual turnover, whichever is higher, for prohibited AI practices in the European Union. This begs the question, who places these fines, and how do they decide what practices pose too big a risk?
This is where the European Artificial Intelligence Act (AI Act) comes into play. The AI Act is the first major and comprehensive legal regulation of AI anywhere in the world; with enforcement and maintenance of the Act being overseen by the AI office, ensuring that the EU has tools to detect infringements and react to new threats. The laws underpinning AI will shape its function in our businesses, studies, and daily practices; it is therefore necessary to have a level of insight into how the AI Office functions, and what it prohibits. Although the Act is only currently applicable under EU jurisdiction, it aims to set a worldwide standard for AI practices. Thankfully, the Act prioritises transparency, with all laws laid out publicly.
So, what is the AI Office, and how does it work? The AI office is a body that sits within the technology-focused department (called DG CONNECT) of the European Commission. Since its establishment in February 2024, the office now consists of over 125 staff members, who are specialists in technology, policy, law and the economy. Within the Office are a range of operating units from innovation to AI in health and life sciences. On 2 August 2026, the AI Office formally gained exclusive authority over General Purpose AI (GPAI) models. They have the power to access data from any GPAI models running in the EU and can demand mitigation or entirely withdraw any high-risk systems. They are also able to hand out fines to GPAI providers of up to three per cent of their annual worldwide (not just EU) turnover, or €15m.
To aid with the legislation, the Scientific Panel and the Advisory Forum were formed to sit alongside the AI office, with the goal of advising the Office on its governing decisions. On 1 June, the Scientific Panel was appointed to provide advice on AI, with a focus on GPAI, and consists of 60 independent experts from across the EU, with no more than three members per EU State. While the Forum is constructed to represent the industry and stakeholders, the Scientific Panel is intentionally constructed without conflicts of interest. The Panel of AI experts will be key in defining the AI Act; therefore, their political and economic independence is crucial.
The AI experts in the panel have the broad task of building the safety harness that underpins AI, evaluating what constitutes malpractice. To do this, they develop tools and benchmarks to evaluate the capabilities of AI models, including known and unforeseen risks. Their findings are used to advise or alert the AI Office, which will inform regulations or enforcements.
What is ‘risk’ in the scope of AI? The concept of risk in this context spans from usage of data to systemic risks. Let’s cover what is currently being regulated, and where there may be exceptions. Any inappropriate use of analysing or inferring the behaviour, social scoring or vulnerabilities of individuals is banned. Following on, any likely harmful deployment of subliminal or manipulative techniques with the objective of distorting behaviour is also banned. Also prohibited: compiling facial recognition databases, inferring emotions through CCTV (in educational and workplace settings), and ‘real-time’ remote biometric identification (RBI). The main exception to the prohibition of RBI is in cases of urgent police investigation. But from these definitions it is clear that the task of defining and classifying risk is widely scoped itself, researchers must be up to date and comprehensive in their understanding of good and bad practices with AI.
The AI Office, and the Act, are still very much in the early days of their development. The Scientific Panel is in its first five months of office, with a task that in difficulty should not be underestimated. The safety structures are being built, and time will tell how the AI Act will contribute to our future relationship with Artificial Intelligence.
Photo by Fabian Kleiser on Unsplash
