EU AI Act — risk classification

Unacceptable risk (Article 5 — prohibited)

These systems are banned across the EU. Listed at Article 5.

• Subliminal manipulation or deception that causes harm.
• Exploitation of vulnerabilities (age, disability, social/economic situation) causing harm.
• Social scoring by public authorities (with limited exceptions).
• Real-time remote biometric identification in public spaces for law enforcement (limited exceptions).
• Predictive policing based solely on profiling or personality traits.
• Untargeted scraping of facial images to build recognition databases.
• Inferring emotions in workplace and education (limited exceptions).
• Biometric categorisation that infers sensitive attributes.

High risk (Annex III — strict)

Listed in Annex III. Strict requirements: data governance, documentation, transparency, human oversight, accuracy, robustness, cybersecurity, conformity assessment.

See High-Risk walk for the eight categories.

Limited risk (Article 50 — transparency)

Interaction transparency: users must know they are talking to an AI system.

• Chatbots / conversational agents.
• Emotion recognition systems and biometric categorisation.
• Deepfake generation / manipulation.
• Synthetic content (text, audio, image, video) generated by AI.
• AI-generated content being published: marking requirement.

CSOAI's Article 50 Passport is a free signed record for the transparency obligation.

Minimal risk (no specific obligations)

Spam filters, AI in video games, basic inventory optimisers. No specific obligations in the Act. Voluntary codes of conduct encouraged.

Where CSOAI measures

CSOAI publishes measurement harnesses for:

• Article 50 transparency (the Article 50 Passport flow).
• Article 50 provenance survival (ProvBench).
• Closed-book statutory citation (the AI Act frozen-split harness).
• Article 43 internal-control self-assessment evidence.

What we do not measure: the legal classification of "is this system Annex III?". That is a regulator's call, not a measurement body's.