In this article14 sections
- The hardest line in the EU AI Act
- (a) Subliminal manipulation
- (b) Exploitation of vulnerabilities
- (c) Social scoring
- (d) Criminal risk prediction based on profiling
- (e) Untargeted facial-image scraping
- (f) Emotion recognition in workplaces and educational institutions
- (g) Biometric categorisation inferring sensitive characteristics
- (h) Real-time remote biometric identification in public spaces
- Two new bans from 2 December 2026
- What counts as "using" a prohibited system?
- The fine tier you really do not want
- Three practical steps for SMEs right now
- References
The hardest line in the EU AI Act
The EU AI Act prohibited practices listed in Article 5 are not a grey zone. They are a flat ban. No risk assessment, no documentation, no conformity declaration makes them legal. If your organisation places on the market, puts into service or uses any system that falls into one of the categories below, you are exposed to the highest fine tier in the entire regulation: up to €35 million or 7% of global annual turnover, whichever is higher (for SMEs and start-ups, whichever is lower).
Eight prohibitions have applied since 2 February 2025, six months after the Act entered into force. So this is not theoretical preparation. It is current law. The Digital Omnibus on AI (Regulation (EU) 2026/1744) added two more, which apply from 2 December 2026; they are covered at the end of the list.
Let's walk through each prohibition in Article 5(1) with workplace-relevant examples, because HR Directors and COOs are the ones most likely to encounter these systems in practice.
(a) Subliminal manipulation
Systems that use techniques operating below the threshold of conscious awareness, or purposefully manipulative or deceptive techniques, to materially distort behaviour in a way that causes or is reasonably likely to cause significant harm are banned. Think audio or visual cues embedded in onboarding videos designed to make candidates feel anxious or compliant without realising it.
Ordinary persuasive design that users can consciously recognise is not automatically caught here. But a system built to manipulate or deceive people into harmful decisions is out.
(b) Exploitation of vulnerabilities
This targets AI that exploits a specific group's vulnerabilities based on age, disability, or social or economic situation. A debt-collection chatbot that detects financial distress and applies escalating pressure tactics would fit squarely here. So would a workplace wellbeing app that identifies employees under personal stress and nudges them toward choices that benefit the employer rather than the employee.
For HR teams, the practical question is: does the tool you are procuring identify vulnerability signals and act on them in ways the person cannot reasonably resist?
(c) Social scoring
AI systems that evaluate or classify people over time based on their social behaviour or personal characteristics, where the resulting score leads to detrimental treatment in unrelated contexts, or treatment that is unjustified or disproportionate, are banned. The original proposal limited this to public authorities; the final text applies to everyone, including private companies.
For employers, the risk is a "trustworthiness" or "culture fit" score built from employees' behaviour outside work, such as social media activity, and then used in decisions about their jobs. If a tool combines data from unrelated contexts into a score that affects how people are treated, check carefully whether it crosses this line.
(d) Criminal risk prediction based on profiling
AI systems that assess the likelihood that a person will commit a crime based solely on profiling or personality traits are banned. Predictive profiling that uses demographic data, social media behaviour, or psychometric proxies to flag individuals as future offenders has no legal place in Europe.
In recruitment, this matters. Some pre-employment screening vendors have marketed tools that claim to identify "risk profiles" in candidates. If the underlying model is predicting antisocial or criminal behaviour from personality proxies, it is caught by this provision.
(e) Untargeted facial-image scraping
Article 5(1)(e) bans the creation or expansion of facial recognition databases through untargeted scraping of the internet or CCTV footage. The prohibition hits the collection layer, not just use.
This is relevant if your organisation is considering building an internal identity verification database from photos scraped off LinkedIn or your own security camera archives. Even if the downstream use seems benign, the scraping method is prohibited.
(f) Emotion recognition in workplaces and educational institutions
This one lands directly on HR. AI systems that infer the emotions of individuals in the workplace or in educational institutions are banned. Emotion AI, affect recognition and mood-detection software in employee monitoring platforms are caught. The only exceptions are systems used for medical or safety reasons.
The prohibition covers inference, not just measurement. A system that watches employees via webcam and flags those it classifies as "disengaged" or "stressed" is inferring emotional state. That is prohibited under Article 5(1)(f). The Commission's guidelines read "workplace" as also covering candidates during selection and hiring, so emotion scoring in a video job interview is prohibited too.
If you are currently using a video interview platform that scores facial expressions, or a productivity monitoring tool that tracks emotional indicators, you need to review it now. Vendors sometimes bury emotion-detection features inside broader analytics packages. Ask explicitly: does this system infer or classify emotional states?
(g) Biometric categorisation inferring sensitive characteristics
Systems that use biometric data to categorise individuals by race, political opinion, trade union membership, religious or philosophical beliefs, or sexual orientation are banned. This applies to real-time and retrospective analysis.
The concern here is not storing biometric data for identity verification. It is inferring protected characteristics from biometric signals. A system that analyses facial geometry to deduce ethnicity or religious affiliation of individuals is caught, regardless of the stated intent. The narrow exceptions concern labelling or filtering lawfully acquired biometric datasets and certain law-enforcement uses.
(h) Real-time remote biometric identification in public spaces
The use of real-time remote biometric identification in publicly accessible spaces for the purpose of law enforcement is banned, with narrow exceptions such as searching for victims, preventing an imminent threat, or locating suspects of specific serious crimes. This provision is aimed at police and other law-enforcement bodies.
Private employers and venue operators are not caught by point (h) when they act for their own purposes. That does not make live face-matching a free choice: such biometric identification is high-risk under Annex III (with the high-risk rules applying from 2 December 2027), and the GDPR's strict rules on biometric data already apply today.
Two new bans from 2 December 2026
The Digital Omnibus on AI added two prohibitions to Article 5(1), which apply from 2 December 2026:
-
(ba) Non-consensual intimate deepfakes. AI systems that generate or manipulate realistic images, video or audio of an identifiable person's intimate parts, or of an identifiable person engaged in sexually explicit activities, without that person's explicit consent ("nudification" apps).
-
(bb) Child sexual abuse material. AI systems that generate or manipulate child sexual abuse material.
For providers, the ban applies where this is the intended purpose, or a reasonably foreseeable and reproducible outcome of a system without reasonable safeguards. For deployers, use is prohibited where they use the system for the purpose of generating such material. Most businesses will never touch this, but if you offer image-generation features to customers, check the safeguards your provider has in place.
What counts as "using" a prohibited system?
You do not need to have built the system. Article 5 prohibits not only placing these systems on the market and putting them into service, but also their use. If your organisation uses a system that falls under Article 5, it is exposed, even if you bought it off the shelf.
This is a critical point for procurement. "The vendor told us it was compliant" is not a legal defence. You need to verify what the system actually does.
The fine tier you really do not want
The EU AI Act enforcement framework places Article 5 violations at the top of the penalty scale. Tier 1 means €35 million or 7% of total worldwide annual turnover, whichever is higher. For SMEs and start-ups, the lower of the two applies (Article 99(6)). For a 50-person company with €8 million in revenue, the maximum is therefore €560,000. For a 150-person company with €30 million in revenue, it is €2.1 million.
National market surveillance authorities are responsible for enforcement. In the Netherlands, the Autoriteit Persoonsgegevens is the intended supervisor for prohibited practices, but it is not yet designated and has no AI Act powers until the Dutch implementing law is in force. In Belgium, no supervisor had been formally designated at the time of writing. The GDPR, enforced by the AP and the APD/GBA, already covers much of the same ground where personal data is involved.
Three practical steps for SMEs right now
1. Audit your current AI tools for emotion-detection features. Ask every vendor directly: does this system infer or classify emotional states? Get the answer in writing.
2. Review any pre-employment screening or candidate-assessment tools that claim to assess personality, risk, or behavioural traits. Map what the model is actually predicting and from what input data.
3. Check your CCTV and access-control systems. If any facial recognition component was added, confirm it does not build a database through untargeted scraping or infer sensitive characteristics.
The Commission's guidelines on prohibited AI practices (February 2025) give worked examples for the original eight bans. They are not binding, and they do not cover the two new bans, but they are the best guide to how authorities will read Article 5.
Start with your existing tool inventory. One afternoon of honest review is usually enough to know whether you have an Article 5 problem. That is a much better position than discovering it when a complaint lands.
Not sure which of your current AI tools might cross an Article 5 line? The free 2-minute compliance check at comply.khairos.ai gives you a quick read on where you stand across the full EU AI Act, including the prohibited practices tier, with no legal jargon and no obligation.
References
Part of our guide to High-Risk & Prohibited AI
Not sure where your company stands?
Answer 15 questions about how your team uses AI. You get a compliance score and the gaps to close first, as a PDF.