Does the EU AI Act legally define four risk categories?

No. The familiar four-tier model is a practical explanation, not a set of four statutory labels. The Act separately prohibits specified practices, regulates defined high-risk systems, imposes transparency duties in particular cases and leaves many other uses without system-specific requirements. Each use must be tested against the relevant article.

What AI practices are prohibited under Article 5?

Article 5 contains a defined list, not a general ban on risky AI. One prohibition covers subliminal, purposefully manipulative or deceptive techniques that materially distort behaviour and cause, or are reasonably likely to cause, significant harm. Other prohibited practices have their own elements and exceptions, which require separate analysis.

When is a product-related AI system high-risk?

Under Article 6(1), a system is high-risk when it is a safety component, or itself a product, covered by Annex I Union harmonisation law and that product must undergo third-party conformity assessment. Both parts of the test matter; being used inside a product is not enough on its own.

Which systems have transparency duties?

Article 50 includes duties for AI systems intended to interact directly with people and for specified synthetic or manipulated content. For direct interaction, providers must ensure people are informed that they are interacting with AI unless that fact is obvious to a reasonably well-informed, observant and circumspect person.

Are ordinary office AI tools unregulated?

Not necessarily. A use that is neither prohibited nor high-risk and has no specific transparency duty may still sit within other AI Act provisions or other law, including data protection, employment and consumer rules. Classification should record purpose, affected people, role and change history rather than applying one label to a whole vendor.

EU AI Act Risk Categories: Four Practical Tiers Explained

Q: Can an Annex III system avoid high-risk classification?

Yes, in limited circumstances. Article 6(3) allows an Annex III system to fall outside high-risk status where it poses no significant risk of harm to health, safety or fundamental rights, including because it does not materially influence decisions. Profiling of natural persons remains high-risk.

Quick Answer: The four EU AI Act risk tiers are a practical map, not four legal labels. Start with prohibited practices, then test high-risk product and Annex III rules, then specific transparency duties. Most other uses have no system-specific rules, but other AI Act provisions and laws may still apply.

Summary in a mind map

EU AI Act risk categories
│
├─ Prohibited
│   ├─ Test every element of Article 5
│   └─ Controls cannot rescue a prohibited practice
├─ High-risk
│   ├─ Annex I product route or Annex III use
│   └─ Article 6(3) exception is narrow
├─ Transparency
│   ├─ Direct interaction may need a notice
│   └─ Synthetic content has separate duties
└─ Other uses
    ├─ No system-specific tier rules
    └─ Other AI Act provisions and laws may apply

Four tiers are a map, not the statute

The four-tier model helps a business sort AI uses, but the EU AI Act does not create four provisions named unacceptable, high, limited and minimal risk. It instead contains a list of prohibited practices, tests for high-risk systems, specific transparency obligations and a large remainder of other uses.

That distinction prevents shortcuts. A team cannot classify a vendor once and apply the result to every feature. The same platform might draft internal text, screen job applicants and operate a public chatbot. Purpose and deployment context can place those uses on different legal paths.

The practical sequence is therefore a set of legal tests. Check Article 5 first because a prohibited practice cannot be rescued by stronger controls. Then assess Article 6 and the relevant annexes. After that, test Article 50. Finally, record which other rules still apply.

Practical tier	Legal question	Typical outcome
Prohibited	Does every element of an Article 5 practice apply?	do not place, use or put into service
High-risk	Does Article 6(1) or 6(2) apply, subject to 6(3)?	full high-risk regime for the relevant role
Transparency	Does Article 50 cover the interaction or output?	inform people or mark/disclose content
Other uses	None of those tests applies	no system-specific tier, but other law remains

Prohibited practices come first

Article 5 is a defined list. It does not prohibit an AI system merely because somebody describes it as dangerous. Each practice has elements concerning technique, purpose, context, effect and sometimes an exception.¹

Article 5(1)(a), for example, prohibits placing on the market, putting into service or using an AI system that deploys subliminal, purposefully manipulative or deceptive techniques and materially distorts behaviour in a way that causes, or is reasonably likely to cause, significant harm.¹ All parts matter. A persuasive interface is not automatically within this provision, while a harmful deceptive technique cannot be relabelled as low risk.

A small business should describe the actual behaviour rather than copy a marketing category. Record what the system does, who experiences it, what decision or behaviour it influences and what harm could follow. Escalate a plausible Article 5 match before procurement or deployment continues.

Two routes lead to high-risk status

The first route concerns product safety. Under Article 6(1), a system is high-risk where it is a safety component of a product, or is itself a product, covered by Annex I Union harmonisation legislation and that product must undergo third-party conformity assessment.¹ Both the Annex I connection and the conformity-assessment condition are necessary.

The second route concerns uses listed in Annex III, including specified cases in biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration and justice. The activity must fit the wording of the listed use; a broad resemblance is not enough.

The Commission states that high-risk rules for the listed areas apply from 2 December 2027. Rules for systems integrated into products such as robotics and industrial machinery apply from 2 August 2028.³ These dates affect readiness planning, not the substance of the classification tests.

The Annex III exception is narrow

An Annex III listing does not always end the analysis. Article 6(3) says an Annex III system is not high-risk where it does not pose a significant risk of harm to health, safety or fundamental rights, including by not materially influencing decision-making.¹

The provision gives a route to a reasoned exclusion, not a casual low-risk label. The business must examine the specific task and impact. A system performing a narrow procedural or preparatory task may differ from one that ranks candidates or determines access to a service.

Profiling of natural persons is always high-risk when the system falls within Annex III.¹ A team must not use the exception to remove such profiling from high-risk status. The classification record should identify the Annex III point, the system's task, its influence on decisions and the reason for any Article 6(3) conclusion.

Transparency duties form a separate route

Some systems carry duties because people interact with them or encounter generated or manipulated content. Article 50(1) requires providers of systems intended to interact directly with natural persons to ensure those people are informed that they are interacting with AI, unless this is obvious to a reasonably well-informed, observant and circumspect person.¹

This is not a lesser version of high-risk status. A public chatbot may have a transparency duty without being high-risk. A high-risk system may also have separate transparency requirements. Each applicable rule should appear as its own line in the compliance record.

The notice must work in the actual channel. A voice service needs an audible route; a chat needs a clear visual disclosure. The exception for obvious interaction is contextual and should not become a default assumption.

How to classify all other uses

The Commission describes AI-enabled video games and spam filters as examples of minimal or no-risk uses for which the Act introduces no system-specific rules.² The practical fourth tier collects uses that do not trigger the earlier tests.

“Other” does not mean legally irrelevant. Provider duties for general-purpose models, AI literacy provisions and rules outside the AI Act may still matter. Personal data use can trigger GDPR or UK GDPR duties. Workplace decisions can engage employment law. Consumer-facing claims can engage consumer protection.

A useful register therefore records more than a tier. It identifies the system, purpose, affected people, data, provider or deployer role, Article 5 check, Article 6 route, Article 50 duty, other applicable law, owner and review date. A change in purpose or functionality triggers a fresh classification.

What to do next

This four-tier explanation cannot classify a system without facts about its purpose, role and deployment. It also does not replace the detailed wording and exceptions in Articles 5, 6 and 50. Borderline use cases, profiling and safety components need specific legal and technical analysis.

UK firms are not directly bound by the Act merely because they are established in the UK, but a UK company may face it through EU market activity or its place in a relevant supply chain. Geography and territorial scope therefore require a separate assessment.

Create one register row for every distinct AI use, not every vendor, and reuse it when you answer a vendor's AI Act questionnaire. Run the four tests in order and cite the relevant article beside the answer. Link the result to the broader AI governance picture, the Article 4 literacy duty and the Article 50 transparency rules.