Understanding how AI systems are classified under the EU AI Act's risk-based framework.
The Four Risk Levels
The EU AI Act categorizes AI systems into four risk levels based on their potential impact on safety, fundamental rights, and society.
Risk Assessment Criteria
Classification depends on:
- Use case and purpose of the AI system
- Sector and domain of deployment
- Decision-making authority (automated vs. human-in-the-loop)
- Impact on individuals and fundamental rights
- Potential for harm if the system fails or is misused
Level 1: Unacceptable Risk 🚫
Definition: AI systems that pose a clear threat to safety, livelihoods, and rights of people.
Status: Prohibited - Cannot be developed, deployed, or used in the EU.
Examples
Social Scoring
Government-operated systems that evaluate or classify individuals based on behavior, social status, or personal characteristics.
Why Prohibited: Violates human dignity and can lead to discriminatory treatment.
Subliminal Manipulation
AI that uses techniques beyond human consciousness to materially distort behavior in a harmful way.
Example: Audio signals imperceptible to humans designed to influence purchasing decisions.
Exploiting Vulnerabilities
Systems that exploit vulnerabilities of specific groups (age, disability, economic situation).
Example: Voice-activated toy that encourages dangerous behavior in children.
Real-Time Remote Biometric Identification
Live facial recognition in publicly accessible spaces (limited law enforcement exceptions).
Example: Mass surveillance cameras with real-time face matching.
Exceptions
Limited exceptions exist for law enforcement in cases of:
- Searching for missing persons
- Preventing imminent terrorist attacks
- Prosecuting serious crimes
Even then, strict safeguards and prior judicial authorization are required.
Level 2: High Risk ⚠️
Definition: AI systems likely to adversely affect safety or fundamental rights.
Status: Permitted with strict compliance requirements.
High-Risk Categories
1. Biometric Systems
- Biometric identification (except real-time remote, which is prohibited)
- Biometric categorization based on sensitive attributes
- Emotion recognition systems
2. Critical Infrastructure
- AI in road traffic and supply of water, gas, heating, electricity
3. Education and Vocational Training
- Determining access to educational institutions
- Assessing students
- Detecting cheating
4. Employment
- Recruitment and selection
- Task allocation decisions
- Monitoring and evaluating performance
- Promotion and termination decisions
5. Essential Services
- Credit scoring affecting access to credit
- Assessment for emergency services eligibility
- Risk assessment for life and health insurance pricing
6. Law Enforcement
- AI for assessing crime recidivism risk
- Polygraphs and similar tools
- Evaluating reliability of evidence
- Crime prediction or profiling
- Detecting deep fakes
7. Migration and Border Control
- Assessing security risks posed by individuals
- Verifying travel documents
- Assisting authorities in examining applications
8. Justice and Democracy
- Influencing election outcomes
- Recommending laws to courts
Compliance Requirements
High-risk systems must meet:
Risk Management
- Continuous identification and mitigation of risks
- Testing throughout the lifecycle
Data Quality
- Relevant, representative datasets
- Examination for bias
- Data governance measures
Technical Documentation
- Detailed system design description
- Development and testing processes
- Performance metrics
Transparency
- Clear user information
- System capabilities and limitations
- Instructions for use
Human Oversight
- Ability to understand system output
- Power to override decisions
- Intervention mechanisms
Accuracy and Robustness
- Appropriate performance levels
- Resilience to errors
- Cybersecurity measures
Record-Keeping
- Automatic logging of events
- Traceability of operations
Level 3: Limited Risk ⚡
Definition: AI systems with specific transparency obligations.
Status: Permitted with transparency requirements.
Examples and Requirements
Chatbots and Virtual Assistants
Requirement: Inform users they are interacting with AI
Example Disclosure:
"You are chatting with an AI assistant. Responses are generated automatically."
Emotion Recognition
Requirement: Inform individuals that emotion detection is being used
Example:
"This system uses AI to detect emotions from facial expressions for customer feedback analysis."
Biometric Categorization
Requirement: Inform individuals when categorization is used
Example:
"This entry system uses age estimation technology."
Deep Fakes and Synthetic Content
Requirement: Clearly label AI-generated content
Example:
"This image was generated using AI / This video contains artificially generated content."
Transparency Best Practices
- Clear and conspicuous notice
- Before or at the time of interaction
- Plain language understandable by average users
- Easily accessible information about the AI system
Level 4: Minimal Risk ✅
Definition: AI systems with low or no risk to rights and safety.
Status: No specific legal obligations (voluntary codes of conduct encouraged).
Examples
- Spam filters
- AI-enabled video games
- Inventory management systems
- Weather forecasts
- Simple recommendation systems (non-profiling)
- Grammar checkers
- Image enhancement tools
Voluntary Measures
While not required, providers may:
- Adopt voluntary codes of conduct
- Follow ethical AI principles
- Implement basic transparency
- Maintain quality standards
How Classification Works in Arcus
Automated Analysis
When you register a system, Arcus AI:
- Analyzes your system description
- Identifies use cases and risk indicators
- Checks against high-risk categories
- Applies EU AI Act criteria
- Determines risk classification
Confidence Scoring
Each classification includes a confidence score:
- High (80-100%): Very confident in classification
- Moderate (60-79%): Reasonably confident, review recommended
- Low (<60%): Expert review strongly recommended
Multi-Jurisdiction Support
Systems can be assessed under multiple frameworks:
- EU AI Act
- UK AI Regulation Approach
- US AI Executive Order
- Other regional frameworks
Edge Cases and Gray Areas
Dual-Use Systems
Some systems may fall into multiple categories depending on use:
Example: Facial recognition
- Consumer photo tagging app → Minimal risk
- Employee time tracking → High risk (employment)
- Law enforcement → High risk (law enforcement) or Prohibited (real-time remote)
Solution: Classify based on primary intended use case.
General Purpose AI
Foundation models and general-purpose AI:
- Have specific requirements under the Act
- Downstream applications classified separately
- Provider and deployer responsibilities differ
Systems with Multiple Components
Example: HR platform with multiple AI features
- Resume screening → High risk
- Interview scheduling → Minimal risk
- Employee chat → Limited risk
Solution: Classify each AI component separately.
Impact of Classification
| Risk Level | Compliance Burden | Time to Comply | Cost Impact |
|---|---|---|---|
| Unacceptable | Cannot deploy | N/A | Cannot use |
| High | Significant | 6-12 months | High |
| Limited | Low | 1-2 months | Low |
| Minimal | None | Immediate | None |
When to Re-Classify
Trigger re-classification when:
- System purpose changes significantly
- New features are added
- Use case expands to new sectors
- User base changes (e.g., starts affecting vulnerable groups)
- Deployment context shifts
- Regulations update (Arcus will alert you)
Getting Help with Classification
Low Confidence?
If Arcus's confidence score is below 60%:
- Request expert review (Professional and Business plans)
- Provide more details about your system
- Consult with compliance team
Disagree with Classification?
If you believe the classification is incorrect:
- Review the reasoning provided by Arcus
- Submit a challenge with justification
- Request human expert review
Remember: Classifications must be justified and documented for audits.
Learn More
Risk classification is a legal determination. While Arcus provides guidance, consult legal professionals for final compliance decisions.