EU AI Act Risk Categories: Prohibited, High-Risk, Transparency and Minimal Risk

The EU AI Act risk categories are four tiers that the Commission uses to describe the act's risk-based approach: unacceptable risk, high risk, transparency risk and minimal or no risk. General-purpose AI models follow a separate track in Chapter V rather than a fifth tier. The four tiers overlap, so one AI system can carry duties from two of them at once.

The EU AI Act is Regulation (EU) 2024/1689, in force since 1 August 2024 and amended by Regulation (EU) 2026/1744 of 8 July 2026, in force since 27 July 2026, which moved the high-risk dates. This page sets out each tier's provision, duties, dates and maximum fines as of 4 November 2026.

Compiled by the AIActGuide.net Editorial Team from the Official Journal text; how we check every statement: methodology.

What Are the EU AI Act Risk Categories?

An EU AI Act risk category is the tier into which an AI system falls according to the provision that applies to it: Article 5, Article 6 with Annexes I and III, Article 50, or none of them. The Commission's AI Act page names the same four levels as of 23 October 2026: unacceptable risk (prohibited), high risk, transparency risk and minimal or no risk.

Commission Labels, Not Defined Terms of the Regulation

The four tier names come from the Commission, not from Article 3: the Guidelines on prohibited AI practices, C(2025) 5052 final of 29 July 2025, list them in paragraph 2. Article 3 defines no "risk category", no "unacceptable risk" and no "limited risk". The label "limited risk" is the common name for the third tier, which the Commission calls "transparency risk".

What Sits Outside the Four Tiers

Only an AI system within Article 3(1) and within the scope of Article 2 can fall into a risk category; general-purpose AI models follow their own rules in Chapter V. The AI system definition of Article 3(1) applies since 2 February 2025, and the Commission explained it in guidelines C(2025) 5053 final of 29 July 2025.

Article 2 also excludes, among others, AI systems used exclusively for military, defence or national security purposes (Article 2(3)) and AI systems and models specifically developed and put into service for the sole purpose of scientific research and development (Article 2(6)). The guide to who the AI Act applies to sets out the Article 2 scope and exclusions. The open-source exclusion of Article 2(12) is narrower than it looks: it does not cover systems placed on the market as high-risk, under Article 5 or under Article 50.

The Four Tiers at a Glance: Legal Basis, Duties, Dates and Fines

Each tier rests on a different provision, binds different operators and applies from a different date. The four tiers and the separate model track compare on six points:

Labels: Guidelines C(2025) 5052, para. 2; provisions as amended by Regulation (EU) 2026/1744.

Tier 1: Unacceptable Risk and the Article 5 Bans

The unacceptable-risk tier covers the AI practices that Article 5 prohibits outright, whatever the operator's size or the system's placing date. Article 5(1), first subparagraph, holds ten lettered points: (a), (b), (ba), (bb) and (c) to (h). The Commission's AI Act page says "The AI Act prohibits nine practices" (last update 19 October 2026), because it counts points (ba) and (bb) as one item.

Points (a), (b) and (c) to (h) apply since 2 February 2025, while points (ba) and (bb) and Article 5(1a) and (1b) apply from 2 December 2026 (Article 113, third paragraph, point (a), as amended). The two new bans therefore apply from AI Act 2 December 2026, not from February 2025. Point (ba) bans AI that generates or manipulates realistic intimate images or sexually explicit depictions of an identifiable person without explicit consent, and point (bb) bans AI that generates or manipulates child sexual abuse material.

No system is grandfathered: Article 111(1) and (2) apply without prejudice to Article 5, so Article 5 reaches every AI system whatever its placing date. Breaches fall in the highest fine tier of Article 99(3), which has applied since 2 August 2025; between 2 February and 1 August 2025 points (a), (b) and (c) to (h) had direct effect without fines (Guidelines C(2025) 5052, paragraphs 429 to 431).

Tier 2: High Risk under Article 6, Annex I and Annex III

An AI system is high-risk when it is a safety component of, or itself, an Annex I product that needs third-party conformity assessment (Article 6(1)), or when its intended purpose matches an Annex III use case (Article 6(2)). High-risk classification is not a ban: Article 6 classifies, Chapter III attaches duties, and only Article 5 prohibits. The two routes are distinct:

• Product route (Article 6(1)): Annex I product or safety component, plus third-party conformity assessment; both conditions are cumulative.
• Use-case route (Article 6(2)): one of the 25 use cases in the eight Annex III areas, subject to Article 6(3).

Annex III is unchanged by Regulation (EU) 2026/1744. Under Article 6(3), an Annex III system is not high-risk when it poses no significant risk of harm to the health, safety or fundamental rights of natural persons and meets any one of four alternative conditions, points (a) to (d), but a system that performs profiling of natural persons is always high-risk. For Annex I Section B products, only Article 6(1), Article 60a and Articles 102 to 112 apply (Article 2(2) as replaced).

The Digital Omnibus on AI moved the high-risk dates: before 27 July 2026, Chapter III Sections 1 to 3 were due on 2 August 2026 for Annex III systems and on 2 August 2027 for Annex I systems. The Commission guidelines on classification under Article 6(5) exist as a draft of 19 May 2026, and the Commission announced on 19 May 2026 that the final guidelines will be adopted by the end of 2026.

Tier 3: Transparency Risk, Often Called Limited Risk (Article 50)

The transparency tier covers AI systems whose function makes disclosure necessary: systems that interact with people, generate synthetic content, recognise emotions or categorise biometrically, or produce deep fakes. The four duties follow functions, not sectors:

• Providers, Article 50(1): inform people that they are interacting with an AI system, unless this is obvious to a reasonably well-informed, observant and circumspect person, and except for systems authorised by law to detect, prevent, investigate or prosecute criminal offences.
• Providers, Article 50(2): mark synthetic audio, image, video or text output in a machine-readable format and make it detectable as artificially generated or manipulated, subject to the Article 50(2) exceptions.
• Deployers, Article 50(3): inform the persons exposed to an emotion recognition or biometric categorisation system.
• Deployers, Article 50(4): disclose deep fakes and AI-generated or manipulated text published to inform the public on matters of public interest, subject to the exceptions Article 50(4) sets out.

Commission guidelines on Article 50, whose content was approved by C(2026) 5054 final and published on 20 July 2026, are non-binding, and their formal adoption is pending. They place recommender systems, spam filters, automated translation or transcription tools and auto-completion outside Article 50(1) (paragraph 31). The Code of Practice on Transparency of AI-generated Content, final since 10 June 2026, is voluntary and gives no presumption of conformity.

Article 50 applies since 2 August 2026 (Article 113, second paragraph), a date Regulation (EU) 2026/1744 left unchanged. The general application date is covered in AI Act 2 August 2026. The only transition is Article 111(4): systems generating synthetic content placed on the market before 2 August 2026 comply with Article 50(2) by 2 December 2026.

Tier 4: Minimal or No Risk and the Article 95 Option

The minimal-risk tier holds every AI system that is not prohibited, not high-risk and not caught by Article 50; the act attaches no tier-specific duty to it. The Commission's AI Act page names AI-enabled video games and spam filters as examples and states that they face no rules under the act.

The AI Office and the Member States encourage codes of conduct that apply some or all of the Chapter III Section 2 requirements to non-high-risk systems (Article 95(1)), with elements such as trustworthy-AI ethics, environmental sustainability and AI literacy (Article 95(2)). Such codes need no approval and create no presumption of conformity, and since 27 July 2026 Article 95(4) requires the interests of SMEs, start-ups and small mid-caps to be taken into account.

No Article 95 code of conduct is listed on Commission pages as of 23 October 2026. The Commission evaluates the impact and effectiveness of voluntary codes of conduct by 2 August 2028 and every three years after that (Article 112(7)).

Models under Chapter V: A Separate Track, Not a Fifth Tier

General-purpose AI models are regulated by Chapter V as models, so they carry their own duties next to the four tiers rather than forming a fifth one. A provider of such a model draws up technical documentation, informs downstream providers, puts in place a copyright policy and publishes a training-content summary (Article 53(1), points (a) to (d)); points (a) and (b) do not apply to free and open-source models with publicly available parameters unless the model has systemic risk (Article 53(2)). A model is presumed to have systemic risk when its cumulative training compute is greater than 10^25 FLOP (Article 51(2)), which adds the duties for GPAI models with systemic risk on top of Article 53.

Chapter V applies since 2 August 2025, models placed before that date comply by 2 August 2027 (Article 111(3)), and the Commission fines model providers under Article 101(1), not under Article 99.

When the Tiers Overlap: Four Combinations

The tiers are not exclusive: one AI system can trigger duties from two tiers at once, and a system that escapes one tier can still fall under another. Four combinations follow from the act and the Commission guidelines:

• Annex III and Article 5: a system that is not high-risk under Article 6(3) can still be prohibited under Article 5 (Guidelines C(2025) 5052, paragraph 38).
• Outside a ban, inside high risk: systems just outside a ban are often high-risk; Guidelines C(2025) 5052 name non-prohibited credit scoring and insurance scoring as high-risk (paragraph 37).
• Emotion recognition: emotion recognition systems are also high-risk unless Article 5(1), point (f), prohibits them (Article 50 guidelines, paragraph 102); point (f) is the ban on emotion recognition at work and in education.
• High risk and transparency: Article 50 applies without prejudice to Chapter III (Article 50(6)), so a high-risk chatbot carries both sets of duties.

A further link runs through the value chain: a distributor, importer, deployer or other third party that modifies the intended purpose of an AI system, including a general-purpose AI system, that has not been classified as high-risk and is already on the market or in service, so that it becomes high-risk, is considered its provider (Article 25(1), point (c)).

An organisation therefore checks each system against every tier, not only the strictest one.

Duties per Risk Category and Operator

Duties attach to the operator's role inside a tier: providers carry most of them, deployers the use-side duties, and Article 4 applies to both in every tier. Article 3(8) names six operators: provider, product manufacturer, deployer, authorised representative, importer and distributor. The provider vs deployer roles decide which role carries which duty, and the table maps the tiers to the roles:

Deployers that are bodies governed by public law or private entities providing public services, and deployers of systems under Annex III, points 5(b) and 5(c), carry out a fundamental rights impact assessment before deploying a high-risk AI system referred to in Article 6(2), except systems in Annex III, point 2 (Article 27(1)). Deployers of high-risk systems also assign human oversight and keep the logs under their control (Article 26).

Employers also inform workers' representatives and the affected workers before putting a high-risk AI system into service or using it at the workplace (Article 26(7)).

Application Dates per Tier after Regulation (EU) 2026/1744

The tiers apply on different dates: Article 5 points (a), (b) and (c) to (h) since 2 February 2025 and points (ba) and (bb) from 2 December 2026, Chapter V since 2 August 2025, Article 50 since 2 August 2026, and Chapter III Sections 1 to 3 from 2 December 2027 (Annex III) and 2 August 2028 (Annex I).

Before 27 July 2026 the high-risk dates were 2 August 2026 (Annex III) and 2 August 2027 (Annex I).

The general date of 2 August 2026 covers only the provisions that points (a) to (d) of Article 113, third paragraph, do not name. The Commission FAQ (last update 20 October 2026) states: "The rules for high-risk AI systems will apply starting 2 December 2027."

The EU AI Act timeline adds the dates outside the tiers, such as the 2 August 2027 deadline for national AI regulatory sandboxes (Article 57(1)).

Maximum Fines per Risk Category

The fine ceiling follows the tier: up to EUR 35 000 000 or 7 % of total worldwide annual turnover, whichever is higher, for Article 5 breaches, and up to EUR 15 000 000 or 3 % for high-risk and Article 50 duties (Article 99(3) and (4)). These amounts are the maxima of the fines that national penalty rules lay down (Article 99(1)). The minimal tier and Article 4 have no EU fine amount.

Fines on SMEs and start-ups are capped at the lower of the two amounts in every tier (Article 99(6)), while fines on small mid-caps are capped at the lower amount only under Article 99(4) and (5) (Article 99(6a)).

Union institutions, bodies, offices and agencies: up to EUR 1 500 000 for Article 5 breaches and EUR 750 000 for other breaches, imposed by the European Data Protection Supervisor (Article 100).

The EU AI Act penalties guide explains how a fine is set within these ceilings under the Article 99(7) factors, such as the gravity and duration of the infringement.

How the Lists Can Change: Article 7 and the Article 112 Review

A system can move between tiers when the lists move: the Commission can amend Annex III by delegated act under Article 7 and assesses every year whether Annex III and Article 5 need amending (Article 112(1)). Delegated acts can add or modify Annex III use cases only within the eight existing areas, using the 11 criteria of Article 7(2), and can remove use cases under Article 7(3).

The Commission's first Article 112(1) report, COM(2026) 234 of 20 May 2026, covers the need to review the list of prohibited practices and Annex III.

Placing an AI System in a Risk Category: First Steps

Five checks place an AI system in its risk category, in this order: is it an AI system, is the practice banned, is it high-risk, does Article 50 apply, and which role does the organisation hold. The checks give an orientation, not a legal assessment:

1. Test the system against the AI system definition of Article 3(1) and the scope of Article 2.
2. Compare the practice with the ten points of Article 5(1).
3. Check Annex I and Annex III, then the four conditions and the profiling rule of Article 6(3).
4. Match the system's functions against Article 50(1) to (4).
5. Assign the operator role under Article 3(8) for each system.

A ban ends the analysis. A step-by-step programme built on these checks is set out in the guide to EU AI Act compliance, for providers and deployers alike.

A provider that relies on Article 6(3) documents its assessment, and the market surveillance authority can evaluate that classification (Article 80).

The EU AI Act checklist sorts the obligations by role and risk class, with every deadline.

EU AI Act Risk Categories: Related Questions

Article 4, Article 2(1) and Article 9 settle how the four tiers relate to AI literacy, non-EU providers and risk management, while the pyramid image and the NIST AI RMF are no source of the tiers.

Is the Risk Pyramid an Accurate Picture?

The pyramid orders the tiers by the weight of their duties but hides the overlaps: one system can sit in two tiers at once. The diagram at the top of this page draws them as overlapping bands instead.

Does Article 4 Apply in Every Tier?

Article 4 applies to all providers and deployers of AI systems, whatever the tier, since 2 February 2025.

The rules on AI literacy under Article 4 reach the minimal tier too. Since 27 July 2026, providers and deployers "shall take measures to support the development of AI literacy".

Do Non-EU Providers Fall into the Same Tiers?

Non-EU providers fall into the same tiers when they place AI systems on the Union market or put them into service there (Article 2(1), point (a)), or when the output is used in the Union (Article 2(1), point (c)).

The guide to the EU AI Act for non-EU companies covers this territorial reach.

How Do the Tiers Relate to Risk Management under Article 9?

The risk management system of Article 9 is a requirement inside the high-risk tier, not a method for classifying systems into tiers.

Providers of high-risk systems run a risk management system under Article 9 once the tier is settled. Article 9(2) describes it as a continuous iterative process planned and run throughout the entire lifecycle of the system.

How Does the NIST AI RMF Compare with the Four Tiers?

The NIST AI RMF is a separate framework and no source for the AI Act tiers.

Comparisons with other frameworks appear under the AI Act compared, including NIST AI RMF vs EU AI Act.

Frequently Asked Questions about the EU AI Act Risk Categories

What are the four risk categories of the EU AI Act?

Unacceptable risk (Article 5), high risk (Article 6 with Annexes I and III), transparency risk (Article 50) and minimal or no risk. The labels come from the Commission guidelines C(2025) 5052, paragraph 2.

Is "limited risk" a term of the AI Act?

No. The act uses no "limited risk" term; the Commission calls the Article 50 tier "transparency risk", and "limited risk" is the common name for the same tier.

Can one AI system fall into more than one category?

Yes. An Annex III system can still be prohibited under Article 5 (Guidelines C(2025) 5052, paragraph 38), and Article 50 duties apply on top of the high-risk rules (Article 50(6)).

Are chatbots limited-risk AI systems?

A chatbot has an Article 50(1) duty to disclose that people interact with an AI system unless that is obvious. It can also be high-risk when its intended purpose matches an Annex III use case.

Is minimal-risk AI completely unregulated?

No tier duty applies, but Article 4 still applies to its providers and deployers, and Article 5 reaches every AI system, whatever its tier or placing date.

Who decides which category applies?

The provider assesses its own system against the act. The market surveillance authority can evaluate a provider's Article 6(3) classification of an Annex III system as not high-risk (Article 80).