AI Act | Risk Classification

Detailed guide for classifying AI systems according to AI Act risk categories.

Category Overview

Prohibited Practices (Article 5)

These AI systems are PROHIBITED and must not be used.

List of Prohibited Practices

#	Practice	Description	Exceptions
1	Subliminal manipulation	AI manipulating human behaviour without awareness	None
2	Exploiting vulnerabilities	Exploiting vulnerabilities (age, disability)	None
3	Social scoring	Evaluating persons based on behaviour	None
4	Real-time biometric ID	Biometric identification in public spaces	Exceptions for law enforcement
5	Predictive policing (individual)	Predicting criminality of individuals	None
6	Emotion recognition (workplace/school)	Emotion recognition in the workplace/school	Health and safety reasons
7	Biometric categorisation	Categorisation by race, religion, orientation	None
8	Facial recognition scraping	Building databases from public sources	None

Self-assessment Checklist

We do not use subliminal manipulation
We do not exploit vulnerabilities of persons
We do not perform social scoring
We do not use real-time biometric ID in public
We do not perform individual predictive policing
We do not recognise emotions of employees/students (without exception)
We do not categorise persons by sensitive attributes
We do not scrape biometric data from public sources

If anything is checked -> Immediately STOP + consult with Legal

High-Risk AI (Article 6, Annex III)

Annex III --- High-Risk Areas

Area	Examples	Article
1. Biometrics	Biometric identification, categorisation	III.1
2. Critical infrastructure	AI in energy, transport, water	III.2
3. Education	Admissions, grading, proctoring	III.3
4. Employment	Recruitment, HR decisions, performance	III.4
5. Essential services	Credit, insurance, social benefits	III.5
6. Law enforcement	Profiling, polygraph, risk assessment	III.6
7. Migration	Visa decisions, border control	III.7
8. Justice	Legal research, sentencing	III.8

Detailed High-Risk AI Examples

High-Risk Obligations

Obligation	Description	Deadline
Risk Management	Documented risk management process	Before deployment
Data Governance	Data quality, bias testing, documentation	Ongoing
Technical Documentation	Model card, training info, limitations	Before deployment
Record Keeping	Audit log, min. 5 years	Ongoing
Transparency	Information for users	Before deployment
Human Oversight	Human-in-the-loop	Always
Accuracy & Robustness	Testing, monitoring	Ongoing
Conformity Assessment	Self-assessment or third-party	Before deployment

Limited Risk / Transparency (Article 50)

Systems Requiring Transparency

System	Obligation
Chatbots	Inform users they are communicating with AI
Generative AI	Label AI-generated content
Deepfakes	Clearly label as artificial
Emotion recognition	Inform affected persons
Biometric categorisation	Inform affected persons

Implementing Transparency

Examples of correct labelling:

CHATBOT:
"You are communicating with an AI assistant. To speak
with a human, type 'operator'."

GENERATED CONTENT:
"This text was created using AI."
"AI-generated image"

DEEPFAKE:
"This content was created or modified using AI."

Minimal Risk

Systems with Minimal Obligations

System	Examples
Spam filters	Email spam detection
Internal analytics	Business intelligence, dashboards
Recommendation	Product recommendations (no impact on rights)
Search	Internal search, document search
Automation	Workflow automation, scheduling

Recommended (Not Mandatory) Practices

Basic documentation
Internal ownership
Basic monitoring
Incident response

Classification Decision Tree

Classification Table

For each AI system from the inventory:

AI ID	Name	Annex III?	Safety component?	User-facing?	Classification	Note
AI-001		[ ] Yes [ ] No	[ ] Yes [ ] No	[ ] Yes [ ] No
AI-002		[ ] Yes [ ] No	[ ] Yes [ ] No	[ ] Yes [ ] No

Next Steps

After classification:

Prohibited -> Immediately deactivate, consult Legal
High-Risk -> DPIA + Full compliance
Transparency -> Implement notifications
Minimal -> Basic documentation