AI Act Risk Levels: Banned, High-Risk, Transparency, Minimal

The risk pyramid

The more a use of AI can affect people's rights and safety, the stricter the rules. The same model can land in different levels depending on how it is used: summarising emails is minimal risk, filtering candidates' CVs is high risk.

1. Unacceptable risk: the banned practices

Article 5 bans, among others, the following practices (in short, with exceptions provided in the text). The bans have applied since 2 February 2025; the last two on the list were added by the Digital Omnibus (Regulation (EU) 2026/1744) and apply from 2 December 2026.

• Subliminal or manipulative techniques that cause harm.
• Exploiting vulnerabilities linked to age, disability or socio-economic situation.
• Social scoring that leads to unjustified unfavourable treatment.
• Untargeted scraping of facial images from the internet or CCTV to build facial recognition databases.
• Emotion recognition in the workplace and in schools, except for medical or safety reasons.
• Biometric categorisation to infer sensitive data such as ethnic origin or political opinions.
• Real-time remote biometric identification in public spaces by law enforcement, except in limited cases.
• Assessing a person's risk of committing crime based solely on profiling.
• Generating realistic intimate images, video or audio of an identifiable person without their consent (so-called "nudify" apps).
• Generating child sexual abuse material.

2. High risk

A system is high risk in two cases: if it is a safety component of an already regulated product (toys, medical devices, machinery, Annex I) or if it falls under one of the Annex III areas:

• Biometrics.
• Critical infrastructure.
• Education and training.
• Employment and workers' management, such as CV screening.
• Access to essential services, such as credit scoring or insurance risk assessment.
• Law enforcement, migration and border control.
• Justice and democratic processes.

These systems must have risk management, data quality, technical documentation, event logging, transparency towards users, human oversight, accuracy and cybersecurity, plus a conformity assessment. The Digital Omnibus postponed these deadlines to 2 December 2027 for Annex III and 2 August 2028 for Annex I: always check the consolidated text.

3. Limited risk: transparency obligations

• Chatbots and assistants: people interacting must know they are talking to an AI.
• Generated or manipulated content, such as deepfakes, must be disclosed as such.
• Emotion recognition systems: the people involved must be informed.

These obligations apply from 2 August 2026. Generative systems already on the market before that date have until 2 December 2026 to mark their output as artificially generated.

4. Minimal risk

Spam filters, video games and simple recommendation systems have no specific obligations. Other laws, such as the GDPR, still apply, and voluntary codes of conduct can be adopted.

A special case: general-purpose AI models

The large models that power assistants such as ChatGPT or Gemini are treated separately. Their providers have documentation and copyright obligations, with additional requirements for models with systemic risk, since 2 August 2025.

How to tell which level you fall into

• What does the system do and what do you use it for?
• In what context: employment, credit, education, health, security?
• Who is affected by its output and how much can it affect the person?
• Does it fall under an Annex III area or a regulated product?