Frequently asked questions about the EU AI Act

Providers, deployers, and other operators can use these short answers to find EU AI Act rules for AI systems and general-purpose AI models. Each answer refers to the page with the full rule.

Disclaimer

This page gives general information. It is not legal advice.

What is the EU AI Act?

The EU AI Act is Regulation (EU) 2024/1689 on artificial intelligence. It is "binding in its entirety and directly applicable in all Member States" (Article 113). It gives rules for AI systems and for general-purpose AI models in the Union.

Article 2 gives a different scope test for each actor. It also gives exclusions, each with its own conditions. Different provisions apply from different dates (Article 113). See EU AI Act scope and Dates.

Is the Digital Omnibus on AI in force?

Yes. Regulation (EU) 2026/1744, the Digital Omnibus on AI, amended the EU AI Act from 27 July 2026. The Official Journal published it on 24 July 2026. It entered into force on the third day after publication (Article 4 of Regulation (EU) 2026/1744).

Entry into force is different from the application date of each amended provision. For example, the new prohibitions in Article 5(1), points (ba) and (bb), apply from 2 December 2026. Article 113 also gives new dates for the high-risk rules, and Article 111 gives new transitional rules. See Digital Omnibus on AI and Dates.

When does the EU AI Act apply?

The EU AI Act has different application dates for different provisions and AI systems. The general date is 2 August 2026 (Article 113). Article 113 gives these other dates:

• 2 February 2025 — Chapters I and II, for example Article 4 and Article 5(1), points (a) to (h).
• 2 August 2025 — Chapter III Section 4, Chapter V, Chapter VII, Chapter XII (but not Article 101), and Article 78.
• 27 July 2026 — Articles 102 to 110.
• 2 December 2026 — Article 5(1), points (ba) and (bb), and Article 5(1a) and (1b).
• 2 December 2027 — Chapter III Sections 1, 2, and 3, but not Article 6(5), for high-risk AI systems in Article 6(2) and Annex III.
• 2 August 2028 — The same sections, for high-risk AI systems in Article 6(1) and Annex I. For Section B of Annex I, Article 2(2) also applies.

Article 111 gives transitional rules for systems and models on the market before a cutoff date:

• Article 111(1) — Annex X large-scale IT components placed on the market or put into service before 2 August 2027 must be "brought into compliance" by 31 December 2030.
• Article 111(2) — Other high-risk systems placed on the market or put into service before their Chapter III date. The EU AI Act applies to their operators only if the systems have "significant changes in their designs" from that date. In all cases, providers and deployers of systems intended for use by public authorities must "take the necessary steps to comply" by 2 August 2030. Recital 39 of Regulation (EU) 2026/1744 addresses subsequent units of the same type and model.
• Article 111(3) — Providers of general-purpose AI models placed on the market before 2 August 2025 must "take the necessary steps" to comply by 2 August 2027.
• Article 111(4) — Providers of synthetic-content systems placed on the market before 2 August 2026 must "take the necessary steps" to comply with Article 50(2) by 2 December 2026.

Article 111(1) and (2) apply "without prejudice to the application of Article 5".

Does the EU AI Act apply to non-EU companies?

Yes, it can. Article 2(1)(a) applies to providers placing on the market or putting into service AI systems in the Union, and to providers placing on the market general-purpose AI models in the Union. It applies "irrespective of whether those providers are established or located within the Union or in a third country". Article 2(1)(c) applies to providers and deployers in a third country "where the output produced by the AI system is used in the Union".

Importers, distributors, product manufacturers, and authorized representatives have their own tests (Article 2(1)(d) to (f)). Article 2(3) to (12) give exclusions, and Article 2(2) limits the rules for Article 6(1) systems of Annex I Section B. Some third-country providers must appoint an authorized representative (Articles 22 and 54). See Article 2.

Is the EU AI Act a risk pyramid?

Not as a set of levels with no overlap. The EU AI Act uses different tests for scope, for prohibited practices, for high-risk classification, and for transparency. One AI system can be in the scope of more than one of these tests.

The Commission refers to "a risk-based approach" with "four different risk categories" (Communication to the Commission "C(2026) 5054 final"). But the obligations come from the tests in the Act. For example, a high-risk AI system can also have Article 50 obligations (Article 50(6)). Chapter V gives different obligations to providers of general-purpose AI models. Article 2 gives the scope tests and exclusions, with the Section B rule of Article 2(2).

See Article 2, Article 5, Article 6 and Article 3(63).

What counts as an AI system under the EU AI Act?

Article 3(1) defines an AI system through its characteristics:

‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments;

The Commission guidelines on the AI system definition ("C(2025) 5053 final") divide it into seven elements (paragraph (9)) and are not binding (paragraph (7)). In the guidelines, "each system must be assessed based on its specific characteristics" (paragraph (6)). Thus, their examples are not automatic exclusions. Article 3(12) gives a different definition for the "intended purpose" of an AI system.

See Article 3(1).

Are generative AI models automatically high-risk?

No. A general-purpose AI model is not automatically a high-risk AI system. Article 6 classifies AI systems, not models. An AI system is high-risk through Article 6(1) or through Annex III (Article 6(2)). Article 6(3) gives conditions in which some Annex III systems are not high-risk.

Article 3(63) defines a general-purpose AI model, and Chapter V gives different obligations to its provider. A model "with systemic risk" (Article 51) is a different test, not a high-risk classification. One provider can have the obligations of a model provider and of a system provider. An example is a provider that integrates its own model into an AI system (Article 3(68)).

See Article 3(63) — general-purpose AI model.

Who is the provider of an AI system?

In the EU AI Act, the provider of an AI system develops the system, or has it developed. It places the system on the market, or puts into service the system, with its own name or trademark, "whether for payment or free of charge" (Article 3(3)). A provider can be a natural or legal person, a public authority, an agency, or a different body.

The provider of a general-purpose AI model develops the model, or has it developed, and places it on the market with its own name or trademark. Putting into service applies only to AI systems. Article 25(1) can make a distributor, importer, deployer, or other third party the provider of a high-risk AI system (Article 25). See Article 3(3).

Who is the deployer of an AI system?

In the EU AI Act, a deployer is a person or body "using an AI system under its authority" (Article 3(4)). A deployer can be a natural or legal person, a public authority, an agency, or a different body. The definition excludes use "in the course of a personal non-professional activity".

The deployer role does not set a role in Union data protection law, which continues to apply (Article 2(7)). Article 2(1)(b) and (c) give the scope tests for deployers in the Union and in a third country. Article 26 gives obligations to deployers of high-risk AI systems, but Article 2(2) excludes them for Article 6(1) systems of Annex I Section B.

See Article 3(4).

When does a deployer or distributor become a provider?

EU AI Act Article 25(1) makes a distributor, importer, deployer or other third party the provider of a high-risk AI system, with Article 16 obligations, in these cases:

• It puts its name or trademark on a high-risk AI system that was placed on the market or put into service before. Contracts can give the obligations to a different party.
• It makes a substantial modification to such a system, and the system stays high-risk in Article 6.
• It changes the intended purpose of an AI system that is not high-risk and that was placed on the market or put into service, also a general-purpose AI system. The change makes the system high-risk in Article 6.

The initial provider is then not the provider of that system, but must cooperate and give information and technical access, unless it "has clearly specified that its AI system is not to be changed into a high-risk AI system" (Article 25(2)). Article 25(3) to (5) add rules for product manufacturers, written agreements, and trade secrets. Article 2(2) excludes Article 25 for Article 6(1) systems of Annex I Section B.

See Article 25.

Do non-EU AI providers need an EU representative?

Yes, in two cases. Providers established in third countries must appoint an authorized representative established in the Union, by written mandate:

• High-risk AI systems — before the provider makes the system available on the Union market (Article 22(1)).
• General-purpose AI models — before the provider places the model on the Union market (Article 54(1)).

The mandate gives the tasks, for example to keep documents for 10 years (Articles 22(3) and 54(3)). The representative must terminate the mandate if it "considers or has reason to consider" that the provider acts contrary to its obligations, and immediately report the termination (Articles 22(4) and 54(5)). Article 54(6) removes the model duty for qualifying free and open-source models without systemic risks. Article 2(2) excludes Article 22 for Article 6(1) systems of Annex I Section B.

See Article 22 and Article 54.

Who must provide AI literacy training?

Providers and deployers must "take measures to support the development of AI literacy" of their staff and of other persons who operate or use AI systems for them (EU AI Act Article 4(1)). Article 4(1) does not prescribe training as the mandatory measure. For the measures, they must think about the knowledge, experience, education, and training of these persons. They must also think about the context of use and the persons on whom the systems are used.

Article 4 does not tell providers or deployers "to guarantee any specific level of AI literacy of any individual". It applies to all AI systems in scope from 2 February 2025, not only to high-risk AI systems (Article 113, point (a)). Article 2(2) excludes Article 4 for Article 6(1) systems of Annex I Section B. The Commission and the Member States help providers and deployers, mainly SMEs (Article 4(2)).

See Article 4.

What AI practices does the EU AI Act prohibit?

Article 5(1) prohibits these AI practices, each on its own conditions and exceptions:

• (a) subliminal techniques "beyond a person’s consciousness", or purposefully manipulative or deceptive techniques. The objective or the effect is materially distorting behavior by "appreciably impairing their ability to make an informed decision". The person then makes "a decision that they would not have otherwise taken", in a manner that causes or is reasonably likely to cause "significant harm".
• (b) exploitation of vulnerabilities due to age, disability, or "a specific social or economic situation", with the objective or the effect of materially distorting behavior in a manner that causes or is reasonably likely to cause "significant harm"
• (ba) realistic intimate or sexually explicit material of an identifiable natural person, without explicit consent
• (bb) "material or performance within the meaning of Article 2, points (c) and (e), of Directive 2011/93/EU", unless a "without right" defense applies in national law
• (c) social scoring that causes "detrimental or unfavourable treatment" in unrelated contexts, or treatment that is "unjustified or disproportionate"
• (d) criminal-offense risk prediction "based solely on the profiling of a natural person or on assessing their personality traits and characteristics", with an exception in point (d)
• (e) the creation or expansion of facial recognition databases through "the untargeted scraping of facial images"
• (f) emotion inference in the workplace and in education institutions, but not for medical or safety reasons
• (g) biometric categorization to infer race, political opinions, trade union membership, religious or philosophical beliefs, sex life, or sexual orientation, with two exclusions
• (h) the use of real-time remote biometric identification in publicly accessible spaces for law enforcement, with three exceptions.

Points (ba) and (bb), with Article 5(1a) and (1b), apply from 2 December 2026, and the other points from 2 February 2025. Article 2(2) excludes Article 5 for Article 6(1) systems of Annex I Section B.

See Article 5.

What does the AI Omnibus change about intimate images and child sexual abuse material?

Regulation (EU) 2026/1744 added two prohibitions to Article 5(1) of the EU AI Act. They apply from 2 December 2026.

• (ba) — AI systems that generate or manipulate "realistic images, videos, audio or similar material" of the intimate parts of an identifiable natural person, or of that person in sexually explicit activities. The prohibition applies where that person did not give "freely-given, specific, informed, unambiguous and explicit consent for that generation or manipulation".
• (bb) — AI systems that generate or manipulate "material or performance within the meaning of Article 2, points (c) and (e), of Directive 2011/93/EU", unless "a ‘without right’ defence applies under national law".

Article 5(1a) prohibits placing on the market and putting into service only in one of these conditions:

• That generation or manipulation is the intended purpose of the AI system.
• The system's "design, training, architecture, capabilities or user-facing functionalities" make that generation or manipulation "a reasonably foreseeable and reproducible outcome, without requiring significant technical modification". Also, the system does not have "reasonable and adequate technical safety measures and other safeguards" to reliably prevent that generation or manipulation, with reasonably foreseeable misuse taken into account, and to correct observed or reported misuse.

Use is prohibited only where the deployer uses the system for that purpose. Article 5(1b) limits "manipulation" in point (ba).

See Article 5(1)(ba) and (bb).

Can police use real-time remote biometric identification in public?

Only on narrow conditions. Article 5(1)(h) prohibits this use for law enforcement, unless it is "strictly necessary" for one of these objectives:

• "the targeted search for specific victims of abduction, trafficking in human beings or sexual exploitation of human beings, as well as the search for missing persons"
• the prevention of a "specific, substantial and imminent threat" to life or physical safety, or of a terrorist attack, on the conditions of point (h)(ii)
• the localization or identification of "a person suspected of having committed a criminal offence" in Annex II, with "a custodial sentence or a detention order for a maximum period of at least four years".

Each use must also obey Article 5(2) to (7):

• Article 5(2) — identity confirmation only, an assessment of harms and rights, national safeguards, and a completed FRIA and registration, with an urgency rule.
• Article 5(3) — prior authorization by a judicial or independent administrative authority "whose decision is binding", on evidence of necessity and proportionality, with a 24-hour urgency rule. After a refusal, use stops and the data is deleted. No decision with an "adverse legal effect" is based only on the output.
• Article 5(4) to (7) — notifications, national authorization rules and stricter national laws, and annual reports.

What does Article 50 require?

Article 50 gives providers and deployers different transparency obligations for specified AI uses:

• Providers, interaction — Make systems that "interact directly with natural persons" inform them that they interact with an AI system, unless this is obvious (Article 50(1)).
• Providers, synthetic content — Mark synthetic audio, image, video and text outputs as artificially generated or manipulated, in a machine-readable format, with technical conditions and exceptions (Article 50(2)).
• Deployers, emotion recognition and biometric categorization — Inform exposed persons, and obey Union data protection law (Article 50(3)).
• Deployers, deep fakes — Disclose artificial generation or manipulation, with a disclosure limit for evidently artistic and equivalent works (Article 50(4)).
• Deployers, public-interest text — Disclose AI generation or manipulation, unless human review or editorial control and editorial responsibility apply (Article 50(4)).

Article 50(1) to (4) also have exceptions for criminal-law uses that law authorizes or permits, each with its own scope and conditions. For example, the Article 50(1) exception does not apply to systems that are "available for the public to report a criminal offence". Article 50(5) gives the time, clarity and accessibility of the information, and Article 50(6) keeps Chapter III and other transparency obligations. Article 50 applies from 2 August 2026, with an Article 111(4) transitional rule for older synthetic-content systems (Dates). Article 2(2) excludes Article 50 for Article 6(1) systems of Annex I Section B.

Is limited-risk AI the same as Article 50?

No. Article 50 applies its own tests for specified uses, not a general limited-risk category. Article 3 has no definition of "limited risk". Each paragraph of Article 50 has its own conditions, actor, and exceptions.

A high-risk AI system can also have Article 50 obligations, because Article 50 does not change the obligations of Chapter III (Article 50(6)). Article 5 can also prohibit an AI system with Article 50 obligations. Thus, compliance with Article 50 does not show that the EU AI Act permits the use. Article 2(2) excludes Articles 5 and 50 for Article 6(1) systems of Annex I Section B.

See Article 50.

Are the Article 50 guidelines binding?

No. The Commission guidelines on Article 50 do not replace the binding EU AI Act. The guidelines tell readers: "These Guidelines are non-binding" (paragraph (5)).

On 20 July 2026, Communication to the Commission "C(2026) 5054 final" approved the content of the draft guidelines. The Communication to the Commission gives this condition: "The Guidelines will be formally adopted by the Commission at a later date, when all language versions are available. It is only from that moment that these Guidelines will be applicable." Article 50 applies from 2 August 2026.

See Guidance status.

What makes an AI system high-risk?

Article 6 gives two routes to high-risk classification:

• Article 6(1), products — The AI system is "intended to be used as a safety component of a product", or is itself a product, in the Annex I legislation. That product must also undergo a third-party conformity assessment. The two conditions apply together, with the limits of Article 6(1a) to (1c).
• Article 6(2), Annex III — The AI system is in a specified use of an Annex III area.

Article 6(3) gives conditions in which an Annex III system is not high-risk, and Article 2(2) limits the rules for Article 6(1) systems of Annex I Section B. Chapter III Sections 1 to 3, but not Article 6(5), apply from 2 December 2027 (Annex III) or 2 August 2028 (Annex I), with Article 111(2) for older systems (Dates).

See Article 6.

Is every Annex III AI system automatically high-risk?

No. In Article 6(3), an Annex III system is not high-risk if it does not pose "a significant risk of harm to the health, safety or fundamental rights of natural persons, including by not materially influencing the outcome of decision making". One of these conditions must also apply (Article 6(3)(a) to (d)). The AI system is intended to:

• "perform a narrow procedural task"
• "improve the result of a previously completed human activity"
• detect decision-making patterns or deviations from prior decision-making patterns, and is not meant to replace or influence the completed human assessment "without proper human review"
• "perform a preparatory task to an assessment relevant for the purposes of the use cases listed in Annex III".

An Annex III system that does profiling of natural persons is always high-risk. A provider that concludes that its system is not high-risk must document its assessment before placing on the market or putting into service (Article 6(4)). It gives the documentation to authorities "upon request", and registers (Article 49(2)).

See Article 6(3).

What are the Articles 8–15 requirements for high-risk AI?

Providers must make sure that their high-risk AI systems comply with Chapter III Section 2 (Article 16(a)), with the intended purpose and "the generally acknowledged state of the art" taken into account (Article 8(1)):

• Article 9 — a risk management system for the full lifecycle
• Article 10 — data and data governance, only for testing data sets if the system does not train AI models (Article 10(6))
• Article 11 — technical documentation, before placing on the market or putting into service, kept up to date
• Article 12 — the "automatic recording of events (logs)"
• Article 13 — transparency, and instructions for use
• Article 14 — human oversight
• Article 15 — accuracy, robustness, and cybersecurity.

Article 4a gives a different conditional permission for special-category data. Articles 8–15 apply from 2 December 2027 (Annex III) or 2 August 2028 (Annex I), with Article 111 transitions (Dates). Article 2(2) excludes Articles 4a and 8–15 for Article 6(1) systems of Annex I Section B.

See Articles 8–15.

Can special-category data be used to detect AI bias?

Yes, but only as an exception and only on the conditions of Article 4a. Providers of high-risk AI systems can process special categories of personal data "to the extent strictly necessary" for bias detection and correction in Article 10(2)(f) and (g). "Appropriate safeguards" and all of these conditions of Article 4a(1) apply:

• no other data, also synthetic or anonymized data, can do the bias detection and correction effectively
• limits on re-use, and "state-of-the-art security and privacy-preserving measures"
• protection against misuse, with strict, documented access controls, and access only for authorized persons with "appropriate confidentiality obligations"
• the data is not "transmitted, transferred or otherwise accessed by other parties"
• deletion when the bias is corrected or at the end of the retention period, "whichever comes first"
• records of why the processing was "strictly necessary", and "why that objective could not be achieved by processing other data".

Article 4a(2) lets other providers and deployers do this processing when "strictly necessary" for specified biases, with all conditions of paragraph 1, and "does not create any obligation". Union data protection law also applies (Article 2(7)). Article 2(2) excludes Article 4a for Article 6(1) systems of Annex I Section B.

See Article 4a.

Who must conduct an AI Act fundamental-rights impact assessment?

Article 27 tells specified deployers to do a fundamental rights impact assessment (FRIA) before the first use of a high-risk AI system in Article 6(2), other than Annex III point 2 (critical infrastructure). These deployers must do the FRIA:

• bodies governed by public law
• private entities that give public services
• deployers of Annex III point 5(b) systems, for creditworthiness and credit scores
• deployers of Annex III point 5(c) systems, for risk assessment and pricing in life and health insurance.

The FRIA has the elements of Article 27(1)(a) to (f). The deployer can use previous assessments "in similar cases", and must update changed elements (Article 27(2)). It notifies the market surveillance authority of the results, but "may be exempt" for the Article 46(1) derogation (Article 27(3)). A deployer can refer to a data protection impact assessment for some obligations, but the other Article 27 obligations stay (Article 27(4)). Article 5(2) also makes a FRIA a prerequisite for real-time remote biometric identification.

See Article 27.

What human oversight does Article 14 require?

Article 14 tells providers to make high-risk AI systems that natural persons can oversee "effectively" during use. The measures are "commensurate with the risks, level of autonomy and context of use". They are of "either one or both" of two types: the provider builds them into the system "when technically feasible", or identifies them for the deployer to use (Article 14(3)).

The persons who do the oversight must be able, "as appropriate and proportionate", to understand, monitor, and interpret the system, and to know about "automation bias". They must also be able not to use the system, to override its output, and to stop it (Article 14(4)).

For Annex III point 1(a) systems, the oversight measures must make sure that "no action or decision is taken by the deployer on the basis of the identification resulting from the system" unless it is "separately verified and confirmed by at least two natural persons with the necessary competence, training and authority". The requirement for "a separate verification by at least two natural persons" does not apply to systems used for law enforcement, migration, border control or asylum, where Union or national law considers "the application of this requirement to be disproportionate" (Article 14(5)). Deployers give oversight to competent persons with "the necessary support" (Article 26(2)). Article 2(2) excludes Article 14 for Article 6(1) systems of Annex I Section B.

See Article 14.

Can I ask for an explanation of an AI-based decision?

Yes, if the conditions of Article 86 apply. An affected person can get "clear and meaningful explanations" from the deployer of "the role of the AI system in the decision-making procedure and the main elements of the decision taken". All of these conditions must apply:

• The deployer took the decision on the basis of the output of a high-risk AI system in Annex III, other than point 2.
• The decision "produces legal effects or similarly significantly affects that person".
• The person considers that the decision has an "adverse impact on their health, safety or fundamental rights".

Union law, or national law that complies with Union law, can give exceptions or restrictions (Article 86(2)). Article 86 applies only where other Union law does not give the right (Article 86(3)).

See Article 86.

What conformity-assessment route applies to high-risk AI?

Article 43 gives a conformity assessment procedure for each category of high-risk AI system:

• Annex III point 1 (biometrics) — Internal control (Annex VI) or a notified body (Annex VII), if the provider applied harmonized standards or, where applicable, common specifications. In the other cases, Annex VII. For systems of law enforcement, immigration, or asylum authorities or Union bodies, the market surveillance authority of Article 74(8) or (9) is the notified body (Article 43(1)).
• Annex III points 2 to 8 — internal control in Annex VI (Article 43(2)).
• Section A of Annex I — the procedure of the sectoral legislation, with the Section 2 requirements and an assessment of the quality management system (Article 43(3)). This also applies to a Section A system in Annex III.

The provider completes the procedure before placing on the market or putting into service (Article 16(f)). Article 46(1) lets a market surveillance authority authorize the placing on the market or putting into service of "specific high-risk AI systems" in its Member State, "for exceptional reasons" that Article 46(1) names. The authorization is "for a limited period while the necessary conformity assessment procedures are being carried out", and these procedures are completed "without undue delay". The authority issues it only if it concludes that the system complies with Chapter III Section 2 (Article 46(3)). Article 46(2) and (7) add urgency and Section A rules (Article 46).

Article 2(2) excludes Article 43 for Article 6(1) systems of Annex I Section B. For the dates, see Dates.

What is the difference between Annex VI and Annex VII?

Annex VI gives internal control. Annex VII includes an assessment by a notified body. EU AI Act Article 43 gives the rules for which procedure a provider must apply to a high-risk AI system.

• Annex VI — Without a notified body, the provider verifies that its quality management system complies with Article 17. It examines the technical documentation to assess compliance with Chapter III Section 2. It also makes sure that "the design and development process" and the Article 72 monitoring agree with the documentation.
• Annex VII — A notified body assesses the quality management system and the technical documentation, with access to data and, on conditions, to the model. It examines changes, does surveillance, and issues a certificate for the period of Article 44(2).

The "reference" of EN 18286:2026 in the Official Journal does not alone make Annex VI available for an Annex III point 1 system. See Annex VI — internal control and Annex VII — notified-body assessment.

When does a high-risk AI system need reassessment?

A high-risk AI system with a previous conformity assessment must undergo a new conformity assessment after substantial modification (EU AI Act Article 43(4)). This applies when the modified system goes to more deployers, and also when the current deployer continues to use it.

A substantial modification is a change after placing on the market or putting into service that the initial conformity assessment did not foresee or plan. The change affects compliance with Chapter III Section 2, or changes the intended purpose (Article 3(23)). For systems that continue to learn, changes that the provider "pre-determined" at the initial assessment, in the Annex IV point 2(f) documentation, are not a substantial modification.

A different actor that makes a substantial modification can become the provider (Article 25(1)(b)). The Article 11(1) duty to keep the documentation up to date is different. Article 2(2) excludes Article 43 for Article 6(1) systems of Annex I Section B. See Article 43(4).

What is the EU declaration of conformity for AI?

Article 47 tells the provider to write an EU declaration of conformity for each high-risk AI system. The declaration is "written machine readable, physical or electronically signed". It contains a statement of compliance with Chapter III Section 2, and the information in Annex V. With it, the provider becomes responsible for that compliance (Article 47(4)).

The provider also translates it for the authorities, and keeps it "at the disposal of the national competent authorities" for 10 years after the system is placed on the market or put into service. The provider gives a copy "upon request", and keeps it up to date "as appropriate". If other Union harmonization legislation also has a declaration, the provider writes one declaration. Article 2(2) excludes Article 47 for Article 6(1) systems of Annex I Section B.

See Article 47.

How must CE marking appear on a high-risk AI system?

Providers must affix CE marking "visibly, legibly and indelibly" (EU AI Act Article 48). If this is not possible or "not warranted" because of the nature of the system, the marking goes on the packaging or the accompanying documentation. The general principles of Article 30 of Regulation (EC) No 765/2008 apply.

These rules also apply:

• Digital systems — "high-risk AI systems provided digitally" use a digital CE marking only if it is easy to access through the interface, or through a machine-readable code or other electronic means (Article 48(2)).
• Notified body — where applicable, its identification number follows the marking, also in promotional material (Article 48(4)).
• Other Union law — the marking also shows compliance with other Union law that has a CE marking (Article 48(5)).

Article 2(2) excludes Article 48 for Article 6(1) systems of Annex I Section B. See Article 48.

Which AI systems must be registered in the EU database?

Article 49 has different registration rules for providers and for specified deployers, with the information in Annex VIII:

• Annex III systems — The provider or its authorized representative registers itself and the system before placing on the market or putting into service (Article 49(1)).
• Article 6(3) systems — A provider that concludes that its Annex III system is not high-risk also registers (Article 49(2)).
• Public deployers — Public authorities and Union bodies, or persons who act for them, register themselves and the use before they put into service or use an Annex III system (Article 49(3)).
• Annex III point 2 — These systems are registered at national level, not in Article 49(1) or (3) (Article 49(5)).

Annex III points 1, 6, and 7 systems in law enforcement, migration, asylum, and border control management are registered in a "secure non-public section", which is not an exemption (Article 49(4)). Public deployers must not use a system that is not registered, and must inform the provider or the distributor (Article 26(8)). Article 6(1) classification alone does not cause registration.

See Article 49.

Does EN 18286 give automatic AI Act compliance?

No. Commission Implementing Decision (EU) 2026/2274 published the "reference" of EN 18286:2026 in the Official Journal on 14 October 2026. The decision entered into force on that day. Article 40(1) gives a presumption of conformity only "to the extent that those standards cover those requirements or obligations".

For EN 18286:2026, the presumption includes only the provisions in Table ZA.1 of Annex ZA of the standard. These provisions are Article 17(1) and the first sentence of Article 11(1), each on the conditions of its row. The standard does not give conformity with the full EU AI Act. Its "reference" alone also does not make Annex VI available for an Annex III point 1 system (Article 43(1)).

See EN 18286 Annex ZA and presumption of conformity and Standards and common specifications.

What is a general-purpose AI model?

Article 3(63) defines a general-purpose AI model by its generality and its capabilities:

‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market;

The Commission GPAI guidelines ("C(2025) 7719 final") give a non-binding "indicative criterion": more than 10²³ FLOP of "training compute", and generation of language, text-to-image, or text-to-video (paragraphs (9) and (17)). A model can satisfy the criterion without the definition, or the definition without the criterion (paragraph (20)). The criterion is different from the 10²⁵ presumption for systemic risk in Article 51(2).

See Article 3(63).

When does a GPAI model have systemic risk?

Article 51(1) classifies a general-purpose AI model as a model with systemic risk in one of these cases:

• It has high-impact capabilities, "evaluated on the basis of appropriate technical tools and methodologies, including indicators and benchmarks".
• A Commission decision, ex officio or after a "qualified alert", finds "capabilities or an impact equivalent" to point (a), on the criteria of Annex XIII.

More than 10²⁵ floating-point operations of training computation gives a presumption of high-impact capabilities (Article 51(2)). The provider notifies the Commission "without delay and in any event within two weeks" after the Article 51(1)(a) condition is met, or after the provider knows that it will be met (Article 52(1)). It can argue that the model "exceptionally" has no systemic risks, and the Commission decides (Article 52(2) and (3)). After a designation, the provider can send a "reasoned request" for a reassessment, at the earliest after six months (Article 52(5)).

See Article 51 and Article 52.

What does Article 53 require of GPAI providers?

Article 53(1) tells providers of general-purpose AI models to do these tasks:

• (a) Keep up-to-date technical documentation with a minimum of the Annex XI information, for the AI Office and authorities "upon request".
• (b) Give up-to-date information with a minimum of the Annex XII elements to providers that intend to integrate the model.
• (c) Make a policy to comply with Union copyright law, also with rights reservations in Article 4(3) of Directive (EU) 2019/790.
• (d) Make public a "sufficiently detailed summary" of the training content, on the AI Office template.

Point (b) applies without prejudice to intellectual property rights and trade secrets. Providers must also cooperate with the Commission and the authorities (Article 53(3)). The free and open-source exception of Article 53(2) applies only to points (a) and (b), and not to models with systemic risk.

See Article 53.

Are open-source GPAI models exempt from the EU AI Act?

No. The EU AI Act gives some exceptions to free and open-source general-purpose AI models, on these conditions:

• The license allows "the access, usage, modification, and distribution of the model".
• The parameters, "including the weights, the information on the model architecture, and the information on model usage", are publicly available.

For such a model, Article 53(2) removes only Article 53(1), points (a) and (b), and Article 54(6) removes the duty to appoint an authorized representative. These two exceptions do not apply to models with systemic risk. Article 2(12) gives a different exclusion for AI systems "released under free and open-source licences", but not for high-risk AI systems or AI systems in Article 5 or 50.

See Article 53(2).

What extra duties apply to systemic-risk GPAI models?

Article 55(1) gives four more obligations to providers of general-purpose AI models with systemic risk, together with Articles 53 and 54:

• (a) "model evaluation" with standardized protocols and documented adversarial testing
• (b) assessment and mitigation of possible systemic risks at Union level, with their sources
• (c) tracking, documentation, and report of serious incidents and possible corrective measures "without undue delay" to the AI Office and, "as appropriate", national competent authorities
• (d) "an adequate level of cybersecurity protection" for the model and its physical infrastructure.

Until a harmonized standard is published, these providers can rely on codes of practice to demonstrate compliance. Compliance with harmonized standards gives "the presumption of conformity to the extent that those standards cover those obligations". Providers that "do not adhere to an approved code of practice or do not comply with a European harmonised standard" must demonstrate "alternative adequate means of compliance" (Article 55(2)). The fixed time limits of Article 73 are not for Article 55(1)(c).

See Article 55.

Is the GPAI Code of Practice mandatory?

No. The General-Purpose AI Code of Practice is a voluntary method to demonstrate compliance with Articles 53 and 55 (Article 56(7)). Until a harmonized standard is published, providers can rely on codes of practice. Providers that "do not adhere to an approved code of practice or do not comply with a European harmonised standard" must demonstrate "alternative adequate means of compliance for assessment by the Commission" (Articles 53(4) and 55(2)).

The Code was published on 10 July 2025, and on 1 August 2025 the Commission and the AI Board found it "an adequate voluntary tool". If a code is not finalized by 2 August 2025, or if the AI Office "deems it is not adequate", the Commission can adopt "common rules" (Article 56(9)). The Code does not remove a legal obligation.

See Article 56.

When do GPAI obligations apply?

Chapter V, with the obligations of providers of general-purpose AI models, applies from 2 August 2025 (Article 113, point (b)). Providers of models placed on the market before 2 August 2025 must comply by 2 August 2027 (Article 111(3)).

Article 101, on fines for these providers, applies from 2 August 2026 (Article 113). The Commission page "The enforcement framework of the AI Act" tells readers that the enforcement powers of the AI Office apply from 2 August 2026. The high-risk dates of Article 113, point (c), are not dates for Chapter V. The enforcement dates give the powers, but do not show that an authority took a measure against a provider.

See Dates and Penalties and enforcement.

What is the Article 72 post-market monitoring plan?

Article 72 tells providers of high-risk AI systems in its scope to "establish and document" a post-market monitoring system, proportionate to the nature of the AI technologies and to the risks. The system collects and analyzes performance data for the full lifetime of the system, to examine continuous compliance with Chapter III Section 2. It uses a post-market monitoring plan in the Annex IV technical documentation (Article 72(1) to (3)).

These rules change the answer for some providers:

• Template — The Commission must adopt guidance with a template "by 2 September 2027" (Article 72(3)). This is a time limit, not a published template.
• Data — "Where relevant", the monitoring includes interaction with other AI systems, but not sensitive operational data of law enforcement deployers (Article 72(2)).
• Sectoral systems — For Annex I Section A systems where "a post-market monitoring system and plan are already established under that legislation", providers can integrate the Article 72 elements into them, if this gives "an equivalent level of protection". The same applies to Annex III point 5 systems that financial institutions place on the market or put into service, if Union financial services law gives these institutions requirements on their "internal governance, arrangements or processes" (Article 72(4)).

Article 2(2) excludes Article 72 for Article 6(1) systems of Annex I Section B. See Article 72.

What counts as a serious AI incident?

Article 3(49) defines a serious incident as "an incident or malfunctioning of an AI system that directly or indirectly leads to any of the following":

• "the death of a person, or serious harm to a person’s health"
• "a serious and irreversible disruption of the management or operation of critical infrastructure"
• "the infringement of obligations under Union law intended to protect fundamental rights"
• "serious harm to property or the environment".

Point (c) refers to the infringement of obligations in Union law, not to an infringement of fundamental rights as such. See Article 3(49) — serious incident.

What are the AI Act serious-incident reporting deadlines?

Providers of high-risk AI systems in scope must report serious incidents immediately (EU AI Act Article 73). The outer limits start when the provider or, where applicable, the deployer "becomes aware of the serious incident":

• 15 days — the general limit, with the report due immediately after the provider establishes "a causal link" or its reasonable likelihood (Article 73(2)).
• 2 days — a widespread infringement or an Article 3(49)(b) serious incident (Article 73(3)).
• 10 days — a death, with the report due immediately after the provider or deployer establishes or suspects a causal relationship (Article 73(4)).

These are maximum limits, not waiting periods, and the period must "take account of the severity of the serious incident". An initial report that is not complete is permitted (Article 73(5)). Articles 73(9), 73(10) and 75(1a) change the recipient or the scope for some systems. Article 2(2) excludes Article 73 for Article 6(1) systems of Annex I Section B. Article 55(1)(c) is a different rule for models.

See Article 73(2)–(4) — reporting times.

Who reports an AI incident: the provider or deployer?

Article 73 gives the reporting obligation to the provider. Article 26(5) gives deployers different notification obligations and a fallback.

• Provider — reports to the market surveillance authorities of the Member States where the incident occurred (Article 73(1)). The AI Office (Article 75(1a)), equivalent reporting (Article 73(9)) and medical-device rules (Article 73(10)) change the recipient or the scope.
• Deployer — immediately informs "first the provider, and then the importer or distributor and the relevant market surveillance authorities". If the deployer is "not able to reach the provider", "Article 73 shall apply mutatis mutandis" (Article 26(5)).

Providers of general-purpose AI models with systemic risk report to the AI Office and, "as appropriate", national competent authorities (Article 55(1)(c)). This report does not replace an Article 73 report. Article 2(2) excludes Articles 26 and 73 for Article 6(1) systems of Annex I Section B.

See Article 73(1) and Article 26 — deployer monitoring.

What must deployers do after a high-risk AI system enters use?

Article 26 tells deployers of high-risk AI systems to monitor the operation and act on specified risks and incidents:

• Use — Use the system in accordance with the instructions for use (Article 26(1)).
• Oversight — Give human oversight to competent natural persons with "the necessary support" (Article 26(2)).
• Monitoring — Monitor the operation. "Where relevant", inform the provider (Article 26(5)).
• Risk — With "reason to consider" that use in accordance with the instructions can cause an Article 79(1) risk, inform the provider or distributor and the market surveillance authority "without undue delay". Then "suspend the use of that system" (Article 26(5)).
• Serious incident — Immediately inform the provider first, then the importer or distributor and the market surveillance authorities. If the provider cannot be reached, Article 73 applies "mutatis mutandis" (Article 26(5)).
• Logs — Keep the logs "under their control" for "a period appropriate to the intended purpose", a minimum of six months (Article 26(6)).

Other law can change the log period, and financial institutions and law enforcement deployers have special rules (Article 26(5) and (6)). Article 26 applies from 2 December 2027 or 2 August 2028 (Dates). Article 2(2) excludes Article 26 for Article 6(1) systems of Annex I Section B.

See Article 26 — deployer monitoring.

When do conformity assessment and post-market duties start?

EU AI Act Articles 40 to 49 (Chapter III Section 5) and Articles 72 and 73 (Chapter IX) apply from 2 August 2026, the general date of application (Article 113, second paragraph). Article 113, points (a) to (d), give them no different date.

Other rules change the start for some readers:

• Chapter III Sections 1 to 3 — Articles 6 to 27, but not Article 6(5), apply from 2 December 2027 (Annex III) or 2 August 2028 (Annex I) (Article 113, point (c)). Examples are Article 16, points (f) and (i), on conformity assessment and registration, and Article 26(5) on serious-incident notices.
• Older systems — Article 111(2) gives a transitional rule, with 2 August 2030 for systems intended for use by public authorities.
• Real-time remote biometric identification — Since 2 February 2025, Article 5(2) makes a completed FRIA and registration prerequisites, with an urgency rule for registration.
• Section B — Article 2(2) excludes these Articles for Article 6(1) systems of Annex I Section B.

See Dates.

What are the maximum EU AI Act fines?

Article 99 sets fines of up to EUR 35 000 000 or, for an undertaking, 7% of its total worldwide annual turnover for the preceding financial year, if that is higher. This maximum is for non-compliance with Article 5 (Article 99(3)). The other maximums also use the higher value:

• EUR 15 000 000 or 3% — non-compliance with Articles 16, 22, 23, 24, 25(2) and (4), 26, and 50, and with the notified-body provisions (Article 99(4)).
• EUR 7 500 000 or 1% — "incorrect, incomplete or misleading information" to notified bodies or authorities "in reply to a request" (Article 99(5)).

Other rules change the fine for some readers. SMEs, also start-ups, get the lower of the two values, and SMCs only for Article 99(4) and (5) (Article 99(6) and (6a)). Member States set the penalty rules, and rules on "to what extent" fines apply to public bodies (Article 99(1) and (8)). Each decision to impose a fine or not, and on its value, takes into account "all relevant circumstances of the specific situation" (Article 99(7)). The EDPS can fine Union institutions, bodies, offices and agencies up to EUR 1 500 000 for Article 5, and EUR 750 000 for other obligations (Article 100). The Commission can fine providers of general-purpose AI models up to 3% or EUR 15 000 000, the higher value, for specified intentional or negligent acts (Article 101).

See Penalties and enforcement.

Which Modulos Frameworks cover the EU AI Act?

Modulos gives the EU AI Act Requirements in two Frameworks: OFF-1 for the Organization and MFF-1 for the AI System. This allocation is a Modulos decision, not a level of legal obligation. The EU AI Act gives obligations to actors for each AI system or general-purpose AI model.

The Requirements for general-purpose AI models are in the two Frameworks. Make these decisions before you set up the Frameworks:

• the Article 2 scope and the Section B rule
• your role
• if you have an AI system, a general-purpose AI model, or the two
• for models, the systemic-risk test and the open-source exceptions.

See Select the Frameworks.

Where does Modulos map incident reporting, monitoring and the FRIA?

The Requirements and provisions table maps provider incident reporting, post-market monitoring, the FRIA, deployer monitoring and notification, and systemic-risk model incident reporting to Modulos Requirements.

A Requirement in Modulos does not make an obligation apply. The legal scope and the responsible actor come from the EU AI Act.

How do EU AI Act Requirement Reviews work in Modulos?

A Modulos Requirement Review examines the completed work at the Requirement level. The Requirement Owner examines the completed Controls and marks the Requirement Fulfilled.

A Fulfilled Requirement is a status in Modulos. It is not a conformity assessment, a declaration of conformity, or a decision of an authority. See Review the Requirements.

How should I scope EU AI Act work in Modulos?

Start with the legal role, scope, and classification decisions. Then set up the Frameworks. Modulos recommends these decisions:

• the Article 2 scope, with its exclusions and the Section B rule of Article 2(2)
• your role for each AI system or model (Article 3 and Article 25)
• the Article 5 and Article 50 tests for each AI system, and its high-risk classification (Article 6)
• for general-purpose AI models, the systemic-risk test (Article 51) and the open-source exceptions (Articles 53(2) and 54(6)).

Then add OFF-1 to the Organization and MFF-1 to each AI System. Examine model obligations and system obligations independently. See Set up the Organization and AI System.