EU AI Act Article 5: Prohibited Practices as Guardrails, Implemented in Practice
EU AI Act Article 5 prohibits eight AI practices, from manipulative influence and social scoring to emotion recognition at the workplace, and at SIFAMO we turn them into rules the gateway checks: classification at registration, blocked feature classes and data combinations, a review on every change of purpose, evidence in the log.
Article 5 is the part of the AI Act that banks and insurers wave off fastest: we do not manipulate anyone, we do not run social scoring. That is true of the intent, less often of the systems. A model that optimises customer contact works out on its own that people in financial trouble respond to pressure. Everyone is addressed: providers who place a system on the market or put it into service, and deployers who use it. The fines are the highest in the regulation (Art. 99(3)). I am an engineer, not a lawyer, and what interests me is how a prohibition becomes a rule a machine can check. The basics are in my EU AI Act guide.
For context: The EU AI Act has been in force since 1 August 2024. The prohibitions and AI literacy obligations have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. The obligations for high-risk systems under Annex III apply, after the postponement by amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, and for Annex I products from 2 August 2028. Article 5 belongs to the first phase: the prohibitions have applied since 2 February 2025, and the Omnibus changed nothing about that.
What Article 5 requires
Article 5 sits in Chapter II. It does not ask for a management system, it draws lines. Art. 5(1) prohibits the placing on the market, the putting into service and the use of AI systems for eight practices:
Point (a), manipulation: subliminal techniques beyond a person’s consciousness or purposefully manipulative or deceptive techniques that materially distort behaviour by impairing the ability to make an informed decision, and that cause or are reasonably likely to cause significant harm. Lawful advertising is not in itself covered, according to recital 29.
Point (b), exploiting vulnerabilities: systems that exploit vulnerabilities due to age, disability or a specific social or economic situation, with the objective or the effect of materially distorting behaviour, with significant harm as the actual or likely result. Intent to harm is not required.
Point (c), social scoring: evaluating or classifying persons or groups over a period of time based on their social behaviour or personal characteristics, where the score leads to detrimental treatment in contexts unrelated to the one in which the data was generated (i), or to detrimental treatment that is unjustified or disproportionate (ii). Credit scoring for a loan is not meant; it is high-risk under Annex III point 5(b).
Point (d), crime prediction: assessing the risk that a person will commit a criminal offence based solely on profiling or personality traits; excluded is support for a human assessment based on objective, verifiable facts.
Point (e), facial databases: creating or expanding facial recognition databases through untargeted scraping of facial images from the internet or CCTV footage.
Point (f), emotion recognition: inferring the emotions of natural persons in the areas of workplace and education institutions, except for medical or safety reasons.
Point (g), biometric categorisation: deducing race, political opinions, trade union membership, religious or philosophical beliefs, sex life or sexual orientation from biometric data; excluded are lawfully acquired datasets and law enforcement.
Point (h), real-time remote biometric identification in publicly accessible spaces for law enforcement, with three narrow exceptions: victims and missing persons, a specific threat to life or a terrorist attack, suspects of offences listed in Annex II.
Art. 5(2) to (7) only govern this exception: a fundamental rights impact assessment under Art. 27, registration under Art. 49, prior judicial or administrative authorisation, notifications, annual reports. Of no relevance to banks and insurers. Art. 5(8): prohibitions under other Union law are unaffected.
The AI Act Explorer also shows an amendment with two further points (ba) and (bb) on sexualised deepfakes and abuse material, according to the Explorer applicable from 2 December 2026; it plays no role in this example. What matters: (a) and (b) describe effects, not techniques. Point (c) describes a change of context.
The technical example
A bank runs a customer analytics platform with three models: a segmentation by behaviour, a next-best-action model that proposes channel, timing and message per contact, and a collections model that steers the approach when a payment is overdue: reminder, call, instalment offer, debt collection. All three read the same data: account transactions, product usage, contact history, complaints, app behaviour. None of this is prohibited. The prohibition does not sit in the purpose but in three drift paths I see regularly.
First path, point (b). The collections model is optimised for incoming payments. It learns that customers with a maxed-out overdraft, returned direct debits and app usage at night respond more strongly to deadline pressure, sends them harsher messages more often and offers instalments later. These are people in a specific economic situation. If under that pressure they take out another loan, that is significant harm. Nobody programmed it. The objective function found it.
Second path, point (c). The segmentation becomes a customer value score built from payment behaviour, complaint frequency, tone in chats and branch visits over years. First marketing uses it, then service for waiting times, then the goodwill team. In the end complaint behaviour decides on goodwill in a credit card dispute. That is the change of context from point (c).
Third path, point (f). The bank introduces voice analytics in the call centre to detect customer mood. The vendor ships the same for the employee side: stress level per agent, traffic light in the team lead’s dashboard. That is emotion recognition at the workplace, prohibited, whatever the feature is called. The customer side does not fall under point (f), but it is high-risk under Annex III point 1(c) and must be disclosed under Art. 50(3).
The data flow is the same for all three: the platform calls the gateway with a use-case ID, the gateway checks registration and data zone, pseudonymises, calls the model and writes the evidence log. That is where we intervene.
How we implement it
A prohibition cannot be tested like an accuracy figure. But the conditions under which a system slides into prohibited territory can be described as data: which features it sees, which action it triggers, whom it affects, where the data comes from. That is what we do in our governance layer. The registry record of the collections model, simplified:
Every use case receives a class at registration: prohibited, high-risk, limited, minimal. Before that, the Article 5 screen runs: eight questions along points (a) to (h), each answered yes or no with a short justification. Plus three fields the gateway reads later: subjects (customer, employee, third party), action class (such as steer contact or infer emotion) and data contexts (such as payment or complaints). The business owner and compliance sign. A prohibited use case gets no route, but a record with the status rejected. It stays: the fact that someone requested emotion recognition for the team and was turned down is evidence in itself.
Scope rules in the gateway
Registration is an intention. It is enforced in the gateway, per call:
Blocked feature classes. In the data catalogue, fields that point to vulnerabilities under point (b) are tagged: age bands, indications of disability, distress indicators such as overdraft utilisation, returned direct debits and incoming social benefits. A use case with the action class steer contact may not receive them. The gateway strips them from the payload and logs that. The collections model may know that a payment is overdue, but not how desperate the customer is.
Blocked data combinations. If a field arrives from a context the use case is not registered for, the gateway rejects the call. The goodwill use case gets no complaint history, service gets no payment status. That is our answer to point (c): the change of context is prevented, not merely forbidden.
Blocked combination of action and subject. Infer emotion plus employee cannot be configured in the gateway, with no exception flag, not even for safety reasons. Whoever genuinely needs that exception starts an occupational health project with its own registration.
Approved templates. The collections model chooses from message templates that the business unit and compliance have signed off, and writes no free text to customers. A model that cannot phrase artificial urgency satisfies point (a) by construction.
Review on change of purpose
The most dangerous moment is not go-live but the second use: the score from marketing wanders into goodwill. So every change to purpose, subjects, action class or data context is a new registration with a new screen and a new signature. The expiry date forces the screen to run again at the latest then, looking at what the model does by now, not at what was requested.
Evidence that the check happened
The evidence log keeps, per decision, the policy version, the version of the Article 5 screen and the stripped fields. Monitoring reads from it out-of-scope attempts and distributions: if contact intensity concentrates on segments that correlate with the blocked features, an alert goes to the business owner. Proxies are real. The auditor sees all use cases with class, screen, signatures and date, the rejected requests, and per decision the rule that applied.
Where the law leaves room, we choose the narrower reading: for us financial distress is a specific economic situation under point (b), and we treat cross-domain scores like social scoring. The cost of the rule is a few blocked fields. The cost of being wrong is in Art. 99(3).
Typical mistakes
Article 5 ticked off once. “We do not do that” as a sentence in the AI policy, no screen per use case, no review on change.
Objective function without guardrails. Maximising incoming payments without taking away the features through which the model finds vulnerabilities.
The one customer score for everything. Built once, used by five departments, owned by nobody.
Employee analytics as coaching. Stress analysis in the call centre under the label of quality assurance. The label does not change the legal position.
Prohibition only in the prompt. “Do not put customers under pressure” is an instruction, not a boundary.
EU AI Act Article 5 checklist
Every AI use case has a documented class and a signed Article 5 screen along points (a) to (h).
Features on age, disability or financial distress are tagged and technically blocked for use cases that steer contact.
Data contexts are registered per use case; cross-domain combinations are rejected by the gateway and logged.
Emotion recognition relating to employees cannot be configured; on customers it is registered as high-risk and disclosed.
Customer communication from models uses approved templates.
Every change to purpose, subjects, action or data context re-triggers screen and signature; the expiry date forces the repeat.
The evidence log holds, per decision, the screen version and the stripped fields; rejected use cases stay in the registry as evidence.
Open questions
Where does significant harm begin? For collections processes there is no measurable threshold. The Commission’s guidelines on prohibited practices from February 2025 help with interpretation but bind no court.
Emotion from text. Art. 3(39) ties emotion recognition to biometric data. Voice is biometric, text is not. Whether frustration detection from employees’ chat transcripts falls under point (f) is open. We block it anyway.
Proxies. The gateway can block fields, but it cannot prevent a model from reconstructing distress from permitted data. Monitoring is a patch; the real lever is research on interpretability and fairness under proxy features.
Frequently asked questions
Does this apply to us as a deployer too?
Yes, without qualification. Art. 5(1) explicitly prohibits the use as well. Whoever switches on a purchased call centre tool with mood analysis for employees is in breach themselves, regardless of the vendor.
Is our credit scoring social scoring?
No, as long as it is done for the credit purpose with the data that is lawful for it. It is then high-risk under Annex III point 5(b). It becomes prohibited when the score is fed from unrelated contexts or leads to detrimental treatment there.
Is a prohibition in our AI policy enough?
As a document yes, as a control no. The policy says what nobody should do. Screen, blocked fields and log show that nobody did it. That is the difference between a statement of intent and AI governance.
EU AI Act Article 5: Prohibited Practices as Guardrails, Implemented in Practice
EU AI Act Article 5 prohibits eight AI practices, from manipulative influence and social scoring to emotion recognition at the workplace, and at SIFAMO we turn them into rules the gateway checks: classification at registration, blocked feature classes and data combinations, a review on every change of purpose, evidence in the log.
Article 5 is the part of the AI Act that banks and insurers wave off fastest: we do not manipulate anyone, we do not run social scoring. That is true of the intent, less often of the systems. A model that optimises customer contact works out on its own that people in financial trouble respond to pressure. Everyone is addressed: providers who place a system on the market or put it into service, and deployers who use it. The fines are the highest in the regulation (Art. 99(3)). I am an engineer, not a lawyer, and what interests me is how a prohibition becomes a rule a machine can check. The basics are in my EU AI Act guide.
For context: The EU AI Act has been in force since 1 August 2024. The prohibitions and AI literacy obligations have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. The obligations for high-risk systems under Annex III apply, after the postponement by amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, and for Annex I products from 2 August 2028. Article 5 belongs to the first phase: the prohibitions have applied since 2 February 2025, and the Omnibus changed nothing about that.
What Article 5 requires
Article 5 sits in Chapter II. It does not ask for a management system, it draws lines. Art. 5(1) prohibits the placing on the market, the putting into service and the use of AI systems for eight practices:
The AI Act Explorer also shows an amendment with two further points (ba) and (bb) on sexualised deepfakes and abuse material, according to the Explorer applicable from 2 December 2026; it plays no role in this example. What matters: (a) and (b) describe effects, not techniques. Point (c) describes a change of context.
The technical example
A bank runs a customer analytics platform with three models: a segmentation by behaviour, a next-best-action model that proposes channel, timing and message per contact, and a collections model that steers the approach when a payment is overdue: reminder, call, instalment offer, debt collection. All three read the same data: account transactions, product usage, contact history, complaints, app behaviour. None of this is prohibited. The prohibition does not sit in the purpose but in three drift paths I see regularly.
First path, point (b). The collections model is optimised for incoming payments. It learns that customers with a maxed-out overdraft, returned direct debits and app usage at night respond more strongly to deadline pressure, sends them harsher messages more often and offers instalments later. These are people in a specific economic situation. If under that pressure they take out another loan, that is significant harm. Nobody programmed it. The objective function found it.
Second path, point (c). The segmentation becomes a customer value score built from payment behaviour, complaint frequency, tone in chats and branch visits over years. First marketing uses it, then service for waiting times, then the goodwill team. In the end complaint behaviour decides on goodwill in a credit card dispute. That is the change of context from point (c).
Third path, point (f). The bank introduces voice analytics in the call centre to detect customer mood. The vendor ships the same for the employee side: stress level per agent, traffic light in the team lead’s dashboard. That is emotion recognition at the workplace, prohibited, whatever the feature is called. The customer side does not fall under point (f), but it is high-risk under Annex III point 1(c) and must be disclosed under Art. 50(3).
The data flow is the same for all three: the platform calls the gateway with a use-case ID, the gateway checks registration and data zone, pseudonymises, calls the model and writes the evidence log. That is where we intervene.
How we implement it
A prohibition cannot be tested like an accuracy figure. But the conditions under which a system slides into prohibited territory can be described as data: which features it sees, which action it triggers, whom it affects, where the data comes from. That is what we do in our governance layer. The registry record of the collections model, simplified:
Classification at registration
Every use case receives a class at registration: prohibited, high-risk, limited, minimal. Before that, the Article 5 screen runs: eight questions along points (a) to (h), each answered yes or no with a short justification. Plus three fields the gateway reads later: subjects (customer, employee, third party), action class (such as steer contact or infer emotion) and data contexts (such as payment or complaints). The business owner and compliance sign. A prohibited use case gets no route, but a record with the status rejected. It stays: the fact that someone requested emotion recognition for the team and was turned down is evidence in itself.
Scope rules in the gateway
Registration is an intention. It is enforced in the gateway, per call:
Review on change of purpose
The most dangerous moment is not go-live but the second use: the score from marketing wanders into goodwill. So every change to purpose, subjects, action class or data context is a new registration with a new screen and a new signature. The expiry date forces the screen to run again at the latest then, looking at what the model does by now, not at what was requested.
Evidence that the check happened
The evidence log keeps, per decision, the policy version, the version of the Article 5 screen and the stripped fields. Monitoring reads from it out-of-scope attempts and distributions: if contact intensity concentrates on segments that correlate with the blocked features, an alert goes to the business owner. Proxies are real. The auditor sees all use cases with class, screen, signatures and date, the rejected requests, and per decision the rule that applied.
Where the law leaves room, we choose the narrower reading: for us financial distress is a specific economic situation under point (b), and we treat cross-domain scores like social scoring. The cost of the rule is a few blocked fields. The cost of being wrong is in Art. 99(3).
Typical mistakes
EU AI Act Article 5 checklist
Open questions
Frequently asked questions
Does this apply to us as a deployer too?
Yes, without qualification. Art. 5(1) explicitly prohibits the use as well. Whoever switches on a purchased call centre tool with mood analysis for employees is in breach themselves, regardless of the vendor.
Is our credit scoring social scoring?
No, as long as it is done for the credit purpose with the data that is lawful for it. It is then high-risk under Annex III point 5(b). It becomes prohibited when the score is fed from unrelated contexts or leads to detrimental treatment there.
Is a prohibition in our AI policy enough?
As a document yes, as a control no. The policy says what nobody should do. Screen, blocked fields and log show that nobody did it. That is the difference between a statement of intent and AI governance.
This post is part of the series on the EU AI Act article by article, all articles at a glance.
Related: EU AI Act · All articles · AI Governance · AI Consultant
Recent Posts
Recent Comments