The EU AI Act uses a risk-based approach. Classifying your AI systems correctly determines your compliance requirements. This guide explains each category with real examples.

The Risk Hierarchy

Prohibited (Banned)
        ↓
High-Risk (Strict Requirements)
        ↓
Transparency Risk (Mid-Risk)
        ↓
Low-Risk (Basic)
        ↓
Unregulated (No Requirements)

Category 1: Prohibited AI (Article 5)

These applications are illegal in the EU, no exceptions.

1.1 Social Credit Scoring

What it is: AI that assigns overall "scores" to people based on their social behavior, used for rewards/punishments.

Examples that are prohibited:

  • China-style system rating citizens overall behavior
  • Company system that tracks "trustworthiness" and restricts services
  • Credit system that includes behavior outside financial domain

Examples that are OK:

  • Credit scoring based only on financial data
  • Loan approval based on income and debt
  • Community moderation (removing harmful posts)

Why: Violates human dignity and creates discriminatory harm.

1.2 Real-Time Facial Recognition in Public Spaces

What it is: Using AI to identify people in real-time in public areas via cameras.

Prohibited uses:

  • Police identifying suspects from street cameras without warrant
  • Shopping centers tracking customers
  • Airports recognizing travelers

Exceptions allowed:

  • Law enforcement with court order and strict safeguards
  • Border control with legal basis
  • Finding missing children/endangered persons
  • Preventing specific crime with legal authorization

Practical note: Most law enforcement use is still prohibited without additional legal justification. The bar is very high.

1.3 Subliminal Manipulation

What it is: Hidden content designed to alter behavior without conscious awareness.

Prohibited examples:

  • Flashing hidden images in videos to influence decisions
  • Audio content below hearing threshold
  • Visual patterns designed to manipulate without awareness
  • Exploitation of psychological vulnerabilities by design

Key: "Without the person being able to perceive" the content.

1.4 Exploitation of Vulnerabilities

What it is: AI deliberately exploiting vulnerable groups (children, disabled, elderly) to harm them.

Prohibited:

  • AI targeting children to manipulate them into addictive behavior
  • Scam systems targeting elderly or cognitively disabled
  • Systems designed to exploit people with gambling addictions

Allowed:

  • Age-appropriate content for children
  • Accessibility features for disabled users
  • Credit checks that consider ability to repay

Category 2: High-Risk AI (Annex III)

High-risk systems require extensive documentation, testing, and controls. If you use high-risk AI, see eu-ai-act-guide.mdx for full compliance requirements.

Subcategory 2.1: Biometric Identification & Categorization

AI that identifies or categorizes people via biometrics.

Includes:

  • Facial recognition
  • Fingerprint matching
  • Iris/retina scanning
  • Gait recognition
  • Voice recognition (for identification, not just activation)
  • Emotion recognition from facial/body analysis
  • Gender/race/age detection from images
  • Medical condition detection from appearance

Examples:

  • Airport facial recognition for border control ✓ High-Risk
  • Phone unlock via face (not high-risk—low consequence, user-initiated)
  • Police facial search database ✓ High-Risk
  • Dating app detecting age from photo ✓ High-Risk
  • Surveillance system identifying "suspicious" people ✓ High-Risk

Why high-risk: Biometric data is permanent, can't be changed, and affects fundamental rights.

Subcategory 2.2: Critical Infrastructure

AI that controls critical systems.

Includes:

  • Power grid management
  • Water system operations
  • Gas distribution
  • Transportation networks
  • Utilities management
  • Emergency services dispatch

Examples:

  • AI managing electricity distribution ✓ High-Risk
  • Smart grid load balancing ✓ High-Risk
  • Traffic light optimization (local city) — Depends (probably not high-risk)
  • Rail switching systems ✓ High-Risk

Why: Failures directly endanger lives and public safety.

Subcategory 2.3: Law Enforcement

AI used by police and courts.

Includes:

  • Crime prediction or pattern analysis
  • Evidence assessment or valuation
  • Suspect identification
  • Risk assessment (recidivism prediction)
  • Case prioritization

Examples:

  • Predictive policing (high-risk) ✓
  • Facial identification for suspects ✓ High-Risk
  • Automated flagging of suspicious transactions ✓ High-Risk
  • Risk assessment for bail decisions ✓ High-Risk
  • Weapon detection in security footage ✓ High-Risk

Why: Directly impacts liberty and justice. Errors have severe consequences.

Subcategory 2.4: Employment & Labor

AI making or informing employment decisions.

Includes:

  • Resume screening or ranking
  • Interview analysis or evaluation
  • Performance evaluation
  • Promotion/bonus decisions
  • Shift scheduling (if affects livelihood)
  • Redundancy decisions
  • Wage decisions

Examples:

  • Resume screening tool ✓ High-Risk
  • Interview analysis rating candidates ✓ High-Risk
  • Performance tracking software ✓ High-Risk
  • AI determining who gets raises ✓ High-Risk
  • Schedule optimization (just efficiency, not firing) — May be high-risk
  • Recommendation system (helps manager, manager decides) — Probably high-risk

Why: Affects fundamental right to work and livelihood.

Subcategory 2.5: Education

AI making educational decisions about individuals.

Includes:

  • Grading or assessment
  • Admission decisions
  • Course or program recommendations
  • Performance evaluation (students)
  • Educational resource allocation
  • Dropout prediction

Examples:

  • AI grading essays ✓ High-Risk
  • University admissions ranking ✓ High-Risk
  • Student performance prediction ✓ High-Risk
  • Course recommendation system ✓ High-Risk (if used for decisions)
  • Automated flagging of struggling students ✓ High-Risk

Why: Affects future opportunities and life trajectory.

Subcategory 2.6: Credit & Financial Services

AI for creditworthiness or financial inclusion decisions.

Includes:

  • Credit scoring
  • Loan approval/denial
  • Interest rate determination
  • Loan terms (amount, duration)
  • Insurance underwriting
  • Insurance pricing
  • Mortgage qualification

Examples:

  • Credit scoring tool ✓ High-Risk
  • Loan approval based on AI ✓ High-Risk
  • Insurance underwriting ✓ High-Risk
  • Automated mortgage denial ✓ High-Risk
  • Banking fraud detection — Not high-risk (just flagging, human reviews)
  • Invoice payment prediction — Not high-risk (operational, not decisional)

Why: Affects financial stability and economic opportunity.

Subcategory 2.7: Benefits & Social Services

AI determining eligibility for social support.

Includes:

  • Welfare eligibility determination
  • Benefit level decisions
  • Social housing assignment
  • Child protection decisions
  • Healthcare rationing
  • Eligibility for state services

Examples:

  • Unemployment benefit eligibility AI ✓ High-Risk
  • Healthcare need assessment ✓ High-Risk
  • Social housing allocation ✓ High-Risk
  • Food assistance determination ✓ High-Risk

Why: Directly affects ability to meet basic needs.

AI used in legal/judicial processes.

Includes:

  • Case outcome prediction
  • Sentencing recommendations
  • Bail/parole decisions
  • Evidence evaluation or prioritization
  • Legal research prioritization
  • Judicial decision assistance

Examples:

  • Sentencing recommendation AI ✓ High-Risk
  • Bail risk assessment ✓ High-Risk
  • Parole eligibility prediction ✓ High-Risk
  • Court case outcome prediction ✓ High-Risk

Why: Affects liberty and justice. Most serious consequences for individuals.

Category 3: Transparency Risk (Article 50)

These systems have mid-level requirements: mostly disclosure and documentation.

3.1 AI-Generated or Manipulated Content

What it is: Content created or substantially altered by AI.

Includes:

  • Deepfakes (synthetic faces/voices)
  • Text generated by large language models
  • AI-created images or art
  • Voice synthesis
  • Video synthesis

Requirement: Disclose that content was AI-generated.

Examples requiring disclosure:

  • Blog post written by ChatGPT ("This article was written with AI assistance")
  • AI-generated product images
  • Deepfake video (even if used for satire)
  • Synthetic voice in narration
  • AI-created music

Not requiring disclosure:

  • Spell checkers and grammar tools (not AI-generated, assisted)
  • Search engine results
  • Simple templates
  • Data visualization

Why: People have right to know if they're interacting with AI, not a human.

3.2 Chatbots & Conversational AI

What it is: AI systems that simulate conversation with humans.

Requirement: Disclose that user is talking to an AI, not a human.

Examples:

  • ChatGPT-based customer support ("Talking to AI assistant")
  • Large language model chatbots
  • Conversational virtual agents

Not including:

  • Traditional chatbots with fixed responses
  • Keyword-matching Q&A systems
  • Simple voice assistants (Alexa, Google)

Why: People should know they're not talking to a human.

3.3 Biometric Emotion Recognition

What it is: AI inferring emotions from facial expressions, body language, tone.

Requirements:

  • Disclose it's being used
  • Document limitations and unreliability
  • Safeguards against misuse

Examples:

  • Job interview analysis detecting "stress"
  • Classroom monitoring detecting "engagement"
  • Lie detection via emotional analysis
  • Recruitment AI analyzing interview demeanor

Why: Emotion recognition is unreliable and can be misused to harm. Employees/candidates should know they're being evaluated this way.

3.4 Content Recommendation Systems

What it is: AI that recommends content to users (not making decisions, just suggestions).

Requirements:

  • Explain main factors affecting recommendations
  • Document algorithm
  • Safeguards against filter bubbles

Note: If recommendation system is used for legal decisions (e.g., "algorithm recommends loan denial"), it becomes high-risk.

Examples:

  • YouTube recommendation algorithm
  • Spotify playlist recommendations
  • Netflix "recommendations for you"
  • E-commerce "similar products"

Why: Algorithms can create filter bubbles and polarization. Transparency helps understand why you see what you see.

Category 4: Low-Risk / General Compliance

These systems have minimal specific requirements (just general GDPR, consumer protection).

4.1 Traditional Machine Learning

  • Spam filters
  • Demand forecasting
  • Price optimization
  • Customer segmentation
  • Churn prediction
  • Inventory management
  • Quality control

Requirement: Basic GDPR if personal data is involved, but no special AI Act requirements.

4.2 Simple Automation

  • Workflow automation (n8n, Zapier)
  • Data transformation
  • Scheduled tasks
  • Report generation

Requirement: If involving personal data, GDPR compliance. Otherwise, minimal requirements.

4.3 Data Analytics

  • Business intelligence
  • Trend analysis
  • Performance dashboards
  • Historical analysis

Requirement: GDPR if analyzing personal data.

Classification Checklist

When classifying a system, ask in order:

  1. Is it on the prohibited list? → PROHIBITED (remove immediately)

  2. Does it involve:

    • Automated decision-making affecting rights? → HIGH-RISK
    • Biometric identification? → HIGH-RISK
    • Employment/hiring/promotion decisions? → HIGH-RISK
    • Credit/financial decisions? → HIGH-RISK
    • Legal/judicial decisions? → HIGH-RISK
    • Education/grading? → HIGH-RISK
    • Welfare/benefits eligibility? → HIGH-RISK
    • Law enforcement use? → HIGH-RISK
    • Critical infrastructure? → HIGH-RISK

  3. Does it involve:

    • AI-generated content? → TRANSPARENCY
    • Chatbots? → TRANSPARENCY
    • Emotion recognition? → TRANSPARENCY
    • Recommendations (not decisions)? → TRANSPARENCY

  4. Otherwise → LOW-RISK

Real-World Classification Examples

Example 1: Resume Screening Tool

1. Prohibited? No
2. Automated decision-making? Yes (screening in/out)
3. Employment-related? Yes

Classification: HIGH-RISK
Compliance needed: Full high-risk compliance

Example 2: Customer Support Chatbot

1. Prohibited? No
2. Automated decision? No (just answering questions)
3. Chatbot? Yes

Classification: TRANSPARENCY
Compliance needed: Disclose "AI assistant"

Example 3: Email Spam Filter

1. Prohibited? No
2. Automated decision? No (just flagging)
3. Emotion/biometric? No
4. Content recommendation? No

Classification: LOW-RISK
Compliance needed: Basic (just GDPR if email data)

Example 4: Sales Price Optimization

1. Prohibited? No
2. Decision? Not really (just suggestions to business)
3. Recommendation? Not individual
4. GDPR data? Probably

Classification: LOW-RISK (or basic transparency)
Compliance needed: GDPR compliance, basic documentation

Checklist

  • List all your AI systems
  • Check each against prohibited list (remove if any)
  • Classify each as high-risk, transparency, or low-risk
  • For each high-risk: Plan compliance work
  • For each transparency: Plan disclosure mechanisms
  • Document your classification decisions
  • Get second opinion from legal counsel (if high-risk)
  • Update classification quarterly (AI capabilities change)
  • Train team on what qualifies as what risk level

Sources: