EU AI Act Article 5 (Prohibited AI Practices)

Article 5 of Regulation (EU) 2024/1689, the EU Artificial Intelligence Act, is the Act’s list of AI practices the EU treats as unacceptable-risk and bans outright — not regulated, not requiring a conformity assessment, simply prohibited. This sits apart from the Act’s much larger high-risk framework (Annex III), which imposes obligations — risk management, data governance, human oversight, conformity assessment — on AI systems that remain lawful to deploy. Article 5 practices are not eligible for that route: if a system’s function matches one of the eight prohibited categories, deploying it in the EU (or affecting people in the EU) is unlawful regardless of sector, deployer type, or stated purpose.

The eight prohibited categories

• Subliminal or manipulative techniques that materially distort a person’s behaviour by impairing their ability to make an informed decision, causing significant harm.
• Exploitation of vulnerabilities due to age, disability, or a specific social or economic situation, again where it materially distorts behaviour and causes significant harm.
• Social scoring — evaluating or classifying people over time based on social behaviour or inferred personal characteristics, where the resulting score is then used to their detriment in a context unrelated to where the data was generated.
• Predictive policing based solely on profiling — assessing someone’s risk of committing a criminal offence based only on profiling or personality-trait assessment, without objective, verifiable facts directly linked to criminal activity.
• Untargeted scraping of facial images from the internet or CCTV footage to build or expand facial-recognition databases.
• Emotion inference in workplace or education settings, except where used for narrow medical or safety purposes.
• Biometric categorisation that infers race, political opinions, trade union membership, religious or philosophical beliefs, sex life, or sexual orientation from biometric data.
• Real-time remote biometric identification (RBI) in publicly accessible spaces for law enforcement — banned by default, with three narrow exceptions (targeted search for victims of trafficking/missing persons, preventing an imminent and specific terrorist threat, or locating a suspect in a specific serious crime carrying at least a 4-year custodial sentence), each requiring prior judicial or independent administrative authorisation and registration in an EU database.

These provisions applied from 2 February 2025 — the earliest compliance date in the whole Act, six months ahead of most other obligations — and the European Commission issued draft guidelines on 4 February 2025 to clarify their scope and give worked examples ahead of the Article 99 penalty regime taking effect. Violations of Article 5 sit in the Act’s highest penalty tier: administrative fines of up to €35 million or 7% of total worldwide annual turnover, whichever is higher — more than double the tier applied to most other AI Act non-compliance.

Why the Act’s research exemptions do not reach Article 5

Research institutions are used to the AI Act carving out space for pure research: Article 2(6) excludes AI systems or models developed and put into service solely for scientific research and development from the Regulation entirely, and Article 2(8) excludes research, testing, and development activity prior to an AI system being placed on the market or put into service. Neither exemption reaches Article 5. Article 2(8)’s own text limits the research/testing exclusion to activity that does not amount to real-world testing — the moment a prohibited-practice-shaped system is tested on real people (student cohorts, research participants, job applicants, the public), that exclusion no longer applies, and Article 5 governs the activity directly. In practice this means the purpose-based exemptions that meaningfully soften the Act’s high-risk obligations for university and institute research programmes (see CASRAI’s guide to EU AI Act obligations and exemptions for research organisations) do essentially nothing for a project that falls within Article 5 — a research purpose does not make an otherwise-prohibited practice lawful.

Worked examples for research institutions

Article 5 issues surface most often in research contexts that involve behavioural inference, biometric data, or scoring of people, even when framed as academic rather than commercial work:

• An emotion-recognition tool built to flag student attention or disengagement during online exams or lectures falls within the Article 5(1)(f) education-emotion-inference ban, regardless of research intent, unless it genuinely fits the narrow medical/safety exception.
• Scraping publicly available photographs or CCTV stills without consent to build a facial-recognition training or benchmarking dataset falls under the Article 5(1)(e) prohibition on untargeted facial-image scraping, whether or not the resulting model is ever commercialised.

By contrast, an AI-assisted evaluation tool that scores candidates or applicants against objective, verifiable, job- or study-related criteria is not itself an Article 5 social-scoring or profiling-based practice — it may still land in the Act’s high-risk category under Annex III (for example, employment- or education-access-related AI systems), which brings its own risk-management, human-oversight, and conformity-assessment obligations, but that is a materially different (and less absolute) compliance path than an outright Article 5 ban.

Related CASRAI content

• EU AI Act: Obligations and Exemptions for Research Organizations
• EU AI Act Article 10 (Data and Data Governance)

This is a fast-moving area of EU regulation. The Commission’s Article 5 guidelines were issued as non-binding draft guidance in February 2025; institutions with an Article 5-adjacent research activity should confirm current status directly against the AI Act Service Desk rather than relying on a static summary.