The Eight Prohibited AI Practices Under Article 5: A Compliance Checklist
Eight practices have been illegal in the EU since 2 February 2025 — with no compliance route, no register and no grace period. Each ban, the commercial uses that accidentally trip it, and a one-page screen for your estate.
Article 5 is the only part of the AI Act where the answer is never “document it.” Eight practices are prohibited outright, applicable since 2 February 2025 — eighteen months before the transparency and penalty regime arrived. No conformity assessment, no registration pathway, no sandbox: the remedy for a hit is removal.
The bans read as abstract legal prose and land as very ordinary product features. This is the checklist, each with the use that actually trips it.
The eight bans, translated
1. Subliminal, purposefully manipulative or deceptive techniques — Art. 5(1)(a)
Systems that materially distort behaviour and cause significant harm. Not “persuasive UX” — the bar is distortion of behaviour with significant harm downstream: addiction loops engineered against vulnerabilities, dark-pattern pricing that pushes people into unaffordable contracts.
Triage: growth features that exploit cognitive bias at scale for harmful outcomes. Recommendation engines that optimise engagement without a harm pathway stay outside; systems steering vulnerable users toward self-harm-adjacent or financially ruinous outcomes are inside.
2. Exploiting vulnerabilities — Art. 5(1)(b)
Age, disability, or social or economic situation. Marketing AI aimed at children with manipulative loops; chatbots engineered to keep financially distressed users rolling over credit.
Triage: who is targeted because of a vulnerability, and does the system exploit it?
3. Social scoring — Art. 5(1)(c)
Evaluating or classifying people leading to detrimental treatment in unrelated contexts, or treatment that is unjustified or disproportionate to the behaviour. The canonical case is China’s social-credit style scoring; the mid-market case is an HR tool that scores “cultural fit” off private-life data and feeds it into decisions about work.
Triage: does one data context (social, private, behavioural) get used to gate an unrelated context (credit, employment, services)? Outside the workplace: insurer pricing from social-media scraping.
4. Criminal offence prediction — Art. 5(1)(d)
Predicting criminal offences solely on profiling or personality traits. Two variants: offence risk from profiling, and risk of a person committing an offence from personality traits. Insurance-adjacent “fraudster likelihood” models scored purely on personal traits drift toward this line; anomaly detection on transactions (behaviour vs. baseline, not personality) does not.
Triage: the word “solely” — is any behavioural evidence involved, or only who the person is?
5. Untargeted facial-image scraping — Art. 5(1)(e)
Scraping the internet or CCTV footage to build facial-recognition databases. This is a sourcing ban: it poisons the dataset pipeline. Any vendor whose face model trained on unconsciously harvested images is built on a prohibited act — a supply-chain question you can ask directly.
Triage: how was the face data acquired? Licensed or consented datasets survive; “we scraped public photos” does not.
6. Emotion inference in workplace and education — Art. 5(1)(f)
Inferring emotions from biometric data in the workplace or educational institutions — narrow exception only for medical or safety reasons (e.g., fatigue detection in truck drivers, a genuine safety case). This is the ban most live today: AI meeting assistants that score “engagement,” sentiment analysis on support-agent voices, proctoring tools that read exam-room affect.
Triage: are you inferring emotion from face/voice/biometrics at work or school? If yes, stop. Documented safety or medical purpose is the only way through.
7. Biometric categorisation inferring sensitive traits — Art. 5(1)(g)
Categorising people by race, political opinions, trade-union membership, religious or philosophical beliefs, sex life or sexual orientation. Not labelling your own lawfully acquired datasets for filtering — the ban is inference of these categories. Smart-camera analytics that estimates “political rally attendee demographics” is the shape of the problem.
8. Real-time remote biometric identification in public spaces — Art. 5(1)(h)
For law enforcement purposes, in publicly accessible spaces, in real time — with narrow, authorisation-conditioned exceptions: targeted search for victims, prevention of a specific terrorist threat, localisation of a suspect of a serious crime. For non-law-enforcement companies this mostly matters as a procurement rule: your building-access vendor is not law enforcement, so real-time public-space facial identification is simply out.
The screen to run this week
One pass over the estate, one row per system:
- Emotion? biometric emotion inference at work/school → Art. 5(1)(f): remove.
- Faces? how were face datasets sourced; real-time public-space identification → 5(1)(e)/(h): remove or re-source.
- Scoring? one context gating an unrelated one, personality-only risk predictions → 5(1)(c)/(d): redesign.
- Targeting? vulnerable groups or manipulative distortion with harm → 5(1)(a)/(b): remove.
- Categorising? inferring sensitive attributes from biometrics → 5(1)(g): remove.
Anything that survives gets a dated line in the register: screened against Article 5, no practice fired, reviewed by whom and when. Anything that hits is a withdrawal ticket, not a remediation task — there is nothing to remediate.
Note the adjacent trap: bans with no grace period sit alongside transparency duties that have passed (Article 50, live since 2 August 2026). Companies routinely roadmap Art. 50 work while running an Art. 5 feature in production — the priority is exactly backwards.
How to check your own position in four minutes
The free exposure scan asks twelve questions about your AI usage and governance and returns an exposure score, your ranked exposures with the exact article that fired, and your first three moves — mapped to the dates that apply. Confirm your email and the full report unlocks.
Frequently asked questions
What AI practices are banned under the EU AI Act?
Eight, in Article 5: manipulative techniques causing significant harm; exploitation of age, disability or social/economic vulnerability; social scoring with unrelated or disproportionate treatment; predicting criminal offences solely by profiling; untargeted scraping of facial images; emotion inference in workplace or education; biometric categorisation inferring sensitive attributes; and real-time remote biometric identification in public spaces for law enforcement (with narrow exceptions). They have applied since 2 February 2025.
Is sentiment analysis on support calls prohibited?
If it infers emotion from voice or other biometric data and it runs in the workplace — yes, Article 5(1)(f), with only a medical or safety carve-out. Keyword-level analysis of call content for quality (what was said, not how the caller felt) is not emotion inference from biometrics and lands outside the ban, though GDPR still applies to the recording.
Can we comply with a prohibited practice instead of removing it?
No. Article 5 has no conformity route, no registration option and no transitional period — the system must be withdrawn or the practice eliminated. Record the removal decision in the register; that record, not a mitigation plan, is the compliant state.
Do the Article 5 bans have a grace period like the transparency rules?
The opposite: the bans applied from 2 February 2025, before general applicability and the Article 50 transparency duties that arrived on 2 August 2026. There is no grandfathering for prohibited practices — legacy features are not exempt.
How do we check our AI vendors for prohibited practices?
Ask three questions in the questionnaire: was face data scraped or licensed; does the model infer emotion from biometrics in your usage context; and does any score gate a decision in an unrelated domain of the person’s life. Vendor answers that are vague on data sourcing are themselves a finding — Art. 5 exposure in a purchased system is your problem to detect before deployment, since removal, not remediation, is the only fix.
Put this guide to work on your own estate.
Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.
- Free forever plan
- No credit card
- EU-hosted · GDPR-native
Build a live AI register in minutes — upgrade only when you need the audit pack.
Keep reading
AI in healthcare under the EU AI Act: clinical decision support and high-risk duties
11 min
Risk & ClassificationThe Annex III High-Risk AI Systems List: All Eight Use Cases Explained with Examples
11 min
Risk & ClassificationAI Act Conformity Assessment Under Article 43: Procedure, Evidence and Timeline
11 min