EU AI Act: Prohibited AI Practices

Prohibited AI Practices [Art 5]

Rule: Eight categories of AI systems are absolutely prohibited in the EU due to unacceptable risks to fundamental rights and EU values. Violations carry fines up to €35M or 7% of global revenue.

Effective: February 2, 2025

Overview of Prohibition

Article 5 establishes “red lines” for AI systems that:

• Violate fundamental rights protected by the EU Charter
• Contravene EU values of human dignity and democracy
• Pose unacceptable risks regardless of safeguards

No risk mitigation allowed: These systems cannot be made compliant through technical measures.

1. Manipulative and Deceptive Techniques [Art 5.1(a)]

Prohibition

AI systems deploying subliminal techniques beyond a person’s consciousness OR purposefully manipulative or deceptive techniques that:

1. Materially distort behavior
2. Appreciably impair ability to make informed decisions
3. Cause or likely cause significant harm

####Examples

Prohibited	Description
Dark patterns	AI-generated interfaces designed to trick users into decisions against their interests
Subliminal audio/visual	Imperceptible stimuli influencing behavior (e.g., hidden messages in advertisements)
Behavioral manipulation	AI nudging people toward harmful actions (e.g., excessive gambling, dangerous challenges)
Deceptive chatbots	AI pretending to be human to manipulate emotional responses for commercial gain

Key Terms

• Subliminal: Below threshold of conscious perception
• Material distortion: Significant change in behavior that wouldn’t have occurred otherwise
• Appreciable impairment: Substantial reduction in decision-making autonomy

Exceptions

None. All forms of manipulative/deceptive AI prohibited.

2. Exploitation of Vulnerabilities [Art 5.1(b)]

Prohibition

AI systems exploiting vulnerabilities of specific groups due to:

• Age (children, elderly)
• Disability (physical or mental)
• Socioeconomic situation

Where such exploitation materially distorts behavior causing or likely to cause significant harm.

Examples

Prohibited	Description
Predatory lending	AI targeting financially vulnerable with harmful credit products
Child manipulation	AI exploiting children’s limited comprehension for commercial purposes
Elderly scams	AI-powered systems designed to confuse or mislead elderly users
Disability exploitation	AI taking advantage of cognitive impairments to extract consent

Key Elements

Must prove:

1. Exploitation of vulnerability (age/disability/socioeconomic)
2. Material distortion of behavior
3. Significant harm caused or likely

Exceptions

Legitimate accessibility tools that assist vulnerable groups (not exploit them) remain permitted.

3. Social Scoring [Art 5.1(c)]

Prohibition

AI systems for evaluation or classification of persons based on:

• Social behavior
• Known, inferred, or predicted personal or personality characteristics

Where such scoring leads to EITHER:

1. Detrimental or unfavorable treatment in contexts unrelated to where data was generated/collected, OR
2. Detrimental treatment that is unjustified or disproportionate to behavior or its gravity

Examples

Prohibited	Description
Citizen trustworthiness scores	Government-run systems rating citizens for access to services
Cross-context penalization	Low social media activity score affecting loan eligibility
Behavioral reputation systems	Scoring based on lifestyle choices unrelated to specific context
Automated social exclusion	AI denying access based on personality traits or social connections

What’s NOT Social Scoring

Permitted	Reason
Credit scoring	Based on financial behavior in relevant context (lending)
Insurance risk assessment	Actuarial calculations using directly relevant data
Fraud detection	Assessing specific transaction risk, not general “trustworthiness”
Game reputation systems	Scoring within single-context environment

Key Distinction

Context matters: Using workplace behavior to evaluate workplace performance = OK. Using workplace behavior to deny housing = PROHIBITED.

4. Predictive Policing (Risk Assessment) [Art 5.1(d)]

Prohibition

AI systems assessing risk of criminal offense based solely on:

• Profiling of a natural person, OR
• Assessing personality traits

Examples

Prohibited	Description
Recidivism prediction based on demographics	AI predicting reoffending from age/race/neighborhood
Personality-based risk scores	Assessing criminality from psychological profiling alone
Predictive hotspot mapping	Identifying individuals as threats based on location patterns
Pre-crime systems	Flagging people as future criminals without specific evidence

Exception: Human-Supported Assessment

Permitted: AI supporting human assessment of involvement in actual criminal activity using:

• Objective, verifiable facts
• Evidence-based analysis
• Human decision-maker has final authority

Example (OK): Police use AI to analyze CCTV footage to identify person matching witness description at crime scene.

Example (PROHIBITED): AI generates “high-risk offender” list based on demographic profiles.

Key Limitation

“Based solely on profiling” — if AI incorporates actual evidence of criminal involvement, not prohibited.

5. Facial Recognition Database Creation [Art 5.1(e)]

Prohibition

Creation or expansion of facial recognition databases through:

• Untargeted scraping of facial images from internet, OR
• Untargeted scraping from CCTV footage

Examples

Prohibited	Description
Web scraping for faces	AI crawling social media/websites to build facial database
Indiscriminate CCTV harvesting	Collecting all faces from public cameras without specific purpose
Mass facial data collection	Building databases from publicly accessible images without consent

What’s NOT Prohibited

Permitted	Reason
Targeted collection for specific investigation	Lawful collection with defined purpose (e.g., missing person search)
Voluntarily submitted photos	Users providing images with informed consent
Law enforcement databases from arrests	Lawfully obtained biometric data in criminal justice context

6. Emotion Recognition [Art 5.1(f)]

Prohibition

AI systems inferring emotions of natural persons in:

• Workplace settings
• Educational institutions

Examples

Prohibited	Description
Employee mood monitoring	AI analyzing facial expressions/voice tone to track worker emotions
Student engagement tracking	Systems detecting boredom or attentiveness in classrooms
Hiring emotion assessment	Analyzing candidate emotions during job interviews
Performance reviews via emotion	Using inferred emotional states as performance metrics

Exceptions

Medical or safety purposes:

Permitted	Purpose
Driver fatigue detection	Preventing accidents by detecting drowsiness
Medical diagnosis support	Clinical assessment of emotional wellbeing
Mental health monitoring	Detecting distress in healthcare settings
Pilot alertness systems	Aviation safety monitoring

Key Distinction

Context + Purpose: Emotion AI for productivity monitoring = PROHIBITED. Emotion AI for safety (fatigue detection) = OK.

7. Biometric Categorization [Art 5.1(g)]

Prohibition

AI systems categorizing individuals based on biometric data to deduce or infer:

• Race or ethnic origin
• Political opinions
• Trade union membership
• Religious or philosophical beliefs
• Sex life or sexual orientation

Exception: Labeling or filtering lawfully acquired biometric data for law enforcement purposes only.

Examples

Prohibited	Description
Sexual orientation inference from faces	AI predicting LGBTQ+ identity from facial features
Race-based profiling	Automated classification by ethnicity for non-LE purposes
Religious belief detection	Inferring faith from appearance/clothing for commercial use
Political affiliation prediction	Deducing political views from biometric characteristics

Law Enforcement Exception

Permitted ONLY for law enforcement:

• Filtering suspect images by physical characteristics in investigation
• Organizing lawfully obtained evidence by features
• Must have legal basis and fundamental rights assessment

NOT permitted: Private sector biometric categorization of protected characteristics.

8. Real-Time Remote Biometric Identification (Public Spaces) [Art 5.1(h)]

Prohibition

Real-time remote biometric identification (RBI) systems in publicly accessible spaces for law enforcement purposes.

Definition:

• Real-time: Identification without significant delay
• Remote: No physical interaction required
• Publicly accessible: Streets, parks, transport hubs, shops, etc.

Strict Exceptions

Law enforcement may use real-time RBI ONLY for:

Exception	Conditions
1. Victim searches	Targeted search for victims of abduction, trafficking, sexual exploitation, or missing persons
2. Imminent threats	Prevention of specific, substantial, imminent threat to life/safety or terrorist attack
3. Criminal suspects	Locating/identifying suspects of offenses punishable by ≥4 years imprisonment

Mandatory Safeguards

Before deployment, law enforcement MUST:

1. Fundamental Rights Impact Assessment (FRIA)

   • Assess proportionality
   • Identify affected persons
   • Evaluate alternatives

2. EU Database Registration

   • Register system in centralized database
   • Public transparency

3. Prior Judicial Authorization

   • Court or independent authority approval
   • Case-by-case basis
   • Urgent cases: ex-post authorization

4. Targeted Use Only

   • Limited to specific individuals
   • Time/geographic limits
   • Minimize false positives

What’s NOT Prohibited

• Post-event facial recognition (non-real-time) for investigating past crimes
• Border control biometrics (not “publicly accessible” under directive definition)
• Airport security checks (controlled access, not public space)
• Private premises (owner’s property, not “publicly accessible”)

Penalties [Art 99]

Violation	Fine
Article 5 breach (prohibited practices)	Up to €35,000,000 OR 7% of total worldwide annual turnover (whichever higher)
Intentional violations	Higher end of penalty range
Repeated violations	Aggravating factor

No de minimis exception: Applies to organizations of all sizes.

Enforcement Timeline

Date	Milestone
August 1, 2024	AI Act enters into force
February 2, 2025	Article 5 prohibitions become enforceable
February 4-6, 2025	European Commission publishes guidelines on prohibited practices
August 2, 2026	Full AI Act enforcement (high-risk systems, conformity assessment, etc.)

Compliance Checklist

Organizations should:

• Audit all AI systems for Article 5 prohibited practices
• Cease use of any prohibited systems immediately
• Review emotion recognition in workplace/education (unless medical/safety)
• Verify biometric systems don’t categorize by protected characteristics
• Check facial recognition doesn’t scrape internet/CCTV indiscriminately
• Ensure no social scoring across unrelated contexts
• Verify predictive policing uses objective facts, not profiling alone
• Eliminate manipulation/deception techniques in AI interactions
• Document safeguards for vulnerable group interactions
• If law enforcement: ensure real-time RBI has judicial authorization + FRIA

Interaction with Other Laws

Law	Overlap with Article 5
GDPR	Prohibited biometric processing reinforces GDPR Art 9 (special category data)
DSA	Manipulative dark patterns also regulated under Digital Services Act
Charter of Fundamental Rights	Article 5 implements Charter protections (dignity, non-discrimination, privacy)
National criminal law	Law enforcement exceptions subject to national procedural safeguards

Gray Areas and Guidance

Is my system prohibited? Decision tree:

1. Is it one of the 8 categories?

   • No → Not prohibited under Art 5 (may be high-risk under Title III)
   • Yes → Continue to step 2

2. Does an exception apply?

   • Medical/safety (emotion recognition)
   • Law enforcement (biometric categorization, real-time RBI with safeguards)
   • Human-supported assessment with objective facts (predictive policing)

3. If exception applies, are safeguards in place?

   • Judicial authorization (RBI)
   • Fundamental rights assessment
   • Database registration
   • Proportionality review

4. If no exception OR safeguards missing:

   • PROHIBITED — cease use immediately

Common Misconceptions

Myth	Reality
”Small companies exempt from Article 5”	NO — applies to all operators placing AI on EU market
”Can make prohibited AI compliant with safeguards”	NO — no risk mitigation possible for prohibited practices
”Only applies to high-risk AI systems”	NO — prohibitions apply regardless of risk classification
”Post-event facial recognition also banned”	NO — only real-time RBI in public spaces prohibited (with exceptions)
“Can use emotion AI if employees consent”	NO — workplace emotion recognition prohibited even with consent (unless safety)

Citation

Article 5 — Prohibited AI Practices, Regulation (EU) 2024/1689

Related:

• EU AI Act Service Desk: Article 5
• European Commission Guidelines on Prohibited Practices (Feb 2025)
• Future of Privacy Forum: Red Lines Under the EU AI Act