AI Act prohibitions start to apply

A law that arrives in stages

The EU AI Act regulates artificial intelligence by risk. A small set of uses is banned outright; other systems are permitted only if they meet transparency or other requirements; the rest is left largely untouched. Rather than applying all at once, the act's obligations take effect in steps, so the bans, the rules for general-purpose AI models and the bulk of the remaining provisions begin to apply at different moments ^[1].

The first of those moments fell in early 2025, and it concerned the sharpest edge of the regulation: the practices the act forbids altogether, together with a duty to make sure people who use AI systems understand them.

What started on 2 February 2025

On 2 February 2025, the prohibitions on certain AI systems started to apply, and so did the act's requirements on AI literacy ^[1]. From that date, an AI practice falling inside the prohibited list is no longer something a provider or deployer can treat as an ordinary design choice; and organisations covered by the literacy requirement are expected to ensure that the people working with AI systems have a sufficient understanding of them.

The Commission's guidelines

Two days later, on 4 February 2025, the European Commission published guidelines on the AI practices prohibited by Article 5 of the act ^[2]. The guidelines set out how the Commission reads the bans, covering practices that include subliminal or manipulative techniques, the exploitation of vulnerabilities, social scoring, predictive policing based on profiling, untargeted scraping of facial images, emotion inference in workplaces and schools, biometric categorisation, and real-time remote biometric identification for law enforcement ^[2].

The guidelines are a statement of the Commission's interpretation rather than a source of new law. They are not legally binding, and the Court of Justice of the European Union holds final authority over how the act is interpreted ^[2]. For anyone trying to work out whether a particular system sits inside or outside a prohibition, they are a reference point rather than a ruling.

The deadlines that followed

A second stage arrived on 2 August 2025, when the obligations for general-purpose AI models, together with the act's provisions on governance, confidentiality and penalties, started to apply ^[1]. The same date was the deadline by which Member States had to designate their national competent authorities — the bodies that supervise the rules inside each country ^[1].

One more stage was set for later: the remainder of the act, with the exception of Article 6(1), was scheduled to start applying on 2 August 2026 ^[1]. Article 6(1) is thus carved out of that date and follows its own timetable.

Why the date matters

The arrival of the prohibitions changed the question organisations had to ask about an AI system. Instead of only considering what a system does and how well it does it, they had to consider whether the use itself is one the act forbids — and, separately, whether the people operating it understand it well enough. The AI literacy requirement is easy to overlook next to the bans, but it applies broadly: it is about the humans in the loop, not only about the systems.

The guidelines published in the same week made the banned practices concrete by naming the categories the Commission has in mind, from manipulation and social scoring to emotion inference at work and at school ^[2]. Their non-binding character matters too: they guide, but the final word on interpretation belongs to the Court of Justice of the European Union ^[2]. Compliance therefore rests on the text of the act as ultimately interpreted, not on the guidelines alone.

Because the act unfolds in stages, the February 2025 date is best understood as the opening of a longer sequence rather than a single compliance deadline — with the general-purpose model rules and national supervisory arrangements following in August 2025, and the remaining provisions, apart from Article 6(1), set for August 2026 ^[1].