Lab 10: EU AI Act Compliance

Time: 50 minutes | Level: Architect | Docker: docker run -it --rm zchencow/innozverse-ai:latest bash

Overview

The EU AI Act is the world's first comprehensive AI regulation, effective August 2024. This lab covers risk classification, high-risk AI requirements, GPAI obligations, conformity assessment, and comparison with NIST AI RMF — with a practical compliance checker.

Architecture

┌──────────────────────────────────────────────────────────────┐
│                    EU AI Act Risk Pyramid                    │
├──────────────────────────────────────────────────────────────┤
│  🔴 UNACCEPTABLE RISK - PROHIBITED                          │
│  Social scoring, real-time biometric surveillance (public), │
│  subliminal manipulation, emotion recognition (work/school)  │
├──────────────────────────────────────────────────────────────┤
│  🟠 HIGH RISK - REGULATED (Title III)                       │
│  Critical infrastructure, education, employment, credit,    │
│  law enforcement, migration, justice administration          │
├──────────────────────────────────────────────────────────────┤
│  🟡 LIMITED RISK - TRANSPARENCY OBLIGATIONS                 │
│  Chatbots (disclose AI), deepfakes (label), emotion AI      │
├──────────────────────────────────────────────────────────────┤
│  🟢 MINIMAL RISK - FREE USE                                 │
│  Spam filters, AI games, recommendation systems             │
└──────────────────────────────────────────────────────────────┘

---

Step 1: Risk Tier Classification

Unacceptable Risk (Article 5 - PROHIBITED):

System	Prohibition
Social scoring by governments	Mass surveillance, scoring citizens by social behavior
Real-time biometric ID in public	Facial recognition in public spaces (exceptions: terrorism, missing persons)
Subliminal manipulation	Bypass conscious decision-making to harm users
Exploiting vulnerabilities	Target vulnerable groups (age, disability)
Predictive policing	Assess crime risk based on profiling
Emotion recognition (work/schools)	Analyze emotions of employees or students

High-Risk (Annex III Categories):

Category	Examples
Critical infrastructure	AI in electricity grids, water systems, transport
Education	Student admission, assessment systems
Employment	CV screening, promotion decisions, work monitoring
Essential services	Credit scoring, insurance risk, social benefits
Law enforcement	Polygraphs, deepfake detection, crime prediction
Migration	Asylum decisions, visa applications
Justice	Court decisions, evidence evaluation
Biometric categorization	Categorize people by race, religion, political views

💡 If your AI system directly influences a HIGH-RISK decision, you must comply with all Title III requirements. "Influencing" includes systems where output is not legally binding but practically determinative.

---

Step 2: High-Risk AI Requirements (Title III)

Organizations deploying high-risk AI must implement:

1. Risk Management System (Article 9):

- Continuous risk identification and analysis
- Testing procedures for each foreseeable risk
- Residual risk evaluation after mitigation
- Documentation of risk management activities
- Annual review minimum

2. Data Governance (Article 10):

- Training data: relevant, representative, error-free
- Documented data collection and processing practices
- Bias detection and mitigation in training data
- Special category data: only when strictly necessary with safeguards

3. Technical Documentation (Article 11):

- System architecture and design choices
- Training methodology and data sources
- Performance metrics and test results
- Cybersecurity measures
- Environmental impact (energy consumption)

4. Transparency (Article 13):

- Clear instructions for use
- AI system identification (users know they're using AI)
- Capabilities and limitations disclosure
- Human oversight requirements

5. Human Oversight (Article 14):

- Technical capability for human overseers to intervene
- Understanding what system does and can't do
- Ability to disregard/override AI outputs
- Monitoring for anomalies

6. Accuracy, Robustness, Cybersecurity (Article 15):

- Appropriate accuracy levels for intended purpose
- Resilience to errors, faults, and inconsistencies
- Robustness to adversarial manipulation
- Redundancy/fail-safe mechanisms

---

Step 3: GPAI (General Purpose AI) Obligations

GPAI models (GPT-4, Claude, Gemini, Llama 2/3) have specific obligations.

Tier 1: All GPAI models:

- Technical documentation (capabilities, limitations, training data)
- Copyright compliance for training data
- Summary of training data (even if confidential)
- EU AI Office compliance cooperation

Tier 2: GPAI with Systemic Risk (>10^25 FLOPs training compute):

Additional requirements:
- Adversarial testing (red-teaming)
- Incident reporting to EU AI Office
- Cybersecurity protection
- Energy efficiency reporting

GPAI Threshold Context:

GPT-4: ~10^25 FLOPs → Systemic risk tier
LLaMA-3-70B: ~10^23 FLOPs → Standard tier
Claude 3.5 Sonnet: Likely systemic risk tier

---

Step 4: Conformity Assessment & CE Marking

High-risk AI must undergo conformity assessment before market placement.

Assessment Paths:

System Type	Assessment Route
High-risk (most)	Internal assessment with documentation
Biometric ID systems	Third-party Notified Body assessment
Remote biometric (law enforcement)	Notified Body + national authority
GPAI systemic risk	EU AI Office oversight

CE Marking Process:

1. Risk classification → high-risk determination
2. Technical documentation preparation
3. Conformity assessment (internal or third-party)
4. EU Declaration of Conformity
5. CE mark affixed to product
6. Register in EU database (Article 60)
7. Market surveillance compliance

Registration Database:

All high-risk AI systems must register in EU AI systems database:
- System name and version
- Provider details
- Intended purpose
- Countries of operation
- Status (in use, withdrawn, etc.)

---

Step 5: NIST AI RMF Comparison

NIST AI RMF (US, Voluntary) vs EU AI Act (EU, Mandatory):

Dimension	NIST AI RMF	EU AI Act
Nature	Voluntary framework	Legal regulation
Jurisdiction	US (global influence)	EU (extraterritorial)
Approach	Risk-based, principles	Prescriptive requirements
Core Functions	GOVERN, MAP, MEASURE, MANAGE	Risk classification → requirements
Enforcement	Self-certification	Market surveillance, fines
Fines	None	Up to €35M or 7% global revenue
Timeline	Published 2023	Effective Aug 2024 (phased)

NIST AI RMF Core Functions:

GOVERN: Establish policies, culture, accountability
MAP: Categorize AI risks for specific contexts
MEASURE: Analyze and assess AI risks quantitatively
MANAGE: Prioritize and implement risk responses

Key Alignment:

Both emphasize: risk management, transparency, human oversight, documentation
EU AI Act adds: legal enforcement, CE marking, Notified Bodies, specific requirements

---

Step 6: Implementation Timeline

EU AI Act Phased Rollout:

Date	Milestone
Aug 2024	Act enters into force
Feb 2025	Prohibited practices take effect (Article 5)
Aug 2025	GPAI obligations take effect
Aug 2026	High-risk AI obligations (Annex III systems)
Aug 2027	High-risk AI obligations (Annex II systems)

Compliance Roadmap for Enterprise:

Now (2024-2025):
  → Inventory all AI systems
  → Classify by risk tier
  → Identify systems requiring compliance
  
2025:
  → Stop deploying prohibited systems
  → Begin high-risk documentation
  
2026:
  → Full high-risk compliance required
  → CE marking, registration, oversight mechanisms

---

Step 7: AI Impact Assessment

Based on DPIA (Data Protection Impact Assessment) but for AI risks.

AI Impact Assessment Template:

1. System Description
   - Purpose, functionality, outputs
   - Who makes decisions? AI, human, combined?
   
2. Stakeholder Analysis
   - Who is affected? Any vulnerable groups?
   - What are the potential harms?
   
3. Risk Assessment
   - Probability × Impact for each risk
   - Fundamental rights impact
   
4. Mitigation Measures
   - Technical: explainability, human oversight
   - Organizational: training, policies
   
5. Residual Risk Sign-off
   - DPO approval
   - Legal counsel review
   - Business owner accountability

---

Step 8: Capstone — Compliance Checker

docker run --rm zchencow/innozverse-ai:latest python3 -c "
class EUAIActChecker:
    def __init__(self, system_name):
        self.system_name = system_name
        self.scores = {}
    
    def check_risk_tier(self, use_case_category):
        unacceptable = ['social_scoring', 'subliminal_manipulation', 'real_time_biometric_public', 'emotion_recognition_workplace']
        high_risk = ['critical_infrastructure', 'education', 'employment', 'essential_services', 'law_enforcement', 'migration', 'justice', 'biometric_categorization']
        limited_risk = ['chatbot', 'deepfake', 'emotion_recognition_other']
        if use_case_category in unacceptable:
            return 'UNACCEPTABLE', 'System is PROHIBITED under EU AI Act'
        elif use_case_category in high_risk:
            return 'HIGH_RISK', 'Must comply with Title III requirements'
        elif use_case_category in limited_risk:
            return 'LIMITED_RISK', 'Transparency obligations apply'
        return 'MINIMAL_RISK', 'No specific obligations'
    
    def assess_high_risk_requirements(self, answers):
        requirements = {
            'risk_management_system': ('Has documented risk management system?', 3),
            'data_governance': ('Training data governance documented?', 3),
            'technical_documentation': ('Technical documentation complete?', 2),
            'record_keeping': ('Automatic logging enabled?', 2),
            'transparency': ('User transparency measures in place?', 3),
            'human_oversight': ('Human oversight mechanisms exist?', 3),
            'accuracy_robustness': ('Accuracy and robustness tested?', 3),
            'cybersecurity': ('Cybersecurity measures implemented?', 2),
        }
        total_score = 0
        max_score = sum(w for _, (_, w) in requirements.items())
        print(f'  High-Risk AI Requirements Assessment:')
        for req, (question, weight) in requirements.items():
            status = answers.get(req, False)
            score = weight if status else 0
            total_score += score
            icon = chr(10003) if status else chr(10007)
            print(f'    [{icon}] {question} (weight={weight})')
        pct = total_score / max_score * 100
        return total_score, max_score, pct

checker = EUAIActChecker('Enterprise Credit Scoring AI')
risk_tier, obligation = checker.check_risk_tier('employment')
print(f'=== EU AI Act Compliance Assessment ===')
print(f'System: {checker.system_name}')
print(f'Risk Tier: {risk_tier}')
print(f'Obligation: {obligation}')
print()

answers = {
    'risk_management_system': True,
    'data_governance': True,
    'technical_documentation': False,
    'record_keeping': True,
    'transparency': True,
    'human_oversight': False,
    'accuracy_robustness': True,
    'cybersecurity': True,
}
score, max_score, pct = checker.assess_high_risk_requirements(answers)
print()
compliance = 'COMPLIANT' if pct >= 80 else ('PARTIAL' if pct >= 60 else 'NON-COMPLIANT')
print(f'  Overall Score: {score}/{max_score} ({pct:.1f}%) -> {compliance}')
print()
print('NIST AI RMF vs EU AI Act:')
print('  NIST: GOVERN | MAP | MEASURE | MANAGE (voluntary, risk-based)')
print('  EU AI Act: legal requirements, conformity assessment, CE marking')
print('  Both: risk management, transparency, human oversight principles')
"

📸 Verified Output:

=== EU AI Act Compliance Assessment ===
System: Enterprise Credit Scoring AI
Risk Tier: HIGH_RISK
Obligation: Must comply with Title III requirements

  High-Risk AI Requirements Assessment:
    [✓] Has documented risk management system? (weight=3)
    [✓] Training data governance documented? (weight=3)
    [✗] Technical documentation complete? (weight=2)
    [✓] Automatic logging enabled? (weight=2)
    [✓] User transparency measures in place? (weight=3)
    [✗] Human oversight mechanisms exist? (weight=3)
    [✓] Accuracy and robustness tested? (weight=3)
    [✓] Cybersecurity measures implemented? (weight=2)

  Overall Score: 16/21 (76.2%) -> PARTIAL

NIST AI RMF vs EU AI Act:
  NIST: GOVERN | MAP | MEASURE | MANAGE (voluntary, risk-based)
  EU AI Act: legal requirements, conformity assessment, CE marking
  Both: risk management, transparency, human oversight principles

---

Summary

Concept	Key Points
Risk Tiers	Unacceptable (banned) → High → Limited → Minimal
High-Risk Requirements	Risk management, data governance, documentation, transparency, human oversight
GPAI	All models need documentation; >10^25 FLOPs → systemic risk tier
CE Marking	Conformity assessment required before market placement for high-risk AI
Timeline	Prohibited: Feb 2025; GPAI: Aug 2025; High-risk: Aug 2026
NIST vs EU	Voluntary framework vs mandatory law; aligned principles, different enforcement
Fines	Up to €35M or 7% global annual turnover

Next Lab: Lab 11: AI Cost Optimization →