Builds on AR-000200 (The EU AI Act Explained) and AR-000210 (High-Risk AI Systems) for shared vocabulary and the four-tier risk structure — not re-derived here.
Executive summary
What this means
Who this applies to
Regulatory context
Requirements
What organizations must do
- Screen every AI system and use case against the nine Article 5 categories before any risk-tier classification exercise — a system caught by Article 5 cannot be used at all.
- Assess actual and foreseeable effects on users, not only stated purpose — several categories key on whether a practice "causes or is reasonably likely to cause" harm, not on intent.
- Treat the Article 5(1)(h) exceptions as a compliance-heavy pathway, not a simple carve-out, if considering real-time remote biometric identification.
- Recognize Article 5 has applied since 2 February 2025 — this is an existing compliance gap for unscreened portfolios, not a future deadline.
Common mistakes
Treating Article 5 as merely the "extreme end" of the high-risk tier is a common conceptual error — it is a structurally different mechanism (an outright ban with narrow, tightly conditioned exceptions in one category only) rather than a stricter version of the Chapter III Section 2 requirements.
Assuming the Article 5(1)(d) predictive-policing exception covers any human-in-the-loop system is a second mistake — it only applies where the underlying assessment is already based on objective, verifiable facts directly linked to criminal activity.
Assuming the Article 5(1)(h) law-enforcement exceptions also cover private-sector or non-law-enforcement public-sector use is a third mistake — the prohibition and its exceptions are scoped specifically to law enforcement purposes.
Open questions
- Whether the Digital Omnibus changed the wording of the original eight Article 5 categories, the Article 99 fine figures, or other parts of Article 113. The AI Act Service Desk's pages for Articles 5, 99, and 113 all carry the same Commission disclaimer seen on other provisions: each "has been amended by the Digital Omnibus on AI" and the displayed text has not yet been updated. The ninth category and the 2 February 2025 date are independently corroborated by SRC-0011, so treated as settled — but whether anything else changed is unconfirmed.
- Whether the Commission has updated its prohibited-practices guidelines to cover the new nudification/CSAM category — not confirmed on this research pass.
- How the "reasonably likely to cause significant harm" standard will be applied to everyday AI-driven personalisation and recommendation systems — not yet settled by Commission guidance.