The EU AI Act regulates AI according to how much harm it could cause. The same technology can land in different categories depending on what it’s used for: a language model writing marketing copy is minimal risk, while the same model ranking job applicants is high risk.
Here are the four categories, with examples and how to classify your own systems.
1. Unacceptable risk: banned
These practices are prohibited because they’re considered a clear threat to people’s safety or rights. They have been banned since 2 February 2025. Examples include:
- Social scoring that leads to unfair treatment.
- AI that manipulates people or exploits their vulnerabilities in ways that cause significant harm.
- Emotion recognition in workplaces and schools, except for medical or safety reasons.
- Untargeted scraping of facial images to build facial recognition databases.
From 2 December 2026, two more bans apply: AI that creates sexual deepfakes of identifiable people without their explicit consent, and AI that generates child sexual abuse material.
See the full list in our guide to EU AI Act prohibited practices.
2. High risk: allowed with strict requirements
High-risk AI systems can be used, but only if they meet detailed requirements. There are two routes into this category.
Route 1: AI in regulated products (Annex I)
AI that is a safety component of, or is itself, a product already covered by EU safety laws, such as medical devices, machinery, toys, lifts or vehicles, when that product needs a third-party conformity assessment.
Route 2: AI used in sensitive areas (Annex III)
AI used in these areas is high risk:
- Biometrics, such as remote biometric identification and emotion recognition where it’s allowed.
- Critical infrastructure, such as safety components in energy, water or traffic.
- Education, such as deciding admissions, grading or detecting cheating.
- Employment, such as screening CVs, ranking candidates, deciding promotions or monitoring workers.
- Access to essential services, such as credit scoring, life and health insurance pricing, public benefits and emergency call triage.
- Law enforcement, migration and border control, and the administration of justice and democratic processes.
There’s an important exception. An Annex III system is not high risk if it doesn’t pose a significant risk of harm, for example if it only performs a narrow procedural task or improves the result of a human activity already completed. AI that profiles people is always high risk. If you rely on the exception, you must document your reasoning.
What high risk means in practice
Providers must have risk management, data governance, technical documentation, logging, transparency, human oversight, and accuracy, robustness and cybersecurity controls, plus a quality management system and a conformity assessment. Deployers must follow the instructions for use, assign competent human oversight, monitor the system, keep logs and inform affected people.
These duties apply from 2 December 2027 for Annex III systems and from 2 August 2028 for Annex I products. Both dates were pushed back by the Digital Omnibus on AI, which entered into force on 27 July 2026.
3. Transparency risk: tell people
Some AI is allowed with light obligations, but people must know they’re dealing with AI:
- Chatbots must tell users they’re interacting with an AI, unless it’s obvious.
- Deepfakes, meaning AI-generated or manipulated images, audio or video of real people, places or events, must be disclosed.
- AI-generated text published to inform the public on matters of public interest must be disclosed, unless it has been human-reviewed and someone takes editorial responsibility.
- Providers of generative AI must mark outputs in a machine-readable way so they can be detected as AI-generated.
These duties have applied since 2 August 2026. One exception: providers whose generative AI was already on the market before that date have until 2 December 2026 to add the machine-readable marking.
4. Minimal risk: no new obligations
Spam filters, AI in video games, inventory forecasting and writing assistants used internally all typically fall here. The Act adds no new legal duties, though the AI literacy duty still applies to any organisation whose staff use AI at work.
General-purpose AI models: a separate track
Large models that can do many tasks, such as the models behind popular chatbots, have their own rules. Their providers must keep technical documentation, share information with the businesses building on them, respect EU copyright law and publish a summary of training content. Models with “systemic risk” have extra duties, including evaluations and serious incident reporting.
How to classify your AI systems
For each system in your AI inventory, ask in order:
- Is it a banned practice? If yes, stop using it.
- Is it covered by Annex I or listed in Annex III? If yes, it’s probably high risk, so check whether the exception applies and document your answer.
- Does it interact with people, generate content, or recognise emotions or categorise people from biometric data? If yes, transparency duties apply.
- Otherwise it’s minimal risk. Record that, and review it if the use changes.
Official sources
- EU AI Act, Regulation (EU) 2024/1689: the official text on EUR-Lex
- Digital Omnibus on AI, Regulation (EU) 2026/1744: the July 2026 amendment that moved the high-risk dates
- AI Act Service Desk: the European Commission’s AI Act explorer, compliance checker and FAQs
More primary sources, with a short note on each, are on our Sources page.
Last reviewed 25 October 2026. Based on Regulation (EU) 2024/1689, including Articles 5, 6 and 50 and Annexes I and III, as amended by Regulation (EU) 2026/1744 (Digital Omnibus on AI). This article is general information, not legal advice.