Drafting a law for a technology that will not hold still

industrial scale photography, clean documentary style, infrastructure photography, muted industrial palette, systematic perspective, elevated vantage point, engineering photography, operational facilities, A vast, tiered lattice of weathered titanium beams emerging from the ocean at low tide, each level tuned to different wave frequencies, filtering turbulent swells into orderly channels; dawn light glancing off wet surfaces, casting long shadows across salt-crusted foundations; mist rising where structured outflows meet chaotic surf, symbolizing regulation shaping uncontrollable technological motion. [fal-ai/z-image/turbo]
The AI Act’s design hinges on adaptability: obligations tied to function, not architecture, and a high-risk list intentionally open to future applications. Whether those mechanisms can track uses not yet conceived remains an unresolved condition of the regulation.
In Strasbourg, Dragoş Tudorache and Brando Benifei spent the negotiation as the two MEPs co-leading the EU's AI Act. Tudorache is a former Romanian Interior Minister who sits with the liberal Renew Europe group. Benifei, a socialist, was one of the youngest Italian MEPs ever elected. Their joint work turned an idea that predates the public arrival of generative AI into law. Tudorache traces the file back well before ChatGPT's launch in 2022. The European Commission had a high-level expert group on AI, its applications and potential risks before 2019. That year, President von der Leyen announced a legislative proposal. A special committee was set up in 2020, and in 2021 the lawmakers began to "unpack the concept of AI", as Tudorache puts it. The central difficulty, in his account, was speed. AI development was fast and unpredictable, and the rules had to avoid being outpaced by it. The answer was to write obligations that do not depend on any particular technique. "If you define obligations of transparency, those will stay relevant, no matter how complex the algorithms will become," Tudorache says. A second device complements the first. The list of high-risk applications, annex three, was "deliberately left open", so that new uses and their risks can be added without rewriting the law. Whether that open list can keep up with uses no one has yet described is the question the design hands forward. It is the part of the regulation that will be tested after the vote. The Act is built on a risk-based approach, in which the riskier the application, the more scrutiny it faces. Benifei points to the sharpest cases. The use of real-time biometric cameras in public spaces is limited to pursuing suspects of very significant crimes. Emotion recognition is banned in workplaces and schools. Approval by Parliament and entry into force are separate steps, and the second was still expected in May 2024. Until then, the text is a set of commitments rather than a working regime. The lawmakers present the restrictions as the condition for the technology's use, not a brake on it. "The purpose of this legislation is not to stifle innovation, but rather to build trust," Benifei says. "Our model, where consumers are heavily protected and human rights are central, will not be hindered by this disruptive technology. Instead, it will be integrated to develop our AI model."

† Analysis: the writer's conclusion from the reported facts, not a fact stated in the source.

  • The open-ended nature of the high-risk applications list (Annex III) is a design choice intended to ensure the regulation remains future-proof against technological developments that are currently unforeseen.
  • The legislative approach prioritizes building societal trust as a foundational requirement for successfully integrating AI into the European economic and social model.
Published November 14, 2026