EU AI Act risk categories: the overview

At the top of the pyramid are practices the legislator considers so dangerous that no obligation, however strict, could offset them. Article 5 has banned them outright since 2 February 2025: they may not be placed on the market, put into service or used. The ban applies to providers and deployers alike; there is no transition period and no grandfathering.

What matters is the use, not the technology. Facial recognition is not banned as such, but untargeted scraping of facial images from the internet or CCTV footage to build databases is. Emotion recognition is not banned across the board, but it is in the workplace and in education – with exceptions for medical and safety reasons. Outside those settings it falls under the transparency obligations.

The Digital Omnibus (Regulation (EU) 2026/1744) extends the list: from 2 December 2026, AI that generates or manipulates intimate images of identifiable people without their consent is also banned, as is AI used to generate child sexual abuse material. For providers of generative image and video tools, safeguards need to be in place by then.

For most companies the ban is not the core issue, but it is a mandatory check. If you use software that analyses the mood of call-centre staff, for example, that feature must be switched off – even if it is just a side feature of a larger tool. So check every system against Article 5 first, before you think about the other categories.

High-risk AI is the category with the most obligations. Article 6 provides two ways in. First, Annex I: the AI system is itself a product, or the safety component of a product, covered by EU harmonisation legislation – such as machinery, toys, lifts or medical devices – and requires a third-party conformity assessment. After the Digital Omnibus, these obligations apply from 2 August 2028.

Second, Annex III: AI systems used in one of eight sensitive areas – biometrics, critical infrastructure, education and vocational training, employment and workers management, access to essential private and public services (including creditworthiness assessment and risk assessment and pricing in life and health insurance), law enforcement, migration and border control, and the administration of justice and democratic processes. For these systems the obligations apply from 2 December 2027.

Not every system in an Annex III area is automatically high-risk. Article 6(3) exempts systems that pose no significant risk because they only perform a narrow procedural task, improve the result of a previously completed human activity, detect decision-making patterns without replacing human assessment, or perform a preparatory task. A system that profiles natural persons is always high-risk. Anyone relying on the exemption must document the assessment (Art. 6(4)) and register the system in the EU database (Art. 49(2)). The Digital Omnibus has slimmed down the required entries: the summary of grounds and the list of Member States no longer have to be registered (Annex VIII, Section B, points 7 and 9 deleted by Art. 1(42) of Regulation (EU) 2026/1744; Recital 22). The documentation duty remains.

Commission guidelines: Article 6(5) requires the Commission to provide guidelines with practical examples of systems that are and are not high-risk. It published a draft on 19 May 2026. The targeted consultation was extended to 23 July 2026, and the Commission plans to adopt the final version by the end of 2026. As of 7 October 2026 there is no final version. The draft is guidance, not binding (para. 6).

Key points of the draft: human involvement does not change the classification under Article 6(2) (paras. 70–71). The Article 6(3) exemptions are exhaustive and must be read narrowly, and profiling always rules them out (paras. 88–89). Software marketed broadly that excludes high-risk uses only in its terms of service stays high-risk (para. 12). Several components that together shape an individual decision are assessed as one system, agentic systems included (para. 75).

The obligations are split by role. Providers carry most of the load: risk management (Art. 9), data governance (Art. 10), technical documentation (Art. 11), record-keeping (Art. 12), transparency and instructions for use (Art. 13), human oversight (Art. 14), accuracy, robustness and cybersecurity (Art. 15), quality management (Art. 17), conformity assessment (Art. 43) and registration in the EU database (Art. 49). Deployers must use the system according to the instructions, assign human oversight to competent people, keep an eye on input data and logs, and inform the people affected (Art. 26). Public bodies and certain private deployers, for example in credit scoring or insurance pricing, also carry out a fundamental rights impact assessment (Art. 27).

This overview keeps high-risk deliberately short. All eight Annex III areas with examples, borderline cases and a step-by-step exemption check are in our Annex III guide.

Many overviews call the third category “limited risk”. The term does not appear in the Regulation, but it has become the usual shorthand for the transparency obligations in Article 50. It covers AI systems whose main risk is that people do not realise they are talking to a machine or looking at artificially generated content.

Article 50 covers four cases. Providers must design systems that interact directly with people so that those people are informed (para. 1), and must mark synthetic audio, image, video and text output in a machine-readable way (para. 2). Deployers must inform people when they use emotion recognition or biometric categorisation (para. 3), and disclose that deepfakes or AI-generated text on matters of public interest were artificially created (para. 4). Paragraph 5 requires the information to be clear, timely and accessible.

The obligations have applied since 2 August 2026. For machine-readable marking by systems placed on the market before that date, the Digital Omnibus introduced a transition period until 2 December 2026. There is no transition period for chatbot notices, emotion recognition or deepfakes.

Important: Article 50 is not an alternative to high-risk, but an extra layer. A chatbot that shortlists applicants is high-risk and must also disclose that it is an AI. Conversely, a notice does not help where a ban applies: emotion recognition in the workplace stays prohibited even if staff are informed.

The broad base of the pyramid consists of AI systems with minimal risk: spam filters, product recommendations in online shops, predictive maintenance in production, spell checking or inventory optimisation. The AI Act imposes no specific obligations on them. In practice, many of the AI applications companies use land here.

“No specific obligations” does not mean “nothing to do”, though. Since 2 February 2025, Article 4 has required providers and deployers of all AI systems – whatever their category – to take measures to support AI literacy. The Digital Omnibus rewrote the provision: companies take measures to support the AI literacy of the people using AI on their behalf. They do not have to guarantee a particular level of competence.

Providers and deployers can also sign up to voluntary codes of conduct under Article 95, for example to apply high-risk requirements voluntarily. And because almost every AI processes personal data, the GDPR applies in full – regardless of how the AI Act classifies the system.

What matters is the rationale: a minimal-risk classification should be documented, with the intended use and a note on why neither Article 5, Annex I or III nor Article 50 applies. If the purpose changes, the classification needs another look. This documentation is what is most often missing in audits.

Large language, image and multimodal models do not fit the pyramid, because they have no fixed intended use. The AI Act therefore regulates them in a separate Chapter V. The obligations fall on the provider of the model – the company that develops such a model and places it on the market – and have applied since 2 August 2025.

Every GPAI provider must keep technical documentation, give downstream providers who build the model into their own AI systems the information they need, implement a policy to comply with copyright law and publish a summary of the content used for training. Models with systemic risk – the most capable models, identified using the criteria in Article 51 – must additionally undergo model evaluation, assess and mitigate systemic risks, report serious incidents and ensure adequate cybersecurity.

Models placed on the market before 2 August 2025 have until 2 August 2027. Since 2 August 2026, the Commission has been able to fine GPAI providers; unlike AI systems, supervision here sits at EU level.

For most companies GPAI is a supply-chain question rather than a set of their own obligations: if you use ChatGPT or a comparable model via an API, you are usually not the provider of the model. The risk category of your own system then depends on its intended use. If you substantially modify a model and offer it under your own name, however, check whether you take on Chapter V obligations yourself.