If your company uses AI with customers, employees or visitors, Article 5 is one of the parts of the Regulation where non-compliance would be current, not future. Its eight original points have applied since 2 February 2025, admit no exception for size or sector, and have no calendar left to wait for; the two inserted by Regulation (EU) 2026/1744 will apply from 2 December 2026.
That changes how it reads. Everything else covered here is preparation; this is a list of things you don't do, today.
The original points are eight, from a) to h). The two the Digital Omnibus added are different and have their own article; this one covers the original eight, with a test that's usually missing: which ones can touch an ordinary private company.
The figure sets out the ten points of Article 5(1) of Regulation (EU) 2024/1689 — the eight original points, (a) to (h), and points (ba) and (bb) inserted by Regulation (EU) 2026/1744 — each with its scope and exceptions: subliminal or manipulative techniques materially distorting behaviour and causing significant harm; exploiting vulnerabilities due to age, disability or a specific social or economic situation; intimate content without consent; child sexual abuse material; social scoring with detrimental treatment in unrelated contexts or unjustified treatment; predicting crime from profiling alone, except supporting a human assessment already based on objective facts; facial recognition databases from untargeted scraping; inferring emotions at work or in education, except for medical or safety reasons; biometric categorisation of sensitive traits; and real-time remote biometric identification for law enforcement, except for the objectives the point lists.
The ones that don't touch anyone who isn't looking for it
Four are ruled out just by reading them.
d) Predicting crimes based solely on profiling or on assessing personality traits and characteristics — with an express exception for systems that support a human assessment “which is already based on objective and verifiable facts directly linked to a criminal activity”. e) Facial recognition databases built by untargeted scraping from the internet or CCTV. h) Real-time remote biometric identification in publicly accessible spaces for the purposes of law enforcement, unless and in so far as it is strictly necessary for one of the three objectives the point itself lists. And f), emotion inference in the workplace and in education institutions — with an exception for medical or safety reasons — already covered here for healthcare, education and staffing, because it's the one that crosses operations the most.
And one usually assumed dangerous that isn't
This one is worth pausing on, because intuition gets it wrong.
Point g) is often cited in connection with biometric access control. Read exactly what it prohibits:
“biometric categorisation systems that categorise individually natural persons based on their biometric data to deduce or infer their race, political opinions, trade union membership, religious or philosophical beliefs, sex life or sexual orientation”.
The point adds its own exception: it does not cover any labelling or filtering of lawfully acquired biometric datasets, or the categorising of biometric data in the area of law enforcement.
What's prohibited is inferring sensitive attributes from biometrics. A turnstile that compares a fingerprint or a face against a register to open a door infers nothing on that list: it checks who you are. That's a different operation, and point g) doesn't reach it.
That doesn't leave it unregulated — it processes biometric data, with everything the GDPR requires, and if instead of verifying it identified people remotely it could fall under Annex III, point 1(a), which expressly excludes verification — but it isn't a prohibited practice, and treating it as one confuses a serious obligation with an absolute infringement.
The three that can touch you, and where the threshold sits
a) Subliminal or purposefully manipulative techniques
The territory is advertising and AI-driven interface design. But the literal wording sets cumulative conditions: subliminal techniques or “purposefully manipulative or deceptive techniques”, that — by objective or by effect — “materially” distort behaviour, “appreciably” impairing the ability to make an informed decision and causing a decision that would not otherwise have been taken, and that cause — or are reasonably likely to cause — “significant harm” to that person, another person or a group.
Ordinary advertising persuasion doesn't reach that bar, and saying otherwise would be alarmist. What does come close is deliberately deceptive design, optimised by AI, to get someone to sign up for something they didn't want, when the result causes — or is reasonably likely to cause — significant harm, to that person or to others.
The word that decides it is “significant”. Without harm of that magnitude, actual or reasonably likely, there's no point a).
b) Exploiting vulnerabilities
The most reachable of the three, and the least examined. It prohibits exploiting vulnerabilities “due to their age, disability or a specific social or economic situation”, with the objective or the effect of materially distorting behaviour, in a way that causes — or is reasonably likely to cause — significant harm to that person or another person.
Ad targeting by economic situation doesn't fall under this point on its own. The line is whether that vulnerability is exploited to induce a harmful decision: offering expensive financing precisely to whoever the model flags as financially vulnerable sits much closer to this point than to any high-risk debate.
c) Social scoring
It's associated with the public sector, and the literal wording doesn't restrict it to that sector. It prohibits evaluating or classifying people “over a certain period of time” based on their social behaviour or known, inferred or predicted personal or personality characteristics, when the resulting score leads to either or both of the following:
“(i) detrimental or unfavourable treatment of certain natural persons or groups of persons in social contexts that are unrelated to the contexts in which the data was originally generated or collected” — or “(ii) detrimental or unfavourable treatment of certain natural persons or groups of persons that is unjustified or disproportionate to their social behaviour or its gravity”.
Scenario (i) is the one that can appear without any bad intent: a score built from data in one context — payment behaviour, use of a service, internal activity — that ends up determining how that person is treated in another, unrelated context. When two systems cross and the output of one conditions something unconnected, this point is worth a look.
What to do with this, which is little and specific
You don't need a project. You need three questions about the systems you already have.
Does any of them segment or decide using age, disability or economic situation as a signal? If the answer is yes, the next question is what decision it produces and who it harms.
Is any internal score about people used outside the context in which it was generated? That's scenario (i) of point c), and it shows up when tools get integrated, not when they get bought.
Does any system categorise people on the basis of their biometric data to deduce their race, political opinions, trade union membership, religious or philosophical beliefs, sex life or sexual orientation? That, and only that, is point g).
And a final note that orders the reading: for a private company, Article 5 isn't documented or managed. Outside the exceptions each point sets out, there's no mitigating measure, no impact assessment, no room for proportionality. Whatever falls inside it stops being used. It's the only part of the Regulation where the work isn't being able to explain what you do, but not doing it.
Content in line with Articles 5 and 113 and Annex III of Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744 (Official Journal of the EU, 24 July 2026).
This article is for informational purposes only and does not constitute legal advice.