---
title: "EU AI Act risk categories: overview with examples | AI-Konform"
description: "EU AI Act risk categories at a glance: prohibited AI, high risk, transparency obligations and minimal risk, with examples, obligations and dates."
url: https://get-ai-conform.com/en/eu-ai-act-risikoklassen
language: en
updated: 2026-11-10
source_revision: 1a1303c5ce70
maintainer: "AI-Konform, New Life Digital UG (haftungsbeschränkt), support@get-ai-conform.com"
---

# EU AI Act risk categories: overview with examples and obligations

The EU AI Act (Regulation (EU) 2024/1689) divides AI into four levels according to its risk: prohibited practices (Art. 5), high-risk AI (Art. 6 with Annexes I and III), limited risk with transparency obligations (Art. 50) and minimal risk without specific obligations. The term “risk category” does not appear as such in the Regulation; it is the common name for these levels. General-purpose AI models (GPAI) have rules of their own in Chapter V. Regardless of the level, Art. 4 applies to providers and deployers: measures on AI literacy.

[Simon Maiwald](https://www.linkedin.com/in/simonmaiwald/) · Founder of New Life Digital, former lecturer at DHBW · Last updated: November 10, 2026 · Sources: [Regulation (EU) 2024/1689](https://eur-lex.europa.eu/eli/reg/2024/1689/2026-07-27/eng), [Regulation (EU) 2026/1744](https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng)

## The risk pyramid of the EU AI Act

The higher the level, the stricter the rules. The graphic shows the four levels with their legal basis, next to them the separate rules for GPAI models and, below, the obligation that applies at every level.

Source: AI-Konform, based on Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744, as of 10 November 2026. [Open the graphic as SVG (German version)](https://get-ai-conform.com/images/eu-ai-act-risikoklassen.svg): for embedding with the source credit get-ai-conform.com.

One obligation applies in every risk category

You need measures on AI literacy under Art. 4 regardless of the classification. The AI-Konform training is one such measure: in 26 languages, with a certificate of participation per person and an audit report for the company.

from €24.69 per person, net of VAT (no subscription)

## Overview: risk categories, obligations and dates

The table summarises what applies at which level and since when. For the Regulation as a whole, with roles and penalties, see the page [The EU AI Act at a glance](https://get-ai-conform.com/en/eu-ai-act).

EU AI Act risk categories: legal basis, examples, obligations and dates
| Level | Legal basis | Examples | Obligations | Since / from |
| --- | --- | --- | --- | --- |
| Unacceptable risk | Art. 5 | Social scoring, emotion recognition in the workplace, manipulative techniques | prohibited, for providers and deployers | since 2 February 2025; two further prohibitions from 2 December 2026 |
| High risk | Art. 6 with Annexes I and III | Pre-screening of job applicants, creditworthiness assessment, AI as a safety component of a product | Providers: risk management, data quality, documentation, conformity assessment, registration. Deployers: use in line with the instructions for use, human oversight, keeping logs (Art. 26) | Annex III from 2 December 2027, Annex I from 2 August 2028 |
| Limited risk | Art. 50 | Chatbots, deepfakes, AI-generated text on matters of public interest | Notice that an AI is involved, labelling of content | since 2 August 2026 |
| Minimal risk | no specific provision | Spam filters, translation aids, draft texts in the office | no obligations under Art. 5, 6 or 50; voluntary codes of conduct (Art. 95) | no separate date |
| GPAI models (separate rules) | Chapter V | large language models behind chatbots and writing assistants | Providers of the model: technical documentation, copyright policy, summary of the training content; more in the case of systemic risk | since 2 August 2025, existing models by 2 August 2027 |
| All levels: AI literacy | Art. 4 | every AI system used professionally | measures for the AI literacy of the people involved | since 2 February 2025, new wording since 27 July 2026 |

One system can trigger several obligations at once, for example as high-risk AI that also talks to people. What matters is the intended purpose, not the tool: the same program can be minimal risk in the office and high risk in recruitment.

## Which AI practices are prohibited?

[Art. 5](https://eur-lex.europa.eu/eli/reg/2024/1689/2026-07-27/eng#art_5) prohibits certain practices altogether. The prohibition covers placing on the market, putting into service and use, so it also applies to companies that merely use such a system. The following have been prohibited since 2 February 2025:

- **Manipulation:** subliminal, purposefully manipulative or deceptive techniques that materially distort people’s behaviour and cause them significant harm or are reasonably likely to do so (point (a)).
- **Exploiting vulnerabilities:** AI that exploits age, a disability or a specific social or economic situation, with the same consequence (point (b)).
- **Social scoring:** evaluating people over a period of time on the basis of their social behaviour or their characteristics, where this leads to unjustified or out-of-context detrimental treatment (point (c)).
- **Predicting criminal offences** based solely on profiling or personality traits (point (d)).
- **Facial recognition databases:** creating or expanding them through untargeted scraping of facial images from the internet or from CCTV footage (point (e)).
- **Emotion recognition in the workplace and in education institutions**, except for medical or safety reasons (point (f)).
- **Biometric categorisation** to infer sensitive characteristics such as political opinions, trade union membership, religion or sexual orientation (point (g)).
- **Real-time remote biometric identification** in publicly accessible spaces for law enforcement, with narrow exceptions (point (h)).

From 2 December 2026, two prohibitions are added, introduced by the [Digital Omnibus on AI](https://get-ai-conform.com/en/digital-omnibus-ki): AI that generates or manipulates non-consensual intimate depictions of identifiable persons, and AI that generates or manipulates depictions of child sexual abuse (points (ba) and (bb)). Companies that use image generators only infringe the prohibition if they use the system for that purpose (Art. 5(1a), point (b)). Infringements of Art. 5 fall into the highest fine tier: up to €35 million or 7 % of global annual turnover (Art. 99(3)).

## When is an AI system high-risk AI?

Under [Art. 6](https://eur-lex.europa.eu/eli/reg/2024/1689/2026-07-27/eng#art_6) there are two routes into the high-risk level. First, AI as a safety component of a product, or as the product itself, that is covered by EU product legislation listed in Annex I and must undergo a third-party conformity assessment, for example medical devices, toys or lifts. These obligations apply from 2 August 2028. Second, AI in one of the eight areas of Annex III, from 2 December 2027:

1. **Biometrics:** remote identification, categorisation according to sensitive characteristics, emotion recognition (in so far as not already prohibited).
2. **Critical infrastructure:** safety components for digital infrastructure, road traffic and the supply of water, gas, heating and electricity.
3. **Education:** admission, evaluation of learning outcomes, assessment of the appropriate level of education, monitoring during tests.
4. **Employment:** targeted job advertisements, screening and evaluation of applications, decisions on promotion or termination, task allocation, performance monitoring.
5. **Essential services:** public assistance benefits, creditworthiness of natural persons (fraud detection excepted), risk and pricing in life and health insurance, emergency calls and triage.
6. **Law enforcement**, in so far as permitted under Union or national law.
7. **Migration, asylum and border control**, likewise only within the permitted framework.
8. **Administration of justice and democratic processes:** assisting courts, influencing elections.

Not every system listed in Annex III is automatically high-risk. Under Art. 6(3) the classification does not apply if the system poses no significant risk, for instance because it only performs a narrow procedural task, improves the result of a previously completed human activity, detects decision-making patterns without replacing the human assessment, or performs a preparatory task. If the system performs profiling of natural persons, it is always considered high-risk. Anyone relying on the exception documents their assessment before placing the system on the market (Art. 6(4)) and registers the system (Art. 49(2)).

Most obligations fall on providers: risk management, data quality, technical documentation, record-keeping, transparency, human oversight, accuracy and robustness (Art. 9 to 15), plus conformity assessment and registration. Companies that merely use high-risk AI are [deployers](https://get-ai-conform.com/en/ki-betreiber) and have obligations of their own under Art. 26: use the system in line with the instructions for use, assign human oversight to persons with the necessary competence, monitor its operation, keep the logs and, before use in the workplace, inform the workers and their representatives (Art. 26(7)). Certain deployers also carry out a fundamental rights impact assessment (Art. 27), for example public bodies and companies that use AI for creditworthiness assessment or for risk and pricing in life and health insurance. Infringements cost up to €15 million or 3 % of global annual turnover (Art. 99(4)).

For high-risk systems that were placed on the market or put into service before these obligations started to apply, the Regulation applies only if their design is significantly changed afterwards. Systems intended for use by public authorities must in any case be brought into line by 2 August 2030 (Art. 111(2)).

## What applies at limited risk?

For AI that interacts with people or generates content, the transparency obligations under [Art. 50](https://eur-lex.europa.eu/eli/reg/2024/1689/2026-07-27/eng#art_50) have applied since 2 August 2026:

- **Paragraph 1, providers:** systems that speak or write with people are designed so that people know they are interacting with an AI, unless this is obvious.
- **Paragraph 2, providers:** AI-generated images, videos, audio and text are marked in a machine-readable format. For systems placed on the market before 2 August 2026, a transition until 2 December 2026 applies (Art. 111(4)).
- **Paragraph 3, deployers:** anyone who uses emotion recognition or biometric categorisation informs the persons concerned.
- **Paragraph 4, deployers:** deepfakes, and AI-generated text intended to inform the public on matters of public interest, are disclosed; for text this does not apply if a person reviews it editorially and takes responsibility for it.

Typical cases and what the notice can look like are explained on the page [AI labelling obligation under Art. 50](https://get-ai-conform.com/en/ki-kennzeichnung). Infringements likewise fall into the tier of up to €15 million or 3 % (Art. 99(4)).

## What applies at minimal risk?

All other AI systems are minimal risk: no obligations under Art. 5, 6 or 50 apply to them. This includes many everyday office tools, such as spam filters, translation and spelling aids, or draft texts that employees check before using them. Voluntary codes of conduct are possible (Art. 95). One obligation remains: measures on [AI literacy under Art. 4](https://get-ai-conform.com/en/ki-kompetenz-pflicht), because it does not depend on the risk level. How to record these measures is shown on the page [AI literacy obligation](https://get-ai-conform.com/en/ki-kompetenz-pflicht). The AI Act does not provide for a specific fine for Art. 4; more on this on the [fines page](https://get-ai-conform.com/en/bussgelder).

## Are GPAI models a risk category of their own?

Some overviews list them as a fifth category. Legally, Chapter V governs the providers of general-purpose AI models, such as the large language models behind chatbots. Their obligations have applied since 2 August 2025; for models placed on the market before that date, providers have until 2 August 2027 (Art. 111(3)). Fines imposed by the Commission under Art. 101 have been possible since 2 August 2026. A company that merely uses a chatbot based on such a model is not the provider of the model. Its use is governed by the level of the specific purpose, and Art. 4 always applies.

## Typical AI applications and their classification

The examples show the usual classification. What counts is always the specific purpose and your role as provider or deployer.

Typical AI applications in a company and their classification under the EU AI Act
| Application | Usual level | What your company should do |
| --- | --- | --- |
| ChatGPT or Copilot for text and research in the office | minimal risk for your use; the model behind it is governed by Chapter V, at the provider | Measures on AI literacy (Art. 4), rules in an [AI policy](https://get-ai-conform.com/en/ki-richtlinie-vorlage) |
| Chatbot on the website | limited risk (Art. 50(1)) | The notice that an AI is responding is the provider’s obligation (Art. 50(1)); as deployer, make sure it appears; anyone who has the bot developed under their own name can be a provider themselves |
| AI pre-screening of job applications | high risk (Annex III, point 4) | from 2 December 2027 deployer obligations under Art. 26: inform applicants (Art. 26(11)); when used in relation to workers, inform their representatives in advance (Art. 26(7)) |
| Creditworthiness assessment of private customers | high risk (Annex III, point 5), fraud detection excepted | Deployer obligations under Art. 26 and fundamental rights impact assessment (Art. 27) |
| Emotion recognition for employees | prohibited (Art. 5(1), point (f)), except for medical or safety reasons | do not use |
| Spam filter | minimal risk | Measures on AI literacy (Art. 4) |
| Image generator for advertising images and graphics | minimal risk; limited risk as soon as deepfakes are created | Disclose images that look like authentic recordings of real persons, objects, places or events (Art. 50(4)) |
| AI diagnostic software in a medical device | high risk (Annex I), if a third party must assess conformity | Provider obligations from 2 August 2028; as deployer, obligations under Art. 26 |

## Determining the risk category: four questions

Go through these questions in order for every AI system in your company. A list of all the tools in use, with their purpose, is the best starting point.

1. **Does the purpose fall under a prohibition in Art. 5?** Then the use is not allowed, regardless of all further questions.
2. **Is the system part of a product covered by Annex I, or is it used in an area listed in Annex III?** Then it is usually high-risk AI; for Annex III, check the exception under Art. 6(3).
3. **Does the system talk to people, or does it generate images, videos, audio or text?** Then the transparency obligations under Art. 50 apply in addition.
4. **What role does your company have?** Anyone who develops a system or offers it under their own name is a provider; anyone who uses it professionally is a deployer. Anyone who uses an AI system, including a general-purpose AI system such as ChatGPT, for a purpose listed in Annex III can themselves become the provider of a high-risk system under Art. 25(1), point (c). The specific obligations depend on this.

If none of the first three questions applies, the system is minimal risk. The dates for all levels are in the [EU AI Act deadline calendar](https://get-ai-conform.com/en/fristen).

## Frequently asked questions about the risk categories

### How many risk categories are there in the EU AI Act?

The usual division is into four levels: unacceptable risk (prohibited), high risk, limited risk and minimal risk. Some overviews count general-purpose AI models (GPAI) as a fifth category; the Regulation governs them in a chapter of their own. The Regulation itself does not use the term “risk category” for this division.

### Which risk category does ChatGPT fall into?

That depends on the use, not on the tool. The language model behind ChatGPT is a general-purpose AI model; the obligations for it lie with the provider of the model (Chapter V). If your company uses ChatGPT for text or research, that is usually minimal risk. If you use it, for example, to pre-screen job applications, the use can fall under Annex III, with much stricter obligations. Anyone who uses an AI system, including a general-purpose AI system such as ChatGPT, for a purpose listed in Annex III can themselves become the provider of a high-risk system under Art. 25(1), point (c). Art. 4 applies in every case.

### Is a chatbot high-risk AI?

As a rule, no. A chatbot for customer enquiries falls under the transparency obligations of Art. 50(1): people must be able to tell that they are talking to an AI, unless this is obvious. It only becomes high-risk AI if it is used in an area listed in Annex III, for example for decisions on job applications or loans. More on the page [AI labelling under Art. 50](https://get-ai-conform.com/en/ki-kennzeichnung).

### Who classifies an AI system?

The Regulation does not provide for an official classification. The provider classifies its system; if it considers that a system listed in Annex III is not high-risk under Art. 6(3), it documents this before placing the system on the market and registers the system (Art. 6(4), Art. 49(2)). Deployers check whether their own intended use falls into an area listed in Annex III. Market surveillance can review the classification (Art. 80); this is done by the market surveillance authorities that each Member State designates (Art. 70).

### Does the AI literacy obligation also apply at minimal risk?

Yes. Art. 4 is not tied to the risk level: providers and deployers take measures to support the AI literacy of the people who work with AI on their behalf, even if only simple tools are in use. What that means in practice is explained on the page [AI literacy obligation under Art. 4](https://get-ai-conform.com/en/ki-kompetenz-pflicht).

### From when do the obligations for high-risk AI apply?

For AI in the areas listed in Annex III from 2 December 2027, for AI in products covered by Annex I from 2 August 2028. The Digital Omnibus on AI (Regulation (EU) 2026/1744) postponed these dates; the original dates were 2 August 2026 and 2 August 2027. For systems placed on the market or put into service before then, the obligations apply only after a significant change in their design; systems intended for use by public authorities must be brought into line by 2 August 2030 (Art. 111(2)). All dates are in the [deadline calendar](https://get-ai-conform.com/en/fristen).

Note: this overview is for general information only and does not constitute legal advice. Source: [Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744, EUR-Lex](https://eur-lex.europa.eu/eli/reg/2024/1689/2026-07-27/eng).

---

Canonical page: https://get-ai-conform.com/en/eu-ai-act-risikoklassen
All Markdown pages: https://get-ai-conform.com/sitemap.md
