[Reference] EU AI Act—Obligations and Application Timeline
Last updated: August 5, 2026
The AI Act (Regulation (EU) 2024/1689, commonly known as the AI Act) is an EU regulation that establishes a "risk-based" framework, varying the intensity of regulation according to the level of risk posed by AI.
As the third in this series of reference articles, I will organize the objectives, scope, main obligations, enforcement, and application schedule. This will be updated in accordance with institutional developments.
This article reflects the content of the amending regulation (Regulation (EU) 2026/1744, commonly known as the Digital Omnibus on AI) which entered into force on July 27, 2026. The background and specific issues of the amendment will be addressed in a separate article.
■ In a nutshell: The overall picture of the AI Act
The AI Act entered into force on August 1, 2024, and obligations will be applied in stages from February 2025 to August 2030.
While this series has covered the DMA and DSA, the AI Act differs in nature from them.
If the DMA is a "market law" dealing with fairness and contestability in digital markets, and the DSA is a "social law" dealing with the societal risks of online services, then the AI Act is a law based on the methodology of product safety regulation.
The fact that the AI Act is in the lineage of product safety regulation is indicated by the text itself. Recital 9 explicitly states that the regulation of high-risk AI is established in a manner consistent with the "New Legislative Framework," which consists of Regulation (EC) No 765/2008, Decision 768/2008/EC, and Regulation (EU) 2019/1020.
Conformity assessment (Article 43), EU declaration of conformity (Article 47), CE marking (Article 48), and market surveillance (Chapter 9) are methods the EU has used in the regulation of machinery, medical devices, and other products.
The characteristic of the AI Act lies in the fact that, in addition to health and safety, it places fundamental rights squarely as protected legal interests.
Recital 1 states that the objective is to protect public interests, including democracy, the rule of law, and environmental protection, at a high level while respecting the fundamental rights set out in the Charter of Fundamental Rights of the EU. Recital 48 lists human dignity, non-discrimination, freedom of expression, and workers' rights, noting that the degree of adverse impact on rights under the Charter is particularly important in determining high-risk classification.
The scope of what is protected has expanded, but the method of protection is close to product safety regulation. The importance of developing harmonized standards and common specifications for making high-risk AI system regulation function in practice stems from this structure.
▷ What is "AI"—AI systems and AI models
The AI Act does not contain a definition of "AI" itself. What are defined are "AI systems" and "general-purpose AI models."
An AI system is a machine-based system 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.
Recital 12 states that this definition should be based on the characteristics that distinguish AI systems from simple traditional software or programming approaches, and does not include systems that perform automatic processing following rules defined solely by natural persons.
The key characteristic is the "capability to infer." This includes not only machine learning but also logic- and knowledge-based approaches.
Because it is defined by function rather than technology, whether your company's software is subject to the regulation is determined not by whether you call it AI, but by whether it meets the requirements of Article 3(1).
On the other hand, there is no general definition of "AI model." What is defined is the general-purpose AI model (GPAI model).
The relationship between the two is explained in Recital 97. An AI model is an essential component of an AI system, but it does not constitute an AI system on its own. It only becomes an AI system when other components, such as a user interface, are added.
It is easy to understand if you think of a large language model itself as a 'model,' and the product or service that incorporates it as a 'system.'
When a company integrates its own GPAI model into its own AI system and places it on the market, both the obligations as a GPAI model provider and the obligations as an AI system provider may become relevant.
■ Scope: To whom does it apply?
▷ Role-based regulation
The AI Act is not a law that designates specific companies like the DMA.
Obligations are allocated according to 'roles,' such as provider, deployer, importer, distributor, and authorized representative.
Article 25 stipulates that for AI systems already placed on the market or put into service, distributors, importers, deployers, or other third parties shall be considered a new provider if they affix their name or trademark to a high-risk AI system, make a substantial modification to a high-risk AI system, or change the intended purpose of an AI system that was not previously classified as high-risk, making it a high-risk AI system.
Companies that procure external AI systems and provide them under their own brand may fall within this scope.
▷ Four tiers of risk and the GPAI model layer
The AI Act is generally explained in the following four tiers:
Tier 1: Prohibited practices (Article 5)
Tier 2: High-risk AI systems (Article 6)
Tier 3: AI systems subject to transparency obligations (Article 50)
Tier 4: Minimal or no-risk AI systems
Intersecting these vertically is the regulation concerning GPAI models (Chapter V).
Therefore, explaining it only by the 'four risk classifications' may lead to overlooking the relationship between the regulations concerning AI systems and those concerning GPAI models.
There are two pathways for high-risk AI systems.
One is for AI systems that are safety components of products, or are products themselves, covered by the EU product safety legislation listed in Annex I, and for which a third-party conformity assessment is required under the relevant product legislation (Article 6(1)).
The other is for standalone AI systems listed in the eight areas of Annex III (Article 6(2)).
The target areas are as follows.
Biometric identification
Critical infrastructure
Education and vocational training
Employment and worker management
Essential private and public services
Law enforcement
Migration, asylum, and border control
Administration of justice and democratic processes
Regarding the first pathway, the 2026 amendment clarified the definition and scope of 'safety components'.
Amended Article 3(14) defines a safety component as a component that fulfills a safety function for a product or AI system, or whose failure or malfunction endangers the health, safety, or property of persons and clarifies that 'fulfilling a safety function' refers to cases where the intended purpose is to prevent or mitigate risks to the health, safety, or property of persons.
Furthermore, AI systems used solely for user assistance, performance optimization, service efficiency, automation, convenience, or quality control aspects unrelated to safety are considered not to be safety components (Article 6(1a)). On the other hand, AI systems whose failure or malfunction endangers health or safety are considered safety components (Article 6(1b)).
The second pathway includes the filter in Article 6(3).
Even for AI systems falling under Annex III, they are not considered high-risk if they are limited to narrow procedural tasks, improving the results of tasks already completed by humans, detecting deviations from decision-making patterns, or preparatory tasks, and do not create a risk of significant harm to health, safety, or fundamental rights.
However, AI systems that perform profiling of natural persons are always considered high-risk.
Providers who determine that their system does not fall under high-risk are obligated to document that assessment before placing it on the market or putting it into service, and to register the AI system in the EU database (Article 6(4), Article 49(2)).
▷ Extraterritorial application: Where do Japanese companies fall under the scope?
Article 2(1) establishes three main pathways for application.
(a) Providers who place AI systems on the market or put them into service within the EU, or who place GPAI models on the EU market. This applies regardless of whether they are established within the EU or in a third country.
(b) Deployers of AI systems located within the EU.
(c) Providers and deployers in third countries, where the output generated by the AI system is used within the EU
The core of extraterritorial application is (a). As long as they place AI systems or GPAI models on the EU market, Japanese companies can also become direct addressees of the AI Act.
Recital 21 states that the AI Act applies in a non-discriminatory manner, regardless of the provider's place of establishment.
(c) is a provision to capture cases that fall outside of that.
Recital 22 cites the example of an operator within the EU commissioning AI-based processing to an operator in a third country, where the output is used within the EU. This is a provision to prevent the circumvention of the regulation through the outsourcing of processing outside the EU.
Providers of high-risk AI systems that do not have a place of business in the EU must, in principle, appoint an authorized representative (Article 22).
GPAI model providers outside the EU are also, in principle, subject to the obligation to appoint an authorized representative (Article 54). However, there are exceptions for certain free and open-source GPAI models. GPAI models that pose systemic risks are not subject to this exception.
■ Main Obligations
The addressees of the obligations differ for each article.
Prohibited Practices (Article 5): Both providers and deployers. Placing on the market, putting into service, or use are all subject to prohibition
High-Risk AI Systems (Articles 8–27): Separate obligations for providers and deployers, respectively. Importers and distributors also have obligations
Transparency Obligations (Article 50): Varies by paragraph. In principle, (1) and (2) are for providers, and (3) and (4) are for deployers
GPAI Models (Articles 53 and 55): Providers of GPAI models
▷ ① Prohibited Practices (Article 5)
As of February 2, 2025, the following eight categories are, in principle, prohibited:
Subliminal, intentionally manipulative, or deceptive techniques to distort behavior significantly
Exploitation of vulnerabilities due to age, disability, or specific social or economic situations
Social scoring
Risk assessment or prediction of criminal behavior based solely on profiling or evaluation of personality traits
Untargeted scraping of facial images from the internet or CCTV footage to create or expand facial recognition databases
Emotion inference in workplaces and educational institutions
Biometric categorization to infer sensitive attributes
Real-time remote biometric identification in publicly accessible spaces for law enforcement purposes
Some categories have limited exceptions under the articles.
In the 2026 amendment, certain AI systems that generate or process specific harmful synthetic content without the individual's consent, or content involving serious rights violations against children, were added to the prohibited practices. For details, please refer to the amended Article 5(1)(ba)(bb) and the same article (1a)(1b).
For providers, this applies when such generation or processing is the intended purpose of the AI system, or when it is a reasonably foreseeable and reproducible result, yet reasonable and appropriate safety measures have not been taken.
For deployers, the use of AI systems for the purpose of generating or processing such material is prohibited.
These new prohibited practices will apply from December 2, 2026.
▷ (2) High-Risk AI Systems (Articles 8–27)
Providers of high-risk AI systems are primarily subject to the following obligations:
Risk management system (Article 9)
Data and data governance (Article 10)
Technical documentation (Article 11)
Logging (Article 12)
Transparency and provision of information to deployers (Article 13)
Design of human oversight (Article 14)
Accuracy, robustness, and cybersecurity (Article 15)
Quality management system (Article 17)
Conformity assessment (Article 43)
EU declaration of conformity (Article 47)
CE Marking (Article 48)
Registration (Article 49)
Deployers also have independent obligations (Article 26).
Specifically, these include using the system in accordance with the instructions for use, ensuring relevance and sufficient representativeness in light of the intended purpose when controlling input data, assigning human oversight to natural persons with the necessary competence, training, and authority, and maintaining logs. When used in the workplace, prior notification to the affected workers and their representatives is also required.
In addition, deployers who use high-risk AI systems for public authorities, private entities providing public services, creditworthiness assessment or credit scoring of natural persons, or risk assessment and pricing for life and health insurance must, in principle, conduct a Fundamental Rights Impact Assessment (FRIA) (Article 27).
The obligations of importers and distributors do not end with mere verification.
In addition to verification before placing on the market, importers are required to display their name, registered trade name or trademark, and contact address, and to keep a copy of the certificate of conformity for 10 years after placing the system on the market (Article 23).
If a distributor has reason to believe that an AI system is not in conformity with the AI Act after it has been placed on the market, they must either take the necessary corrective measures, withdrawal, or recall themselves, or ensure that the provider, importer, or other relevant operators do so (Article 24).
▷ (3) Transparency Obligations (Article 50)
The addressees of the obligations in Article 50 differ by paragraph.
Provider Obligations
For AI systems intended to interact directly with natural persons, design them to inform the user that they are interacting with an AI system. This does not apply if it is obvious from the circumstances and context to a natural person (Article 50(1)).
For AI systems that generate synthetic audio, image, video, or text, mark the output in a machine-readable format and make it detectable as artificially generated or manipulated (Ibid. (2)).
Deployer Obligations
When using emotion recognition systems or biometric categorization systems, inform the natural persons concerned (Ibid. (3)).
When generating or manipulating images, audio, or video that constitute a deepfake, disclose that fact. If it constitutes part of an artistic, creative, satirical, or fictional work, disclosure in an appropriate manner that does not hinder the display or enjoyment of the work is sufficient (Ibid. (4) first paragraph).
When AI generates or manipulates text published for the purpose of informing the public on matters of public interest, disclose that fact. However, this does not apply if the content has undergone human review or editorial control and there is a natural or legal person who holds editorial responsibility for the publication (Ibid. (4) second paragraph).
Companies that incorporate generative AI into their own services and provide them may fall under Article 50(2), while companies that use generative AI in their business to publish content may fall under Article 50(4).
Even when both are 'using generative AI,' their positions under the law differ.
Article 50 will apply from August 2, 2026.
However, for AI systems that generate synthetic audio, image, video, or text content placed on the market before August 2, 2026, the compliance deadline for Article 50(2) is set for December 2, 2026 (Article 111(4)).
Please see below for details on transparency obligations.
▷ ④ GPAI Model Providers (Articles 53 and 55)
In principle, all GPAI model providers are subject to the following obligations (Article 53):
Preparation and maintenance of technical documentation
Provision of information and documentation to downstream AI system providers
Development of a policy to comply with EU copyright law
Preparation and publication of a sufficiently detailed summary of the content used for training
Certain free and open-source GPAI models are exempt from the obligations regarding technical documentation and the provision of information to downstream providers.
However, they are not exempt from the policy to comply with EU copyright law or the publication of a summary of training content. Furthermore, this open-source exemption does not apply to GPAI models that present systemic risks.
GPAI models with systemic risks are subject to the following additional obligations (Article 55):
Model evaluation based on standardized protocols and tools
Adversarial testing
Assessment and mitigation of systemic risks at the EU level
Tracking, documenting, and reporting of serious incidents
Adequate cybersecurity protection
GPAI models for which the cumulative amount of computation used for training exceeds 10^25 FLOPs are presumed to have high-impact capabilities (Article 51).
Systemic risk as referred to here (Article 3(65)) refers to risks specific to the high-impact capabilities of GPAI models.
■ Enforcement Framework and Sanctions
Enforcement is divided between member state authorities and the European Commission. Within the European Commission, the AI Office plays a central role.
▷ Market Surveillance Authorities of Member States
In principle, the supervision and enforcement of AI systems are carried out by the market surveillance authorities of each Member State.
The deadline for Member States to designate at least one market surveillance authority and one notifying authority was August 2, 2025.
The preamble to the 2026 amendment explains that, in addition to delays in the preparation of harmonized standards, delays in establishing governance and conformity assessment frameworks at the Member State level have created a compliance burden greater than anticipated.
▷ European Commission / AI Office
For GPAI models, the European Commission has exclusive supervisory and enforcement powers (Article 88). In practice, the AI Office plays a central role.
Following the 2026 amendment, the AI Office also has exclusive supervisory and enforcement powers for the following AI systems (Article 75(1)).
AI systems based on GPAI models, where the model and the system are developed by the same provider or by providers belonging to the same corporate group
AI systems that constitute VLOPs/VLOSEs under the DSA, or AI systems integrated into VLOPs/VLOSEs
However, for the former, exceptions are made for product-related AI systems in Annex I, critical infrastructure in Annex III, point 2, certain AI systems in the fields of law enforcement, border control, and finance, and certain AI systems related to the administration of justice.
Furthermore, the exclusive jurisdiction of the AI Office generally extends to the provider of the subject AI system and to deployers who are the provider themselves or belong to the same corporate group as the provider. Other deployers generally remain under the supervision of Member State authorities.
In addition, the AI Office has been granted powers as a market surveillance authority, including the power to request information, request access to AI systems, and conduct remote or on-site inspections (Article 75a).
A procedural framework for initiating investigations, notifying preliminary findings, commitments, non-compliance decisions, fines, and periodic penalty payments has also been established (Articles 75b to 75d).
▷ Fines
The maximum amounts for major fines are as follows. If the violator is a company, the maximum is generally the higher of a fixed amount or a percentage of the total worldwide annual turnover of the preceding financial year.
Violation of prohibited practices under Article 5: The higher of 7% of total worldwide annual turnover or 35 million euros (Article 99(3))
Violation of other listed obligations: The higher of 3% of total worldwide annual turnover or 15 million euros (Article 99(4))
Provision of incorrect, incomplete, or misleading information: The higher of 1% of total worldwide annual turnover or 7.5 million euros (Article 99(5))
Violation of obligations by GPAI model providers: The higher of 3% of total worldwide annual turnover or 15 million euros (Article 101)
For SMEs and startups, the maximum fines under Article 99(3) to (5) are, in principle, the lower of the turnover percentage or the fixed amount.
With the 2026 amendment, for Small and Medium-sized Companies (SMCs), the fines under Article 99(4) and (5) are capped at the lower of either the percentage of turnover or the fixed amount. However, this special provision for SMCs has not been extended to violations of prohibited practices under Article 99(3).
■ Application Schedule
August 1, 2024: Entry into force of the AI Act
February 2, 2025: General provisions, initial 8 categories of prohibited practices, and AI literacy
August 2, 2025: Notifying authorities and notified bodies, GPAI models, governance, and member state penalty provisions, etc.
July 27, 2026: Entry into force of Regulation (EU) 2026/1744.
August 2, 2026: General application. Transparency obligations under Article 50, market surveillance under Chapter 9, and the European Commission's sanctioning powers regarding GPAI models, etc.
December 2, 2026: Two new categories of prohibited practices. Deadline for compliance with Article 50(2) for generative AI systems placed on the market before August 2, 2026.
August 2, 2027: Deadline for member states to begin operation of AI regulatory sandboxes. Compliance deadline for GPAI models placed on the EU market before August 2, 2025.
December 2, 2027: Application of Chapter 3, Sections 1 to 3 regarding standalone high-risk AI systems in Annex III.
August 2, 2028: Application of Chapter 3, Sections 1 to 3 regarding product-embedded high-risk AI systems in Annex I.
August 2, 2030: Compliance deadline for existing high-risk AI systems intended for use by public authorities.
■ Summary
The AI Act is a regulation that governs the risks of AI to health, safety, and fundamental rights, based on the methodology of product safety regulations.
Obligations are role-based. The applicable articles change depending on whether one is classified as a provider or a deployer.
If the output of an AI system is used within the EU, operators outside the EU may also be subject to the Act (Article 2(1)(c)).
The structure consists of four layers of AI system risk, intersected vertically by regulations concerning GPAI models.
Enforcement is, in principle, carried out by member state market surveillance authorities. GPAI models and certain AI systems falling under Article 75(1) are exclusively supervised and enforced by the European Commission and the AI Office.
While the general application date of August 2, 2026, has been maintained, the application dates for Chapter III, Sections 1 to 3 regarding high-risk AI have been postponed to December 2, 2027, for Annex III types and August 2, 2028, for Annex I types.
There are two points of focus from August 2, 2026, onwards.
First is how the AI Office will exercise its new enforcement powers.
With the scope expanded beyond GPAI models to include AI systems based on certain GPAI models, as well as AI systems that constitute or are integrated into VLOPs/VLOSEs under the DSA, the supervisory relationship when both the AI Act and the DSA apply to the same service has been institutionally clarified.
How this will work in practice remains to be seen.
Second is the development of harmonized standards and other practical means of compliance.
Since the delay in preparing harmonized standards was one of the primary reasons for postponing the application of high-risk AI regulations, it will be crucial whether harmonized standards and guidelines are developed as scheduled for the application dates of December 2027 and August 2028.
■ List of Primary Sources and Official Documents
▼ Legal Texts
AI Act Full Text
Regulation (EU) 2024/1689, OJ L, 2024/1689, 12.7.2024
https://eur-lex.europa.eu/eli/reg/2024/1689/oj
2026 Amending Regulation (Digital Omnibus on AI)
Regulation (EU) 2026/1744, OJ L, 2026/1744, 24.7.2026
https://eur-lex.europa.eu/eli/reg/2026/1744/oj
▼ European Commission Guidelines
Guidelines on the definition of an artificial intelligence system
Guidelines on the definition of an artificial intelligence system established by Regulation (EU) 2024/1689
This document explains the definition of an AI system under Article 3(1) of the AI Act, including its constituent elements and how it is distinguished from conventional software. It serves as a starting point for considering whether your company's software qualifies as an 'AI system' under the AI Act.
https://ai-act-service-desk.ec.europa.eu/sites/default/files/2026-01/guide-definition_en.pdf
Guidelines on prohibited artificial intelligence practices
Guidelines on prohibited artificial intelligence practices established by Regulation (EU) 2024/1689
This document explains the eight categories of prohibited practices originally set out in Article 5 of the AI Act, including their scope, constituent elements, exceptions, and specific examples. However, as it was published in February 2025, it does not reflect the new prohibited practices added by the 2026 amendment.
https://ai-act-service-desk.ec.europa.eu/sites/default/files/2026-01/guide-prohibited_en.pdf
Guidelines on the scope of the obligations for providers of general-purpose AI models
Guidelines on the scope of the obligations for providers of general-purpose AI models under the AI Act
This document explains which models qualify as GPAI models, who is considered a provider of a GPAI model, under what circumstances a person who modifies a model becomes a provider, and the extent of the free and open-source model exception. It is an essential document for reading Articles 53, 54, and 55 regarding GPAI models.
https://digital-strategy.ec.europa.eu/en/library/guidelines-scope-obligations-providers-general-purpose-ai-models-under-ai-act
▼ Practical Materials for GPAI Models
General-Purpose AI Code of Practice
This is an optional practical tool for GPAI model providers to comply with their obligations under the AI Act. It consists of the following three chapters:
Transparency
Copyright
Safety and Security
The Transparency and Copyright chapters apply to all GPAI model providers. The Safety and Security chapter is intended for providers of GPAI models that pose systemic risks.
While this Code is not legislation, the European Commission and the European AI Board have confirmed it as an appropriate voluntary means for GPAI model providers to demonstrate compliance with their obligations under the AI Act.
https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai
▼ Comprehensive Information and Practical Support
European Commission "AI Act"
Regulatory framework for AI
This is the European Commission's comprehensive page summarizing the AI Act's risk classification, GPAI model regulation, enforcement structure, application schedule, and an overview of the 2026 amendments.
https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
European Commission "Navigating the AI Act"
This explains the scope of the AI Act, high-risk AI systems, GPAI models, fundamental rights impact assessments, standardization, governance, and enforcement in a Q&A format.
https://digital-strategy.ec.europa.eu/en/faqs/navigating-ai-act
AI Act Service Desk / Single Information Platform
This is a practical support site from the European Commission where you can check the text of the AI Act, related materials, and Q&As all in one place.
https://ai-act-service-desk.ec.europa.eu/
European AI Office
Information regarding the organization and tasks of the AI Office, supervision of GPAI models, and the Code of Practice is published here.
https://digital-strategy.ec.europa.eu/en/policies/ai-office
List of upcoming guidelines
Supporting the implementation of the AI Act with clear guidelines
The European Commission is preparing additional guidelines on high-risk classification, transparency obligations under Article 50, serious incident reporting, requirements for high-risk AI, obligations of providers and deployers, fundamental rights impact assessments, responsibilities in the AI value chain, substantial modifications, post-market monitoring, and the relationship between the AI Act and data protection laws.
https://digital-strategy.ec.europa.eu/en/news/supporting-implementation-ai-act-clear-guidelines
▷ Points to note when reading official documents
The European Commission's guidelines, Q&As, Code of Practice, and the AI Act Service Desk are official materials useful for understanding the provisions and practical implementation.
However, guidelines and Q&As are not the AI Act itself and, in principle, do not have legal binding force. The final authority for interpreting the law lies with the Court of Justice of the European Union. Especially after the 2026 amendments, depending on the publication or update date of the materials, the amendments may not yet be reflected.
Additionally, the AI Act Service Desk may display pre-amendment versions of articles in some sections. For example, when checking Article 75, the exclusive jurisdiction and enforcement powers of the AI Office as amended in 2026 may not be reflected. For the post-amendment content, it is safer to check Regulation (EU) 2026/1744 directly on EUR-Lex.
*This article represents the author's personal views and does not represent the official views of the organization to which the author belongs. It is based on publicly available information and the author's own knowledge and experience, and does not contain any non-public information from the author's organization. Please consult with a professional for specific decisions.
*On X, I share the latest news and primary information regarding EU digital policy and regulation.