Prohibited AI practices are the uses of artificial intelligence that the EU AI Act bans outright in Article 5, because they pose an unacceptable risk to safety, fundamental rights or democratic values. They form the top tier of the regulation’s risk pyramid: unlike high-risk AI systems, which may be used after meeting strict requirements, prohibited practices may not be placed on the market, put into service or used in the European Union at all.
In AI regulation, the prohibitions are the first part of the AI Act that became applicable, from 2 February 2025. They apply to both providers and deployers, inside and outside the EU when the output is used in the Union, and they carry the highest fines in the regulation.
The eight practices banned by the EU AI Act
Article 5 lists eight categories. Each is defined by its purpose or effect rather than by a technology, so the same model can be lawful in one use and prohibited in another. Several bans contain conditions such as significant harm or specific settings, which narrow their scope.
| Practice | Core condition | Example |
|---|---|---|
| Manipulative or deceptive techniques | Materially distorts behavior and causes or is likely to cause significant harm | Subliminal cues that push users into harmful purchases |
| Exploitation of vulnerabilities | Targets age, disability or a social or economic situation | A toy that encourages children into dangerous behavior |
| Social scoring | Leads to detrimental treatment in unrelated contexts or disproportionate treatment | Denying services based on a general trustworthiness score |
| Predictive policing of individuals | Predicts criminal offenses solely from profiling or personality traits | Flagging people as future offenders from their personality |
| Untargeted facial image scraping | Builds or expands facial recognition databases from the internet or CCTV | Mass collection of faces from social media |
| Emotion recognition at work and in education | Infers emotions of employees or students | Monitoring staff mood through webcams |
| Sensitive biometric categorization | Infers race, political views, union membership, religion, sex life or sexual orientation | Sorting people by inferred religion from photos |
| Real-time remote biometric identification for law enforcement | Live identification in publicly accessible spaces | Live facial recognition across a city square |
Exceptions and narrow carve-outs
Some prohibitions have tightly defined exceptions. Emotion recognition remains possible for medical or safety reasons, such as detecting driver fatigue. Biometric categorization does not cover labeling or filtering of lawfully acquired biometric datasets, for example in law enforcement. Predictive policing is allowed when AI supports a human assessment already based on objective and verifiable facts linked to criminal activity.
The most debated exception concerns real-time remote biometric identification by law enforcement. It may be used only where strictly necessary for three purposes, and only with prior authorization by a judicial or independent administrative authority, a fundamental rights impact assessment and registration of the system. Member states must also opt in through national law.
| Purpose | Scope | Safeguard |
|---|---|---|
| Targeted search for victims | Abduction, trafficking, sexual exploitation and missing persons | Prior authorization per use |
| Prevention of an imminent threat | Specific, substantial threat to life or a genuine terrorist threat | Urgent use requires authorization within 24 hours |
| Locating suspects of serious crime | Offenses listed in the regulation with at least four years’ maximum sentence | Limits on time, place and persons |
Prohibited practices compared with other risk tiers
The AI Act applies different obligations to different levels of risk. Understanding where a system falls determines whether it is banned, regulated or largely free to operate.
| Risk tier | Legal consequence | Typical examples |
|---|---|---|
| Unacceptable risk | Banned under Article 5 | Social scoring, untargeted face scraping |
| High risk | Allowed after conformity assessment and ongoing obligations | Credit scoring, recruitment tools, medical devices |
| Transparency risk | Disclosure duties | Chatbots, deepfakes, emotion recognition outside banned settings |
| Minimal risk | No specific obligations | Spam filters, video game AI |
The boundary between tiers is use-dependent. Emotion recognition is prohibited in a workplace, but in a retail setting it is treated as a transparency case and may be high-risk under Annex III. Remote biometric identification that is not real-time is high-risk rather than banned.
Enforcement and penalties
Violations of Article 5 carry fines of up to 35 million euros or 7 percent of worldwide annual turnover, whichever is higher. For small and medium-sized enterprises the lower of the two amounts applies. National market surveillance authorities enforce the bans, and the penalty provisions have applied since 2 August 2025.
The European Commission has published guidelines on prohibited practices to explain key concepts such as material distortion, significant harm and publicly accessible spaces. The guidelines are not binding, but authorities and courts are expected to take them into account, and they are the best available reference for borderline cases.
Compliance steps and common mistakes
Organizations should screen every AI system against Article 5 before deployment and after any change of purpose. The screening is usually part of an AI inventory, together with risk classification and AI literacy measures.
- Assuming general-purpose tools are exempt. A deployer that uses a general model for a banned purpose violates the prohibition.
- Overlooking the workplace ban. Sentiment analysis of employee calls can fall under emotion recognition if it infers emotions of staff.
- Treating consent as a cure. Consent does not make a prohibited practice lawful.
- Ignoring purpose drift. A permitted analytics tool can become prohibited when it is repurposed, for example for scoring citizens.
- Forgetting non-EU providers. The ban applies whenever the output is used in the Union.
Stripped to the essentials
Most AI tools in Europe are allowed, some need careful checks, and a short list is simply forbidden. These forbidden uses are the prohibited AI practices.
They include secretly manipulating people into harmful choices and giving citizens a general social score. Reading the emotions of workers or students is banned too. So is scanning crowds to identify people in real time, except in a few serious police emergencies approved by a judge. The idea is simple: some uses are too dangerous to fix with paperwork, so they are not allowed at all.
Where it is applied
A call center operator in the EU plans to buy a software module that analyzes employees’ voices during customer calls and scores their stress, frustration and enthusiasm for performance reviews. During its AI inventory review, the compliance team checks the module against Article 5.
Because the tool infers emotions of employees in the workplace and no medical or safety reason applies, the team concludes that it falls under the ban on emotion recognition at work. The company drops the emotion scoring and keeps only a function that measures call duration and topic categories, which does not infer emotions. The decision avoids exposure to fines of up to 7 percent of global turnover and is documented in the inventory for future audits.