Prohibited AI Practice: Definition under EU AI Act Article 5
A prohibited AI practice is an AI use banned outright under EU AI Act Article 5 since 2 February 2025. Breach draws the top €35M or 7% penalty tier. Read more.
A prohibited AI practice is an AI use that Article 5 of Regulation (EU) 2024/1689 bans from being placed on the market, put into service, or used in the EU. It is the most restrictive of the Act's four risk tiers — the so-called unacceptable risk category.
The defining property is categorical, not conditional. Unlike a high-risk AI system, a prohibited practice has no compliance pathway. No conformity assessment, safeguard, governance measure, or risk-management process makes it lawful. The only route to lawfulness is to remove the prohibited functionality.
This entry fixes the meaning of the term. The tier mechanics — how unacceptable risk sits among the four levels — live on the unacceptable-risk tier page, and the clause-by-clause statute walkthrough on the full Article 5 prohibition framework.
Definition: what counts as a prohibited AI practice
The term covers a use or practice, not a technology class. The same underlying model may be perfectly lawful in another deployment; what Article 5 bans is the specified use. A facial-recognition model is not prohibited as software — its untargeted scraping of facial images to build a recognition database is.
That distinction matters because it puts the burden on intended purpose, the same way high-risk classification turns on use rather than architecture. A practice is either inside one of the Article 5 prohibitions or it is not; there is no proportionate-safeguards middle ground, as there is for high-risk systems under Chapter III.
What Article 5 prohibits (the compact list)
The categories at a glance
Article 5(1) sets out the prohibited practices, lettered (a) through (h). The table below states each in a single phrase. For the full per-prohibition analysis — thresholds, definitions, and the narrow carve-outs — follow the full Article 5 prohibition framework and the unacceptable-risk tier.
| Article 5(1) | Prohibited practice (one line) |
|---|---|
| (a) | Subliminal, manipulative or deceptive techniques causing significant harm |
| (b) | Exploiting vulnerabilities of age, disability or socioeconomic status |
| (c) | Social scoring by public or private actors causing unjustified detrimental treatment |
| (d) | Predicting criminal offending from profiling or personality traits alone |
| (e) | Untargeted scraping of facial images to build recognition databases |
| (f) | Emotion recognition in the workplace or education (medical/safety carve-outs) |
| (g) | Biometric categorisation inferring sensitive attributes |
| (h) | Real-time remote biometric identification in public spaces by law enforcement |
Each prohibition carries its own harm threshold or exception inside the statute — see subliminal and manipulative techniques and real-time remote biometric identification for two worked examples.
What is NOT prohibited (common confusions)
Three boundaries trip readers up.
- Private creditworthiness and insurance scoring is not prohibited. Article 5(1)(c) binds public and private actors alike, but it targets scoring based on social behaviour that produces detrimental treatment in unrelated contexts. Commercial credit scoring and life/health insurance risk-pricing are high-risk uses under Annex III, points 5(b) and 5(c) — heavily regulated, not banned.
- Private real-time biometric identification is not banned by 5(1)(h). That prohibition binds law enforcement. Private deployments may still be high-risk and face other duties, but they are not caught by 5(1)(h).
- Recidivism tools built on objective, verifiable facts are not caught by 5(1)(d). The prohibition bites where risk is assessed solely from profiling or personality traits, not where it supports a human assessment grounded in objective evidence.
The new 2 December 2026 prohibition (CSAM / 'nudifier' tools)
Track one live change. A further prohibition lands 2 December 2026, covering AI used to generate child sexual abuse material and so-called 'nudifier' tools, alongside the Article 50 content-marking and transparency duties that take effect the same date.
The practical point: the set of prohibited practices is not frozen at the February 2025 list. The original Article 5(1) prohibitions have applied since 2 February 2025, while this added prohibition applies from 2 December 2026. Treat it as a note to monitor, not a settled boundary — if your roadmap touches synthetic-image generation, the December 2026 line is the one to watch.
In force since 2 February 2025 — and the penalty for breach
The Article 5 prohibitions have applied across every Member State since 2 February 2025. There is no transition period and no grandfathering: a system built before that date enjoys no legacy exemption.
Breach sits at the top of the penalty ladder. Under Article 99(3), a prohibited practice draws fines up to €35 million or 7% of total worldwide annual turnover, whichever is higher. The lower tiers, for context:
| Penalty tier | Trigger | Maximum |
|---|---|---|
| Article 99(3) | Prohibited practice (Article 5) | €35 million or 7% of worldwide turnover |
| Article 99(4) | Most high-risk / obligation breaches | €15 million or 3% of worldwide turnover |
| Article 99(5) | Incorrect or misleading information to authorities | €7.5 million or 1% of worldwide turnover |
Under Article 99(6), for SMEs and start-ups the fine is capped at the lower of the fixed amount or the percentage of turnover. One timing nuance: the prohibitions themselves bind from 2 February 2025, while the Article 99 penalty regime became applicable on 2 August 2025.
How Confir helps
Confir screens your AI uses against the Article 5 prohibitions as part of classification intake. Its deterministic and rule-based engine walks each prohibition — and the public-authority, law-enforcement and harm-threshold boundaries that separate a banned practice from a high-risk one — and returns a documented, reproducible finding: the same logic every time, no model inference, no hallucination. Because the output is traceable, you can show authorities why a use was cleared as high-risk rather than mislabelled prohibited, and re-run the check whenever an intended purpose changes.
Frequently asked questions
What is a prohibited AI practice under the EU AI Act?
A prohibited AI practice is an AI use that Article 5 of Regulation (EU) 2024/1689 bans outright from being placed on the market, put into service, or used in the EU. It is the most restrictive of the Act's four risk tiers — the unacceptable-risk category. Unlike high-risk systems, a prohibited practice has no compliance pathway: no conformity assessment, safeguard, or governance measure converts it into a lawful system. The only route to lawfulness is to remove the prohibited functionality.
What practices does Article 5 prohibit?
Article 5(1) sets out prohibitions lettered (a) through (h): manipulative, subliminal or deceptive techniques causing significant harm; exploitation of vulnerabilities of age, disability or socioeconomic status; social scoring causing unjustified detrimental treatment; predicting criminal offending from profiling or personality traits alone; untargeted facial-image scraping for recognition databases; emotion recognition in workplaces and education; biometric categorisation inferring sensitive attributes; and real-time remote biometric identification in public spaces for law enforcement. Each carries narrow exceptions or harm thresholds within the statute.
When did the prohibited AI practices take effect?
The Article 5 prohibitions have applied across every EU Member State since 2 February 2025, with no transition period or grandfathering. A further prohibition — targeting AI used to generate child sexual abuse material and 'nudifier' tools — applies from 2 December 2026, alongside Article 50 content-marking duties. So the set of prohibited practices is not frozen at the February 2025 list; readers should track the December 2026 addition as a live change.
What is the penalty for a prohibited AI practice?
Breaching Article 5 sits at the top penalty tier under Article 99(3): fines up to €35 million or 7% of total worldwide annual turnover, whichever is higher. Lower tiers apply elsewhere — €15 million or 3% for most high-risk obligation breaches (Article 99(4)), and €7.5 million or 1% for supplying incorrect or misleading information to authorities (Article 99(5)). For SMEs and start-ups, Article 99(6) caps the fine at the lower of the fixed amount or the percentage.
Is private credit scoring a prohibited AI practice?
No. Article 5(1)(c) is not limited to public authorities — it binds private operators too — but it targets evaluation based on social behaviour that causes detrimental treatment in unrelated contexts. Private creditworthiness and credit-scoring AI, and life or health insurance risk-pricing, are not prohibited — they fall in the high-risk tier under Annex III, points 5(b) and 5(c), with their own obligation stack. Mislabelling a commercial scoring tool as 'prohibited social scoring' is a common misreading of the boundary between Article 5 and Annex III.
Related terms
- the full Article 5 prohibition framework — the clause-by-clause statute walkthrough of each prohibition.
- the unacceptable-risk tier — where prohibited practices sit among the four risk levels.
- real-time remote biometric identification — the law-enforcement prohibition and its narrow exceptions.
- social scoring as a high-risk use — why private scoring is high-risk, not prohibited.
- subliminal and manipulative techniques — the Article 5(1)(a) prohibition in detail.
Manage your EU AI Act compliance in one place
Confir automates risk classification, technical documentation, and audit trails for any company. No consultants. No 6-month projects. 14-day free trial.
Start free trial →