Synthèse
One of the worst Commission proposals was stopped. This specific proposal was also a clear example of what is wrong with the current deregulatory wave. The final deal does not fully delete the obligation to register AI systems in the EU database when providers rely on Article 6(3) to classify systems as ‘not high-risk’. This matters because Article 6(3) already gives providers discretion to decide that some systems used in high-risk areas should not be treated as high-risk. Without public regist
One of the worst Commission proposals was stopped. This specific proposal was also a clear example of what is wrong with the current deregulatory wave. The final deal does not fully delete the obligation to register AI systems in the EU database when providers rely on Article 6(3) to classify systems as ‘not high-risk’. This matters because Article 6(3) already gives providers discretion to decide that some systems used in high-risk areas should not be treated as high-risk. Without public registration, companies could have avoided key AI Act obligations with almost no public trace.
But the deal still weakens public transparency. Providers will have to upload less information to the public database. This means regulators, researchers, civil society and affected people will have less information to assess whether providers are classifying their systems correctly, consequently less ability to fight back when AI harms occur. The deal also moves the Machinery Regulation from Section A to Section B of Annex I. In practice, this pushes AI systems embedded in machinery towards sectoral machinery rules, rather than the AI Act’s horizontal high-risk framework.
This is a serious change. The AI Act was designed as a horizontal law because AI-related harms do not fit neatly into product safety boxes. Machinery law may address physical safety, but it was not designed to cover the full range of AI-specific and fundamental rights risks. AI systems embedded in machinery can affect workers’ safety, work pace, task allocation, monitoring, autonomy and ability to challenge decisions. If these systems are framed as optimisation, automation, efficiency or quality control, their rights impact may become harder to capture.