Prohibited AI practices (Article 5): the hard floor that already applies

Article 5 is the hardest provision of the AI Act: not obligations or conditions, but an outright ban. It has applied since 2 February 2025, to providers and deployers alike, and the Digital Omnibus agreement of May 2026 changes nothing about it (see the timeline). Violations risk the regulation's highest fine category: up to €35 million or 7 percent of worldwide annual turnover (Article 99(3)).

The eight prohibited categories

1. Manipulative or deceptive techniques that materially distort a

person's behaviour and cause, or are likely to cause, significant harm, beyond their awareness.

1. Exploitation of vulnerabilities due to age, disability or social or

economic situation, with the same harm test.

1. Social scoring: evaluating people on social behaviour or personal

characteristics where it leads to detrimental treatment in an unrelated context or that is disproportionate. The ban covers public authorities and private parties alike.

1. Predicting criminal offences solely on the basis of profiling or

personality traits.

1. Untargeted scraping of facial images from the internet or CCTV to

build facial recognition databases.

1. Emotion recognition in the workplace and in education, save for

medical or safety reasons.

1. Biometric categorisation to infer sensitive characteristics (race,

political opinions, trade union membership, religion, sexual orientation).

1. **Real-time remote biometric identification in publicly accessible spaces

for law enforcement**, save for exhaustively defined exceptions subject to prior authorisation.

A ninth ban via the Omnibus (from 2 December 2026)

The Digital Omnibus agreement of May 2026 adds a new prohibited category to Article 5: AI systems that generate or manipulate non-consensual intimate imagery ("nudifiers") or child sexual abuse material (CSAM). This ban applies from 2 December 2026. It covers not only systems intended for this purpose, but any system where such output is a reasonably foreseeable consequence without adequate technical safeguards.

Where the edges lie

The Commission coloured in the open concepts on 4 February 2025 with guidelines. Three clarifications stand out. First: the manipulation ban requires a material distortion of behaviour plus (likely) significant harm — ordinary personalisation and advertising do not qualify. Second: the social scoring ban looks at the effect (unrelated or disproportionate detrimental treatment), not at the label a system carries; a "fraud score" can fall under it too. Third: for emotion recognition at work the exception is narrow — well-being monitoring that in fact measures performance falls under the ban.

What this means in practice

The prohibitions apply to all use in the Union, regardless of where the provider sits. For most organisations the practical task is a screening question in the AI register: does an existing or proposed application fall into one of the eight categories, or come dangerously close? For borderline cases — gamification that steers behaviour, scoring in customer acceptance, staff monitoring tools — the guidelines of 4 February 2025 are the first frame of reference. Since the bans took effect, this is the provision supervisors can enforce fastest: no high-risk classification needs to be established, only a prohibited practice.