Back to home
In effect since 2 February 2025

EU AI Act Prohibited Practices: What Article 5 Bans

Article 5 of the EU AI Act outright bans 8 categories of AI practice. Unlike the high-risk obligations (which apply from August 2026), these prohibitions have been enforceable since February 2025 — over a year ago.

11 min read·Article 5, EU AI Act 2024/1689·Updated June 2026

Already enforceable. The Article 5 prohibited practices came into effect on 2 February 2025, when the prohibition chapter of the EU AI Act became applicable — not the August 2026 date that applies to high-risk AI obligations. If your business is using any of the 8 prohibited practices today, enforcement action is already possible.

What Article 5 Actually Says

Article 5 of the EU AI Act creates an absolute prohibition on certain AI practices — not a risk management framework, not a transparency requirement, but a hard ban. These are AI applications that the EU legislature concluded carry risks so fundamental that no safeguard, no documentation, and no risk assessment makes them acceptable.

The eight prohibitions cover: AI systems that manipulate or exploit people psychologically, government social scoring, criminal risk profiling, facial recognition database scraping, emotion monitoring at work and school, biometric categorisation of protected characteristics, and (with narrow exceptions) real-time biometric identification by law enforcement in public spaces.

The prohibitions apply to both providers (businesses that build AI systems) and deployers (businesses that use AI systems in their operations). You cannot use a prohibited AI system even if you did not build it.

The 8 Prohibited Practices

Article 5(1)(a)Applies Broadly

1. Subliminal and Manipulative Techniques

AI that uses techniques operating below conscious awareness or that deliberately deceive or manipulate people in ways that cause them harm.

What IS prohibited

  • AI-driven nudge systems that exploit cognitive biases (scarcity pressure, false urgency, social proof manipulation) to push people toward decisions that harm them
  • Personalisation systems designed to amplify anxiety or fear to drive purchases or behaviour changes that damage the user's interests
  • AI that uses deceptive content tailored to an individual's psychological profile to influence beliefs or behaviour without their awareness

What is NOT prohibited

  • Standard A/B testing or UX optimisation that does not cause harm to users
  • Recommender systems that suggest relevant content without deceptive framing
  • Marketing personalisation that targets interests without exploiting psychological vulnerabilities

SMB relevance: Most standard marketing AI (email segmentation, ad targeting, product recommendations) is not covered here unless it is specifically designed to exploit psychological weaknesses and causes demonstrable harm.

Article 5(1)(b)Applies Broadly

2. Exploiting Vulnerabilities of Specific Groups

AI that specifically targets and exploits people who are vulnerable due to age, disability, or social and economic hardship, in ways that cause them harm.

What IS prohibited

  • AI that identifies elderly users and uses that knowledge to target them with manipulative financial or health pitches
  • Systems that detect users are in economic difficulty and exploit that to push them toward harmful decisions (payday loans, gambling, etc.)
  • Children's apps that use AI to detect emotional states and exploit them to drive in-app purchases or addictive engagement

What is NOT prohibited

  • Accessibility features that detect disabilities to provide better support
  • Age verification or parental controls
  • Welfare-oriented AI that identifies vulnerable individuals to offer help or protection

SMB relevance: The key element is exploitation for harm. Building accessible products or protective safeguards for vulnerable groups is not prohibited — only using their vulnerability against them.

Article 5(1)(c)Public Sector Focus

3. Social Scoring by Public Authorities

Government or public authority systems that rate individuals based on social behaviour and then use those ratings to restrict rights or access to services.

What IS prohibited

  • A national AI system that assigns citizens a trust or social credit score based on their online behaviour, financial history, or social activity
  • Public bodies using AI-derived scores to restrict access to education, healthcare, housing, or employment in ways that are disproportionate or based on unrelated behaviour
  • Law enforcement using an algorithmic score from one domain (e.g. tax compliance) to disadvantage individuals in an unrelated area (e.g. visa applications)

What is NOT prohibited

  • Standard credit scoring by private financial institutions (regulated separately)
  • Risk-scoring for specific, proportionate purposes in regulated contexts (e.g. fraud detection)
  • Individual public authority assessments in their domain (e.g. a benefit agency assessing eligibility for its own programme)

SMB relevance: This prohibition applies to public authorities, not private businesses. If you are a private company — even one that processes citizen data for government clients — this specific prohibition does not directly apply to you as deployer. However, your government clients may not be permitted to use outputs in the way Article 5(1)(c) describes.

Article 5(1)(d)Applies Broadly

4. Predictive Criminal Profiling Based Solely on Characteristics

AI that assesses or predicts the likelihood of an individual committing a crime based solely on their profiling or personal characteristics, rather than objective facts about actual behaviour.

What IS prohibited

  • Risk-scoring individuals as crime risks based on demographic data, postcodes, ethnicity, or social group membership
  • Predictive policing systems that deploy resources based on AI predictions derived from statistical patterns about groups rather than evidence about individuals
  • AI that flags individuals for investigation solely because they fit a demographic profile, without any supporting factual basis

What is NOT prohibited

  • AI analysis of objective, factual evidence in an ongoing investigation
  • Fraud detection that flags specific anomalous transactions based on that person's own account history
  • Recidivism tools that incorporate the full context of an individual's specific case (not just demographic profiling)

SMB relevance: This primarily affects law enforcement and justice-related systems. Private businesses running fraud detection, insurance pricing, or credit risk tools need to ensure their models do not reduce to demographic profiling, but most well-designed fraud/risk tools operate on case-specific evidence rather than group characteristics.

Article 5(1)(e)Applies Broadly

5. Facial Recognition Database Scraping

Building or expanding facial recognition databases by scraping facial images from the internet or CCTV feeds without the knowledge or consent of the people in those images.

What IS prohibited

  • Mass-harvesting photos from social media, websites, or public image repositories to train or expand a facial recognition system
  • Using CCTV footage to identify and add individuals to a recognition database without a lawful basis for doing so
  • Providing tools or APIs that enable others to perform the above at scale

What is NOT prohibited

  • Facial recognition using images that individuals have explicitly and knowingly provided for that purpose
  • Accessing lawfully obtained, properly consented datasets for specific, authorised purposes
  • Individual, case-specific identification based on a single image with a lawful basis (e.g. verifying a known person's identity)

SMB relevance: If you are developing or using any kind of face matching technology, this prohibition is highly relevant. The ban covers the act of building such databases, not just deploying them. Verify that any face dataset you use was lawfully obtained and did not involve mass scraping.

Article 5(1)(f)Check Your Tools

6. Emotion Recognition in Workplaces and Educational Institutions

AI systems that infer the emotional states of employees or students through analysis of biometric signals, video, audio, or behavioural patterns.

What IS prohibited

  • Software that monitors employee webcam footage to infer mood, engagement, or stress levels
  • Proctoring tools that flag students during exams based on inferred emotional or psychological states
  • Hiring platforms that analyse video interviews to score candidates on emotional traits
  • Workplace productivity tools that track keystrokes or mouse movements to infer employee emotional engagement

What is NOT prohibited

  • Emotion recognition for purely personal use by the individual themselves
  • Medical or safety applications that monitor emotional states for genuine health or safety reasons (e.g. detecting driver fatigue for road safety)
  • Basic engagement metrics (time-on-page, click-through rates) that do not involve biometric inference of emotional state

SMB relevance: This is one of the most immediately relevant prohibitions for SMBs. Remote work monitoring tools and AI-enhanced hiring platforms are specifically in scope. If your business uses any software that analyses employee video, voice, or behaviour to infer emotional states, review it carefully against this prohibition.

Article 5(1)(g)Applies Broadly

7. Biometric Categorisation for Sensitive Traits

AI that processes biometric data to categorise individuals according to sensitive protected characteristics: race, ethnicity, political opinions, religious beliefs, trade union membership, sex life, or sexual orientation.

What IS prohibited

  • Facial analysis tools that infer or categorise racial or ethnic origin from facial features
  • Voice analysis AI that attempts to predict political affiliation or religious beliefs from speech patterns
  • Any system that derives or infers special category personal data under GDPR from biometric inputs

What is NOT prohibited

  • Biometric authentication that verifies identity (is this person who they claim to be) without categorising sensitive traits
  • Medical imaging AI that uses biometric data for clinical diagnostic purposes
  • Document verification that compares a biometric against a stored reference for identity purposes

SMB relevance: This prohibition overlaps strongly with GDPR special category data protections. Most SMBs are not operating facial analysis systems at this level, but SaaS tools you deploy might include such capabilities. Review any vendor-supplied identity or analytics tooling carefully.

Article 5(1)(h)Law Enforcement Only

8. Real-Time Remote Biometric Identification in Public Spaces

Live identification of individuals in publicly accessible spaces using biometric data (primarily facial recognition) by law enforcement authorities. This is prohibited except for specific, narrowly defined purposes.

What IS prohibited

  • Law enforcement agencies using live facial recognition feeds to identify individuals in public spaces for general policing purposes
  • Mass surveillance of public spaces using AI to identify individuals from live video without case-specific justification

What is NOT prohibited

  • Searching historical recordings for a specific individual when investigating a serious crime
  • Live identification of specific individuals to prevent specific, imminent, and serious threats to life
  • Live identification for prosecution of specific serious crimes listed in the Act
  • Private businesses' use of access control systems (face ID to enter a building) — this is a different context to public-space surveillance

SMB relevance: This prohibition applies specifically to law enforcement use in publicly accessible spaces. Private businesses operating standard access control or customer-facing identity verification are in a different regulatory category. However, if you build or sell technology to law enforcement, this provision applies to your customers and may affect your contracting and compliance obligations.

Enforcement: What to Expect

Who enforces it?

National competent authorities (NCAs) in each EU member state are responsible for enforcement. Most are still designating their NCAs and building enforcement capacity. The European AI Office oversees GPAI model compliance at EU level.

What are the fines?

Violations of Article 5 (prohibited practices) attract the highest tier of fines: up to €35,000,000 or 7% of global annual turnover, whichever is higher. SMBs receive proportional consideration, but there is no guaranteed exemption. See the full fines guide →

GDPR as a precedent

GDPR enforcement was slow in its first 1–2 years, then accelerated significantly. The EU AI Act will likely follow a similar pattern: limited enforcement initially, but building as NCAs gain experience and as high-profile cases establish precedents. Acting now rather than waiting reduces your exposure window.

Complaints can trigger investigation

Employees, customers, or competitors can file complaints with their national NCA. Unlike regulatory own-initiative investigations, complaint-triggered reviews can move faster. Workplace monitoring tools and consumer-facing AI manipulation are natural complaint triggers.

SMB Compliance Checklist for Article 5

Run through these checks for your current AI systems and vendor tools.

Review any AI used in your hiring process — especially video interview analysis or proctoring tools
Check whether your employee monitoring, productivity, or engagement tools use biometric or emotional inference
Review your customer-facing AI for manipulative patterns: does it exploit emotional or cognitive vulnerabilities to drive decisions that harm users?
If you process facial images, confirm your dataset sources are lawfully obtained and not scraped
Review AI vendor contracts and DPAs — if a vendor's tool performs any of the 8 prohibited functions, you as the deployer share exposure
Use the risk classifier to check your AI systems against the full EU AI Act risk spectrum

Check Your AI Systems Against the Full Act

The prohibited practices in Article 5 are the most extreme end of the spectrum. Our free risk classifier covers all four risk levels and tells you your specific obligations — including whether any of your systems raise Article 5 concerns.

Run the Free Risk Classifier

Frequently Asked Questions

When did the Article 5 prohibitions come into effect?

The EU AI Act entered into force on 1 August 2024. Article 5 (prohibited practices) became enforceable six months later, on 2 February 2025. This means these prohibitions are already live — they are not part of the August 2026 deadline. If your business is currently using any of these AI practices, it is already potentially subject to enforcement.

Who enforces the Article 5 prohibitions?

Each EU member state is required to designate a national competent authority (NCA) responsible for supervising and enforcing the EU AI Act. These authorities can investigate, require remediation, and impose fines. The European AI Office handles enforcement against general-purpose AI model providers at EU level.

Do these prohibitions apply to AI tools I use but did not build?

Yes. The EU AI Act places obligations on both providers (those who build AI systems) and deployers (those who use them in their business). If you deploy an AI system that falls under Article 5 — even one purchased from a third-party vendor — you can still be in violation. Review the AI tools your business actually uses.

Does this apply to businesses outside the EU?

The EU AI Act applies wherever the AI system's output is used within the EU, or where the business deploying it is targeting EU users. A non-EU company deploying prohibited AI to EU users is still in scope. If your business has EU customers or employees, assume the Act applies to you.

What about the exception for "safety purposes" in emotion recognition?

Article 5(1)(f) specifically exempts emotion recognition used for "safety" reasons — for example, detecting driver fatigue or monitoring a patient's distress levels for clinical reasons. This exception is narrow and purpose-specific. Using it as a general cover for employee surveillance monitoring is not how regulators will interpret it.

Most of these sound extreme — does any of this realistically affect my small business?

Prohibitions 1, 2, 6, and 7 are the most likely to be relevant to SMBs through third-party tooling. Specifically: AI-enhanced hiring or proctoring platforms (prohibition 6), customer analytics tools that use emotional inference (prohibition 1), or biometric verification tools that categorise protected traits (prohibition 7). Run through the checklist below and review your key AI vendors.

Related Guides

This guide provides general information about the EU AI Act for educational purposes only. It does not constitute legal advice. The interpretation of Article 5 will be shaped by regulatory guidance and enforcement decisions as the Act is applied. For advice specific to your business, consult a qualified legal professional. Last updated June 2026.