前文

THE EUROPEAN PARLIAMENT AND THE COUNCIL OF THE EUROPEAN UNION,

Having regard to the Treaty on the Functioning of the European Union, and in particular Article 114 thereof,

Having regard to the proposal from the European Commission,

After transmission of the draft legislative act to the national parliaments,

Having regard to the opinion of the European Central Bank,

Having regard to the opinion of the European Economic and Social Committee,

Having regard to the opinion of the Committee of the Regions,

Acting in accordance with the ordinary legislative procedure,

Whereas:

(1)Regulation (EU) 2024/1689 of the European Parliament and of the Council lays down harmonised rules on artificial intelligence (AI) and aims to improve the functioning of the internal market, to promote the uptake of human-centric and trustworthy AI, while ensuring a high level of protection of health, safety and fundamental rights and supporting innovation. Regulation (EU) 2024/1689 entered into force on 1 August 2024. The entry into application of its provisions is staggered, with all rules entering into application by 2 August 2027.

(2)The experience gathered from the implementation of the parts of Regulation (EU) 2024/1689 that already apply can inform the implementation of those parts that are yet to apply. In this context, the delayed preparation of standards, which provide technical solutions for providers of high-risk AI systems to ensure compliance with their obligations under that Regulation, and the delayed establishment of the governance and the conformity assessment frameworks at national level has resulted in a compliance burden that is heavier than expected. In addition, consultations of stakeholders have revealed the need for additional measures to facilitate and provide clarification on implementation and compliance, without reducing the level of protection for health, safety and fundamental rights from AI-related risks that the provisions of Regulation (EU) 2024/1689 seek to achieve.

(3)Targeted amendments to Regulation (EU) 2024/1689 are necessary to address certain implementation challenges, with a view to the effective, simple and uniform application of the relevant rules.

(4)In order to enable AI innovation in the private and public sectors, it is important that the Commission and Member States’ competent authorities ensure that the supervision, enforcement and monitoring of sectoral and national laws do not create overlaps, inconsistent interpretations or divergent enforcement.

(5)Regulation (EU) 2024/1689 lays down horizontal rules for AI systems to ensure a consistent and high level of protection of public interests as regards health safety and fundamental rights. For high-risk AI systems referred to in Article 6(1), that Regulation applies in conjunction with the Union harmonisation legislation listed in Section A of Annex I. In certain cases, it is possible for such Union harmonisation legislation to lay down requirements that achieve the same or a higher level of protection of relevant public interests as achieved by the specific requirements or obligations laid down in Regulation (EU) 2024/1689. In that case, it should be possible to limit the application of specific requirements or obligations laid down in Regulation (EU) 2024/1689 in order to facilitate compliance, minimise administrative burden and duplications, while preserving the level of protection ensured by that Regulation. Such limitation should be possible where, and to the extent that, the Union harmonisation legislation listed in Section A of Annex I lays down requirements providing for an equivalent level of protection of health, safety or fundamental rights as the requirement or obligation concerned. The Commission should be empowered to adopt delegated acts to supplement that Regulation by identifying such cases and specifying the products concerned, the requirements or obligations that may be limited, and the conditions and scope of any limitation, ensuring that the level of protection provided by Regulation (EU) 2024/1689 is not reduced.

(6)Most Union companies are small and medium-sized enterprises, the majority of which are micro and small enterprises. Enterprises outgrowing the micro, small and medium-sized enterprises (SMEs) definition, namely the ‘small mid-cap enterprises’ (SMCs), play a vital role in the Union’s economy. Compared to SMEs, SMCs tend to demonstrate a faster pace of growth, and a higher level of innovation and digitisation. Nevertheless, they face challenges similar to SMEs in relation to administrative burden, leading to a need for proportionality in the implementation of Regulation (EU) 2024/1689 and for targeted support. To enable the smooth transition of enterprises from SMEs to SMCs, it is important to address in a coherent manner the effect that Regulation may have on their activity once those enterprises become SMCs and are faced with rules that apply to large enterprises. Regulation (EU) 2024/1689 provides for several measures for small-scale operators, which should be extended to SMCs where appropriate while safeguarding the overarching objectives and level of protection afforded under Regulation (EU) 2024/1689. In order to clarify the treatment of SMEs and SMCs in Regulation (EU) 2024/1689, it is necessary to introduce definitions for SMEs and SMCs, which should correspond to the definition set out in the Annex to Commission Recommendation 2003/361/EC and Annex to Commission Recommendation (EU) 2025/1099, respectively.

(7)The notion of a ‘safety component’ is crucial for the classification of certain AI systems as high-risk according to Regulation (EU) 2024/1689. Thus, the notion should be targeted to include only the AI systems which could have an adverse impact on the health and safety of persons or property, in line with the risk-based approach of that Regulation. The definition set out in Article 3, point (14) of Regulation (EU) 2024/1689 does not provide the necessary clarity to allow providers of AI systems to determine whether an AI system qualifies as a safety component and, as a result, risks extending the high-risk classification of AI systems beyond what is justified by the risk-based approach of that Regulation. It is therefore necessary to amend that definition. First, it is necessary to provide clarity on the concept of a ‘safety function’. The safety function should be an intended purpose of the system, which is determined by the provider of the system. An AI system fulfils a safety function where its intended purpose, as determined by the provider, is to prevent or mitigate risks to the health and safety of persons or property. In particular, this does not include AI systems which are intended to solely fulfil functions related to user assistance, performance optimisation, service efficiency, automation, convenience, or non-safety related aspects for quality control operations. The mere fact that an AI system is integrated into or operates within a product that is subject to Union harmonisation legislation does not, in itself, mean that it fulfils a safety function.

(8)Article 4 of Regulation (EU) 2024/1689 currently imposes an obligation on all providers and deployers of AI systems to ensure AI literacy of their staff. Development of AI literacy starting from education and training and continuing in a lifelong learning manner is crucial to equip providers, deployers and other affected persons with the necessary skills to make informed decisions regarding the deployment of AI systems. However, experience shared by stakeholders reveals that a solution imposing stringent obligations to ensure a sufficient level of AI literacy would not be suitable for all types of providers and deployers in relation to the promotion of AI literacy. Moreover, data indicates that imposing such obligations creates an additional compliance burden, particularly for smaller enterprises, whereas AI literacy should be a strategic priority, regardless of regulatory obligations and potential sanctions. In light of that information, Article 4 of Regulation (EU) 2024/1689 should be amended to require providers and deployers to take measures to support the development of AI literacy of their staff and of other persons dealing with the operation and use of AI systems on their behalf. The Commission and Member States should support and facilitate the efforts of providers and deployers of AI systems, including through offering training opportunities, providing informational resources, and allowing exchange of good practices and other initiatives. When complying with this obligation, the Commission and the Member States could take into account the European competence frameworks, for example the Digital Competence Framework for Citizens (DigComp) and the AI Literacy Framework for Primary and Secondary Education. The European Artificial Intelligence Board (the Board) should support the Commission and Member States by adopting recommendations setting out common objectives to be achieved in order to meet their obligation and ensure regular exchange between the Commission and Member States on the topic, while the Apply AI Alliance should allow discussion with the wider community.

(9)Bias detection and correction constitute a substantial public interest because they protect natural persons from the adverse effects of biases, including discrimination. For that reason, Regulation (EU) 2024/1689 provides a legal basis authorising providers of high-risk AI systems to process special categories of personal data in certain exceptional cases and subject to strict safeguards. This legal basis is linked to the obligation on those providers to establish practices concerning the detection, prevention and correction of biases likely to affect the health and safety of persons, have a negative impact on fundamental rights or lead to discrimination prohibited under Union law. Nevertheless, biases likely to have those effects could also result from the actions of the deployers of high-risk AI systems. Furthermore, such biases could also arise in the case of other AI systems or models. For example, biases in eligibility or risk-scoring tools used to assess applications for various types of public permits or licences can restrict rights or effectively prevent certain groups from accessing public services. Accordingly, a substantial public interest exists to allow, exceptionally and where strictly necessary, the processing of special categories of personal data for the purposes of bias detection and correction to providers and deployers of other AI systems and models and deployers of high-risk AI systems. It is therefore necessary to extend the legal basis, established pursuant to Regulation (EU) 2024/1689, to those providers and deployers. That legal basis should be subject to the same limitations, conditions and safeguards that apply in accordance with the existing Article 10(5) of that Regulation, thereby ensuring compliance with Article 9(2), point (g) of Regulation (EU) 2016/679 of the European Parliament and of the Council, Article 10(2), point (g) of Regulation (EU) 2018/1725 of the European Parliament and of the Council and Article 10, point (a) of Directive (EU) 2016/680 of the European Parliament and of the Council. Furthermore, to enable providers of high-risk AI systems to lawfully undertake bias detection and correction activities in preparation for compliance with the requirements for high-risk AI systems, including Article 10(2), points (f) and (g), of Regulation (EU) 2024/1689, the legal basis established by Article 4a of Regulation (EU) 2024/1689 should apply from the date of entry into application of that Regulation.

(10)Article 5 of Regulation (EU) 2024/1689 prohibits certain practices of AI systems that are particularly harmful and abusive, contradict certain Union values and violate certain fundamental rights. Article 5 is to be kept under review, as laid down in Article 112(1) of that Regulation. In light of technological and societal developments since the adoption of that Regulation, including the deployment and widespread use of AI systems generating non-consensual intimate images, videos, audio and similar material (‘non-consensual intimate material’) and child sexual abuse material, it is necessary to amend Article 5 of that Regulation.

(11)Non-consensual intimate material constitutes sexual violence and abuse against individuals, in particular women. AI systems that generate or manipulate such material pose a severe risk to health safety and fundamental rights, including victims’ human dignity, personal autonomy, integrity and private life, with potentially serious lasting psychological and other harms, and enable abuse at scale. The proliferation of such technologies, often described as ‘nudification’ applications, has created an urgent need for an explicit regulatory prohibition. Child sexual abuse material, including wholly or partially synthetic material, constitutes a grave threat to the safety and fundamental rights of children. AI systems generating or manipulating such material pose a grave risk to human dignity and the rights of the child, and risk normalising, amplifying and perpetuating sexual violence against children. Accordingly, an amendment to Article 5 of Regulation (EU) 2024/1689 is necessary both to protect women, children, other individuals and society from seriously harmful practices, thereby pursuing the objectives of that Regulation, and to bring clarity to providers and deployers as to the scope of their obligations, thereby addressing implementation challenges.

(12)It is necessary to define clearly the scope of the prohibition, including in particular the extent of providers’ and deployers’ obligations. This prohibition should not prevent providers from developing the technical capabilities of AI systems to generate or manipulate images, videos, audio or similar material. As concerns providers, the prohibition should be limited to the placing on the market or the putting into service of AI systems that generate or manipulate non-consensual intimate material or child sexual abuse material in two cases. First, it should cover systems intended to generate or manipulate such material. Second, it should cover systems where such generation or manipulation is a reasonably foreseeable and reproducible outcome and there are no reasonable and adequate technical safety measures and other safeguards in place, taking into account reasonably foreseeable misuse, to reliably prevent, and where necessary correct, that outcome and correct observed or reported misuse, including circumvention of such measures. Technical measures and other safeguards to prevent the generation of such material could include data cleaning, refusal training, safe prompt design and output controls, runtime prompt guardrails, content classification and filtering mechanisms, usage restrictions, abuse detection mechanisms, and notice and action mechanisms. Such preventive measures should be reasonable for the specific AI system, and are considered adequate if they align with the state-of-the-art measures and demonstrably prevent or sufficiently reduce in each specific case the likelihood of generating or manipulating such material, taking into account known and reasonably foreseeable misuse, including reasonably foreseeable circumvention of the preventive measures without significant technical modification. For providers retaining effective control over AI systems, for instance through a platform or a web interface, that could include following and reporting methods for misuse cases in full compliance with Union privacy and data protection law. In cases of observed or reported circumvention of the preventive measures or other safeguards, adequate corrective measures should be taken provided that such measures are reasonable, taking into account the specific AI system, including its release and distribution strategy (such as open-source releases). As concerns deployers, the use of an AI system should be prohibited only where the deployer uses an AI system for the purpose of generating or manipulating non-consensual intimate material or child sexual abuse material. This includes cases where a deployer uses or misuses AI systems placed on the market or put into service that lack reasonable and adequate preventive measures or where a deployer circumvents the preventive measures or uses for such purposes lawful AI systems not intended to generate or manipulate such material. The prohibition on use therefore does not cover the use of an AI system for lawful purposes, such as the generation or manipulation of material other than non-consensual intimate material or child sexual abuse material, even in cases where the AI system lacks reasonable and adequate safeguards that should have been put in place by the provider, nor does it cover accidental generation or manipulation of such content. Concerning the prohibition regarding non-consensual intimate material, where an AI system is intended for generation or manipulation of material falling under this prohibition, measures and other safeguards should include appropriate means for the distribution of the AI system aimed at enabling the reliable collection and demonstration of consent of the depicted person to such generation or manipulation, in compliance with Regulation (EU) 2016/679. The prohibition regarding non-consensual intimate material should be limited to realistic depictions of intimate parts, notably the genitals, pubic area, anus, exposed buttocks or exposed female breasts, nipples or areolae, or of sexually explicit activity. This ‘realism’ refers to the depiction of the person’s face, voice or their body in a credible real-life manner, regardless of the realism of the context of that depiction or whether it fully corresponds to the actual voice or appearance of the depicted person. Conversely, it excludes cartoonish or physically impossible depictions of a person’s body. The prohibition of non-consensual intimate material does not affect the generation or manipulation of other forms of nude material, such as material that does not depict identifiable natural persons, realistic partially nude depictions where intimate parts are not revealed and sexually explicit activities are not depicted, or non-realistic artistic nude works that do not realistically depict identifiable natural persons engaged in sexually explicit activity or depict their intimate parts. It also does not cover generative AI applications where intimate parts are not exposed or, if exposed, this is subject to the freely given, specific, informed, unambiguous and explicit consent of the depicted person (for example, try-on applications and medical applications, such as medical anatomical simulations and mammograms); this prohibition does not preclude the exceptional use of AI systems generating or manipulating nude depictions of the intimate parts of an identifiable person, in accordance with fundamental rights law, including data protection law, and applicable medical law, for the purpose of medical diagnosis and treatment by medical professionals where the person concerned is incapable of consent (for instance in an emergency situation). Finally, the prohibition on ‘manipulating’ non-consensual intimate material excludes cases where pre-existing intimate material is manipulated in a way that does not increase the exposure of any depicted intimate parts or alter the nature of any depicted sexually explicit activities, for instance the mere enhancement of an existing image depicting intimate parts or video depicting sexually explicit activities, such as changing the background, adding a text heading or enhancing the contrast or the brightness. Conversely, any manipulation of material, including material that already depicts intimate parts or sexually explicit activity, that increases the level of exposure of any depicted intimate parts or alters the nature of any depicted sexually explicit activities, falls under the scope of this prohibition.

(13)The prohibition on child sexual abuse material should not prevent the placing on the market, putting into service or use of an AI system where a ‘without right’ defence applies under national law, as referred to in Article 5(1) of Directive 2011/93/EU of the European Parliament and of the Council. This includes activities carried out under domestic legal powers, such as the legitimate generation, or manipulation of child sexual abuse material by the authorities in order to conduct criminal proceedings or to prevent, detect or investigate crime, as well as the legitimate use of the AI system in the context of red-teaming and evaluation activities for the purpose of assessing the system’s compliance with the prohibition laid down in this Regulation.

(14)These prohibitions constitute justified interferences with the freedom of expression and information and the freedom to conduct a business. They pursue objectives of general interest and protect the rights and freedoms of others, including those set out in Article 1, Article 3(1) and Articles 4, 7, 8, 21, 23 and 24 of the Charter of Fundamental Rights of the European Union (the Charter). They are closely tailored. In the case of intimate material, they are limited to realistic depictions of identifiable natural persons and to cases where the AI system is used to increase the level of nudity or explicitness; and exclude generation or manipulation with the person’s consent. Moreover, they are limited to requiring providers to implement ‘reasonable and adequate’ measures and safeguards in the case of systems not intended to generate or manipulate the prohibited material. They are also aligned with existing Union law, including Directive 2011/93/EU and Directive (EU) 2024/1385 of the European Parliament and of the Council. The interferences respect the essence of Articles 11 and 16 of the Charter, are prescribed by law, and are proportionate.

(15)The conduct covered by these prohibitions may also violate other law, including criminal law. The prohibitions do not preclude prosecution under such law. However, insofar as an infringement of the prohibitions may result in the imposition of penalties of a criminal nature, which may be laid down pursuant to Article 99(1) of Regulation (EU) 2024/1689, and to the extent that the same conduct is sanctioned under criminal law, including criminal law falling within the scope of Directives 2011/93/EU and (EU) 2024/1385, Member States are required to ensure respect for the ne bis in idem principle in accordance with the Charter.

(16)The prohibitions are without prejudice to remedies available under national laws for individuals to protect their fundamental rights, including rights to their image, privacy and human dignity.

(17)In order to ensure consistency, avoid duplication and minimise administrative burdens in relation to the procedure for the designation of notified bodies pursuant to Regulation (EU) 2024/1689, while maintaining the same level of scrutiny, a single application and a unified assessment procedure should be available for new conformity assessment bodies and notified bodies which are designated pursuant to the Union harmonisation legislation listed in Section A of Annex I to Regulation (EU) 2024/1689, such as Regulations (EU) 2017/745 and (EU) 2017/746 of the European Parliament and of the Council, where such an application and procedure is established pursuant to that Union harmonisation legislation. The single application and unified assessment procedure aims to facilitate, support and expedite the designation procedure in accordance with Regulation (EU) 2024/1689, while ensuring compliance with the requirements applicable to notified bodies in accordance with that Regulation and the Union harmonisation legislation listed in Section A of Annex I thereto. The unified assessment procedure should be carried out with respect to the tasks and responsibilities of the authorities involved. Moreover, it should be clarified that a conformity assessment body that is designated pursuant to more than one piece of Union harmonisation legislation listed in Section A of Annex I should have to apply only once to be designated pursuant to this Regulation.

(18)With a view to ensuring the smooth application and consistency of Regulation (EU) 2024/1689, amendments should be made to it. A technical correction to Article 43(3), first subparagraph, of Regulation (EU) 2024/1689 should be made to align the conformity assessment requirements with the requirements of providers of high-risk AI systems in Article 16 of that Regulation. Moreover, it should be clarified that where a provider of a high-risk AI system is subject to the conformity assessment procedure under Union harmonisation legislation listed in Section A of Annex I to Regulation (EU) 2024/1689, and the conformity assessment extends to the compliance of the quality management system of that Regulation and of such Union harmonisation legislation, the provider should be able to include aspects related to quality management systems pursuant to that Regulation as part of the quality management systems pursuant to such Union harmonisation legislation, in accordance with Article 17(3) of Regulation (EU) 2024/1689. Article 43(3), second subparagraph, of that Regulation should be amended to clarify that notified bodies which have been notified pursuant to the Union harmonisation legislation listed in Section A of Annex I to Regulation (EU) 2024/1689 and which aim to assess high-risk AI systems covered by the Union harmonisation legislation listed in Section A of Annex I to that Regulation, should have the power to assess the conformity of high-risk AI systems under certain conditions for 18 months from 27 July 2026. This amendment is without prejudice to Article 28 of Regulation (EU) 2024/1689, thus conformity assessment bodies that wish to be designated and notified pursuant to that Regulation can submit an application at any time during and after these 18 months. Moreover, Regulation (EU) 2024/1689 should be amended to clarify that where a high-risk AI system is both covered by the Union harmonisation legislation listed in Section A of Annex I to Regulation (EU) 2024/1689 and falls within one of the use-cases listed in Annex III to that Regulation, the provider should follow the relevant conformity assessment procedure as required under that relevant harmonisation legislation.

(19)Regulation (EU) 2024/1689 and Regulation (EU) 2024/2847 of the European Parliament and of the Council complement each other so that the safety and cybersecurity of products with digital elements is ensured. Article 12 of Regulation (EU) 2024/2847 states that where high-risk AI systems and processes put in place by manufacturers fulfil the essential cybersecurity requirements set out in Regulation (EU) 2024/2847, they should be deemed to comply with the cybersecurity requirements set out in Article 15 of Regulation (EU) 2024/1689 in so far as those requirements are covered by the EU declaration of conformity or parts thereof issued pursuant to Regulation (EU) 2024/2847. In order to improve the visibility of the interplay of Regulation (EU) 2024/1689 and Regulation (EU) 2024/2847, the rule of Article 12 of Regulation (EU) 2024/2847 should also be reflected in Regulation (EU) 2024/1689. The interplay between the two instruments should not be affected.

(20)In accordance with Article 6(1) of Regulation (EU) 2024/1689, AI systems are classified as high-risk where an AI system that is a component of a product covered by Union harmonisation legislation listed in Section A of Annex I to that Regulation is a safety component and that product requires a third-party conformity assessment. The requirement that such product must require a third-party conformity assessment, however, does not affect the choice of the manufacturer regarding the conformity assessment procedure for such product. Where Union harmonisation legislation listed in Section A of Annex I allows to choose a conformity assessment procedure based on harmonised standards amongst conformity assessment procedures, this possibility remains applicable also to products in which a high-risk AI system is embedded. Article 6(1) of Regulation (EU) 2024/1689 should not be understood to require products in which a high-risk AI system is embedded to automatically undergo a third-party conformity assessment involving a notified body. Where this possibility is provided for in Union harmonisation legislation, the provider of the product in which the high-risk AI system is embedded could continue to rely on harmonised standards to comply with the requirements of the Union harmonisation legislation and Regulation (EU) 2024/1689 as a conformity assessment procedure.

(21)In order to enhance competitiveness and innovation, it is essential to support economic operators that are required to comply simultaneously with the requirements or obligations set out in Chapter III, Sections 2 and 3 of Regulation (EU) 2024/1689, and with the relevant requirements and obligations laid down in the Union harmonisation legislation listed in Annex I to that Regulation. To support and simplify the regulatory compliance pathways of such economic operators, the Commission should request that, without undue delay, the European standardisation organisations develop standardisation deliverables, including, where appropriate, harmonised standards. Those standardisation deliverables should be based on the harmonised standards published in the Official Journal of the European Union that give the presumption of conformity with the requirements or obligations of Regulation (EU) 2024/1689, as well as any relevant harmonised standards published in the Official Journal of the European Union that give the presumption of conformity with the relevant requirements or obligations set out in the Union harmonisation legislation listed in Annex I to that Regulation. Such standardisation deliverables should help reduce legal uncertainty, avoid unnecessary duplication of conformity assessment activities, testing, documentation and reporting obligations, and lower compliance costs, in particular for small and medium-sized enterprises and start-ups. Timely development of such deliverables is essential in order to provide economic operators with practical and reliable technical solutions, strengthen legal certainty and facilitate the placing on the market, putting into service and use of AI systems in accordance with this Regulation and the Union harmonisation legislation listed in Annex I to that Regulation.

(22)To streamline compliance and reduce associated costs, the registration of AI systems referred to in Article 6(3) of Regulation (EU) 2024/1689 in the EU database pursuant to Article 49(2) of that Regulation should be simplified by streamlining the content required under Annex VIII to that Regulation. While it remains crucial for effective market surveillance and public accountability that such AI systems are registered in the EU database, the registration requirements should be simplified and made more proportionate. This simplification would strike a better balance without undermining the protection laid down by Regulation (EU) 2024/1689. Such AI systems are not considered to be high risk under certain conditions where they do not pose a significant risk of harm to the health, safety or fundamental rights of persons. Furthermore, a provider applying Article 6(3) of Regulation (EU) 2024/1689 remains obligated to document its assessment before that AI system is placed on the market or put into service. National competent authorities should be able to request that assessment.

(23)Articles 57, 58 and 60 of Regulation (EU) 2024/1689 should be amended to strengthen further cooperation at Union level of AI regulatory sandboxes, foster clarity and consistency in the governance of AI regulatory sandboxes, and to extend the scope of real-world testing outside AI regulatory sandboxes to high-risk AI systems covered by the Union harmonisation legislation listed in Annex I to that Regulation. In particular, to allow procedural simplification, where applicable, in the projects supervised in the AI regulatory sandboxes that also include real-world testing, the real-world testing plan should be integrated into the sandbox plan agreed by the providers or prospective providers and the competent authority.

(24)In addition, it is appropriate to provide for the possibility of the European Artificial Intelligence Office (AI Office) to establish an AI regulatory sandbox at Union level for AI systems that are covered by Article 75(1) of Regulation (EU) 2024/1689. To ensure coherence, legal certainty and an efficient allocation of supervisory responsibilities between Union and national levels, the scope of the Union-level AI regulatory sandbox should be clearly defined in order to avoid any overlap with national AI regulatory sandboxes established pursuant to that Regulation. In order to foster innovation and facilitate the uptake of AI, SMEs, including startups, and SMCs should be provided with priority access to the AI regulatory sandboxes established by the AI Office.

(25)Moreover, the provisions on the cooperation between relevant competent authorities for the operation of an AI regulatory sandbox should be clarified in order to ensure their effective functioning. For that reason, the empowerment of the Commission to adopt implementing acts specifying the detailed arrangements for the establishment, development, implementation, operation and supervision of the AI regulatory sandboxes should be extended to also cover governance aspects of such sandboxes. In addition, where AI regulatory sandboxes involve innovative AI systems that process personal data or otherwise fall under the supervisory remit of other national authorities or competent authorities providing or supporting access to data, the relevant competent supervisory authorities should be associated with the operation of the AI regulatory sandbox and involved in the supervision of those aspects to the extent of their respective tasks and powers.

(26)To foster innovation, it is also appropriate to extend the scope of real-world testing outside AI regulatory sandboxes in Article 60 of Regulation (EU) 2024/1689, currently applicable to high-risk AI systems listed in Annex III to that Regulation, and allow providers and prospective providers of high-risk AI systems covered by the Union harmonisation legislation listed in Annex I to that Regulation to also test such systems in real-world conditions, subject to sufficient safeguards. This is without prejudice to other Union or national law on the testing in real-world conditions of high-risk AI systems related to products covered by that Union harmonisation legislation.

(27)It is also appropriate to ensure that real-world testing of high-risk AI systems covered by the Union harmonisation legislation listed in Section B of Annex I to Regulation (EU) 2024/1689 is possible. Those systems are subject to the requirements and procedures of the relevant sectoral legislation and are, for most purposes, not directly subject to that Regulation. Those sectoral acts will, in due course, incorporate requirements corresponding to the requirements set out in Articles 8 to 15 of that Regulation. Therefore, it is appropriate to ensure that Member States can allow real-world testing of these AI systems with a view to assessing and verifying the conformity of those systems with the requirements set out in Articles 8 to 15 of that Regulation. If Member States decide to allow such testing, that Regulation should require them to adopt frameworks setting out the detailed requirements for such testing. That Regulation should provide for essential elements to be contained in such frameworks. When designing such frameworks, Member States should ensure a high level of protection of health safety and fundamental rights of natural persons. Before implementing the framework, Member States should notify it to the Commission. The real-world testing should comply with the relevant Union harmonisation legislation listed in Section B of Annex I to Regulation (EU) 2024/1689, including any applicable provisions regarding testing. However, this should not affect the application of the new article regarding real-world testing.

(28)Article 63 of Regulation (EU) 2024/1689 offers microenterprises who are providers of high-risk AI systems the possibility to benefit from a simplified way to comply with the obligation to establish a quality management system. With a view to facilitating compliance for more innovators, that possibility should be extended to all SMEs, including start-ups.

(29)In light of the important role of the AI Office for the effective and coordinated governance of Regulation (EU) 2024/1689, as further reinforced by this Regulation, and without prejudice to the next Multiannual Financial Framework and to the budgetary procedure, the Commission should allocate adequate human, financial and technical resources to the AI Office to ensure that it can effectively, and within reasonable timeframes, perform its tasks in relation to the enforcement of Regulation (EU) 2024/1689, including the allocation of a sufficient number of permanent personnel with in-depth competences and technical expertise.

(30)Article 69 of Regulation (EU) 2024/1689 should be amended to simplify the fee structure of the scientific panel. If Member States call upon the panel’s expertise, the fees they may be required to pay to the experts should be equivalent to the remuneration the Commission is obliged to pay in similar circumstances.

(31)In order to strengthen the governance system for AI systems, it is necessary to clarify the role of the AI Office in monitoring and supervising the compliance of such AI systems with Regulation (EU) 2024/1689. The Commission has exclusive competence as regards general-purpose AI models under Article 88 of that Regulation. To increase coherence, clarity and effectiveness, and in light of the reach and impacts of AI systems linked to those competences, the scope of the AI Office’s exclusive competence to supervise systems should be refined. In particular, the AI Office should have exclusive competence over AI systems built on general-purpose AI models, not only where both the system and the model are developed by the same provider, but also where they are developed by providers that form part of the same undertaking. However, in certain cases, in particular where there is specific sectoral supervision, responsibility should remain with the relevant national competent authority. Accordingly, certain exceptions should be laid down. The personal scope of this exclusive competence should extend to the providers of those AI systems and to their deployers within the same undertaking. Other deployers should remain subject to national supervision and enforcement. Moreover, this does not include AI systems placed on the market, put into service or used by Union institutions, bodies, offices or agencies, which are under the supervision of the European Data Protection Supervisor pursuant to Article 74(9) of Regulation (EU) 2024/1689.

(32)Additionally, considering the existing supervisory and enforcement system under Regulation (EU) 2022/2065 of the European Parliament and of the Council, it is appropriate to grant the Commission the powers of a competent market surveillance authority pursuant to Regulation (EU) 2024/1689 where an AI system qualifies as a very large online platform or a very large online search engine in accordance with Regulation (EU) 2022/2065, or where it is embedded in such a platform or search engine. This should contribute to ensuring that the exercise of the Commission’s supervision and enforcement powers pursuant to Regulation (EU) 2024/1689 and Regulation (EU) 2022/2065, as well as those applicable to general-purpose AI models integrated into such platforms or search engines, is carried out in a coherent and effective manner. This is also appropriate in light of the importance of such platforms and search engines, in view of their reach, impact and potential to cause complex and large societal harms. The personal scope of this exclusive competence should extend to the providers of those AI systems and to their deployers within the same undertaking. In the case of AI systems embedded in or qualifying as a very large online platform or search engine, the first point of entry for the assessment of the AI systems are the risk assessment, mitigating measures and audit obligations prescribed by Articles 34, 35 and 37 of Regulation (EU) 2022/2065, without prejudice to the AI Office’s powers to investigate and enforce ex post non-compliance with the rules of Regulation (EU) 2024/1689. In the context of the analysis of this risk assessment, mitigating measures and audits, the Commission services responsible for the enforcement of Regulation (EU) 2022/2065 may seek the opinion of the AI Office on the outcome of a potential earlier or parallel risk assessment carried out in accordance with Regulation (EU) 2024/1689 and the applicability of prohibitions pursuant to Regulation (EU) 2024/1689. In addition, the AI Office and the competent national authorities should, in accordance with Regulation (EU) 2024/1689, coordinate their enforcement efforts with the authorities competent for the supervision and enforcement of Regulation (EU) 2022/2065, including the Commission, in order to ensure that the principles of loyal cooperation, proportionality and ne bis in idem are respected, while information obtained pursuant to one Regulation is to be used for the purposes of supervision and enforcement of the other only provided the undertaking agrees. In particular, those authorities should exchange views regularly and take into account, in their respective areas of competence, any fines and penalties imposed on the same provider for the same conduct through a final decision in proceedings relating to an infringement of other Union or national rules, so as to ensure that the overall fines and penalties imposed are proportionate and correspond to the seriousness of the infringements committed.

(33)When supervising and enforcing obligations in relation to AI systems under its competence, the AI Office has the same role and responsibility as a market surveillance authority pursuant to Regulation (EU) 2024/1689. Consequently, it is necessary for the AI Office to have all of the powers and responsibilities that market surveillance authorities have pursuant to that Regulation and Regulation (EU) 2019/1020 of the European Parliament and of the Council (‘the general powers’). These should ensure the appropriate and effective enforcement of the requirements and obligations set out in Regulation (EU) 2024/1689. However, it is necessary to specify, complement, and frame certain essential elements and other aspects of the general powers, as well as their safeguards (‘the specifying provisions’). In particular, it is necessary to lay down provisions governing the relationship between the AI Office and national authorities; provisions specifying, complementing, and constraining the powers to request information and conduct on-site inspections; provisions governing investigations, including the possibility to make commitments binding; and provisions specifying and constraining power to find non-compliance, impose fines and impose periodic penalties. Where a type of general power has been so specified, the AI Office may not circumvent the conditions and limits of those powers by relying on a related general power. Conversely, types of general power that are not specified and framed in respect of the AI Office, such as the power to adopt measures referred to in Article 16(3) of Regulation (EU) 2019/1020, may be relied on by the AI Office. As necessary for the proper implementation of Regulation (EU) 2024/1689, the Commission should be able to adopt implementing acts further defining the rules and the procedures concerning the application of limitation periods and, the access to the file and the negotiated disclosure of information. In exercising all of these powers, the AI Office must comply with the Charter. Additionally, the AI Office is subject to the safeguards and protection for fundamental rights laid down in the specifying provisions.

(34)In addition to those procedural and fundamental rights safeguards the procedural rights provided for in Article 18 of Regulation (EU) 2019/1020 should apply mutatis mutandis to providers of AI systems, without prejudice to more specific procedural rights provided for in Regulation (EU) 2024/1689. When national market surveillance authorities, through the single point of contact, request that the AI Office takes supervisory and enforcement measures with regards to AI systems under its exclusive supervision, the AI Office should no later than four months following the receipt of that request, inform the single point of contact of its intention to exercise its supervisory and enforcement powers or of its reasons for not exercising its powers. If the AI Office decides to exercise its supervisory and enforcement powers, it should also inform the single point of contact about the final outcome of such proceedings and of intermediate developments that the AI Office considers as having a major impact in the investigation, including the decision to open proceedings, to impose a fine, and to withdraw or recall the AI system from the market.

(35)To enable access to the Union market for AI systems which are under the supervision of the AI Office pursuant to Article 75 of Regulation (EU) 2024/1689 and subject to a third party conformity assessment, the Commission should be responsible for the pre-market conformity assessments of those systems.

(36)Article 77 and related provisions of Regulation (EU) 2024/1689 constitute an important governance mechanism, as they aim to enable authorities or bodies responsible for enforcing or supervising Union law intended to protect fundamental rights to fulfil their mandate under specific conditions and to foster cooperation with market surveillance authorities responsible for the supervision and enforcement of that Regulation. It is necessary to clarify the scope of such cooperation, as well as to clarify which public authorities or bodies benefit from it. With a view to reinforcing cooperation, it should be clarified that requests to access information and documentation should be made to the competent market surveillance authority, which should respond to such requests without undue delay, and that the authorities or bodies involved should have a mutual obligation to cooperate. It should be clarified that these provisions are without prejudice to the competences, tasks, powers and independence of the relevant national public authorities or bodies under their mandates. In particular, those provisions do not limit any powers that those authorities and bodies have to request information pursuant to other Union or national law. Accordingly, those authorities and bodies retain any power they have to directly request information from operators pursuant to their mandate or Union or national law.

(37)The requirements for high-risk AI systems laid down in Regulation (EU) 2024/1689 address specific risks inherent to AI systems, including bias, unpredictable model behaviour, poor robustness or accuracy, vulnerabilities to attacks by third parties, and a lack of transparency of the AI system. By addressing AI-specific risks, that Regulation complements the requirements laid down in Union harmonisation legislation listed in Annex I to that Regulation, without duplicating them. Regulation (EU) 2024/1689 provides mechanisms for economic operators to minimise the compliance burden. In particular, Article 8(2) on the interplay with the sectoral legislation, Article 9(10) on risk management and Article 17(3) on quality management, allow economic operators to integrate, when necessary and appropriate, an assessment of AI-specific risks into existing risk and quality management systems. Article 40 of Regulation (EU) 2024/1689 further requires the Commission to specify that harmonised standards developed pursuant to that Regulation are to be consistent with standards developed pursuant to the Union harmonisation legislation listed in Annex I to that Regulation. The Commission should provide guidelines to assist economic operators of high-risk AI systems covered in Annex I to Regulation (EU) 2024/1689 in complying with that Regulation, including by providing guidance on the application of Article 8(2), Article 9(10) and Article 17(3) of that Regulation as mechanisms to minimise the compliance burden, in line with principles of complementarity and proportionality. Those guidelines should be published, at the latest, by 1 August 2027.

(38)To allow sufficient time for providers of generative AI systems subject to the marking obligations laid down in Article 50(2) of Regulation (EU) 2024/1689 to adapt their practices within a reasonable time without disrupting the market, it is appropriate to introduce a transitional period of four months for providers who have already placed their systems on the market before the 2 August 2026.

(39)To provide sufficient time for providers of high-risk AI systems and to clarify rules applicable to the AI systems already placed on the market or put into service before the relevant provisions of Regulation (EU) 2024/1689 apply, it is appropriate to clarify the scope of the grace period provided in Article 111(2) of that Regulation. For the purpose of that Article 111(2), the grace period should apply where the type and model of AI system has already been placed on the market. This means that if at least one individual unit of the high-risk AI system has been lawfully placed on the market or put into service before the date specified in Article 111(2), other individual units of the same type and model of high-risk AI system are subject to the grace period provided in that Article 111(2) and thus may continue to be placed on the market, made available or put into service on the Union market without any additional obligations, requirements or the need for additional certification, as long as the design of that high-risk AI system remains unchanged. For the purposes of the application of the grace period provided in Article 111(2), the decisive factor is the date on which the first unit of that type and model of high-risk AI system was placed on the market or put into service on the Union market for the first time. Any significant change to the design of that AI system after the date specified in Article 111(2) should trigger the obligation of the provider to fully comply with all relevant provisions of that Regulation applicable to high-risk AI systems, including the conformity assessment requirements.

(40)Article 113 of Regulation (EU) 2024/1689 establishes the dates of entry into force and application of that Regulation, in particular that the general date of application is 2 August 2026. For the obligations related to high-risk AI systems laid down in Sections 1, 2 and 3 of Chapter III of Regulation (EU) 2024/1689, the delayed availability of standards, common specifications, and alternative guidance and the delayed establishment of national competent authorities lead to challenges that jeopardise the effective entry into application of those obligations and that risk a significant increase in implementation costs in a way that does not justify maintaining their initial date of application, namely 2 August 2026. Therefore, it is appropriate that the date of application of Sections 1, 2 and 3 of Chapter III is set to 2 December 2027 for AI systems classified as high-risk pursuant to Article 6(2) and Annex III, and to 2 August 2028 for AI systems classified as high-risk pursuant to Article 6(1) and Annex I. The distinction between the entry into application of the rules as regards AI systems classified as high-risk pursuant to Article 6(2) and Annex III and Article 6(1) and Annex I to that Regulation is consistent with the difference between the initial dates of application envisaged in Regulation (EU) 2024/1689 and aims to provide the necessary time for adaptation and implementation of the corresponding obligations. The timely availability of support instruments, including guidance, relevant standards, common specifications and codes of practice is important in order to facilitate compliance and reduce the risk of divergent interpretation and uneven application of the rules across Member States. In order to ensure legal certainty and to avoid further delays in the application of Regulation (EU) 2024/1689, the Commission should ensure that measures in support of compliance with regard to Chapter III, Sections 1, 2, and 3 of that Regulation are in place in due time to ensure timely and effective implementation of the necessary provisions.

(41)In light of the objective to reduce implementation challenges for citizens, businesses and public administrations, it is essential that harmonised conditions for the implementation of certain rules are adopted only where strictly necessary. For that purpose, it is appropriate to remove certain empowerments bestowed on the Commission to adopt such harmonised conditions by means of implementing acts in cases where that is not the case. Article 50(7), Article 56(6), and Article 72(3) of Regulation (EU) 2024/1689 should therefore be amended to remove the empowerments conferred on the Commission to adopt implementing acts. Given that the codes of practice referred to in Article 50(7) and Article 56(6) have limited legal effect, and in particular do not grant a presumption of conformity, it is not strictly necessary for these codes to be approved by an implementing act. Providers should be able to rely, pursuant to Article 53(4) and Article 54(2) of Regulation (EU) 2024/1689, on codes of practice assessed as adequate pursuant to Article 56(6) thereof. The removal of the empowerment to adopt a harmonised template for a post-market monitoring plan in Article 72(3) of Regulation (EU) 2024/1689 has the additional benefit of offering more flexibility for providers of high-risk AI systems to put in place a system for post-market monitoring that is tailored to their organisation. At the same time, recognising the need to offer clarity regarding how providers of high-risk AI systems are required to comply with their obligation set out in Article 72(1) of Regulation (EU) 2024/1689, the Commission should be required to publish guidance, including a voluntary template, on the post-market monitoring plan by 2 September 2027.

(42)The use of artificial intelligence in machinery can help foster innovation and improve the efficiency of those machines. The application of Regulation (EU) 2023/1230 of the European Parliament and of the Council and Regulation (EU) 2024/1689 might lead to overlaps. At the same time, it is important to ensure a level of protection with regard to risks related to the use of artificial intelligence in machinery that is consistent with the level of protection provided in Regulation (EU) 2024/1689 with regard to high-risk AI systems. Given the specific nature of machinery and the machinery sector, and in order to address the need to simplify the regulatory framework for AI-enabled machinery, it is appropriate to move to a sectoral approach by moving Regulation (EU) 2023/1230 from Section A to Section B of Annex I to Regulation (EU) 2024/1689. Accordingly, first, the application of Regulation (EU) 2024/1689 to those machines should be limited to the provisions referred to in Article 2(2) of that Regulation. The reference to Directive 2006/42/EC should be moved from Section A to Section B of Annex I to Regulation (EU) 2024/1689 and updated so as to refer to Regulation (EU) 2023/1230. Second, it is crucial to ensure that Regulation (EU) 2023/1230 incorporates essential health and safety requirements for high-risk AI systems classified pursuant to Article 6(1) of Regulation (EU) 2024/1689 and used as a safety component in machinery or high-risk AI systems constituting machinery that ensure a level of protection consistent with Regulation (EU) 2024/1689. To that end, the Commission should be required to adopt delegated acts amending Annex III to Regulation (EU) 2023/1230 in order to reflect the relevant requirements set out in Chapter III, Section 2, and Articles 17, 19, 72 and 73 of Regulation (EU) 2024/1689. To avoid a legal gap and ensure alignment with the entry into application of the relevant high-risk AI system rules set out in Regulation (EU) 2024/1689, those delegated acts should apply by 2 August 2028. For the same reasons, manufacturers should be free to rely on harmonised standards or common specifications referenced or adopted pursuant to Regulation (EU) 2024/1689 that cover the relevant essential requirements for the presumption of conformity within the meaning of Article 20 of Regulation (EU) 2023/1230 until harmonised standards or common specifications regarding AI are referenced or adopted pursuant to Regulation (EU) 2023/1230.

(43)Conformity assessments of high-risk AI systems pursuant to Regulation (EU) 2024/1689 can require the involvement of conformity assessment bodies. Only conformity assessment bodies that have been designated pursuant to that Regulation are able to carry out conformity assessments, and only for the activities related to the categories and types of AI systems concerned. To enable the specification of the scope of the designation of conformity assessment bodies notified pursuant to Article 30 of Regulation (EU) 2024/1689, it is necessary to draw up a list of codes, categories, and corresponding types of AI systems. The list of codes should take into account whether the AI system is a component of a product or itself a product covered by the Union harmonisation legislation listed in Annex I to Regulation (EU) 2024/1689 (referred to as ‘AIP codes’, for AI systems covered by product legislation), or a system listed in Annex III to that Regulation, which currently concerns only biometric AI systems referred to in point (1) of Annex III (referred to as ‘AIB codes’, for biometric AI systems). Both AIP codes and AIB codes are vertical codes. The AIP codes are reference codes that provide a link to the Union harmonisation legislation listed in Section A of Annex I to Regulation (EU) 2024/1689. The AIB codes are new codes specific to Regulation (EU) 2024/1689 that identify biometric AI systems referred to in point (1) of Annex III to that Regulation. The list of codes should also take into account specific types and underlying technologies of AI systems (referred to as ‘AIH codes’, for horizontal AI system codes). The AIH codes are new AI technology-specific codes and can be applied in conjunction with AIP or AIB vertical codes. The AIH codes cover the underlying types and technologies of AI systems. The list of codes, comprising three categories, should provide for a multi-dimensional typology of AI systems which ensures that conformity assessment bodies designated as notified bodies are fully competent in regard to the AI systems they are required to assess.

(44)Regulation (EU) 2018/1139 of the European Parliament and the Council lays down common rules in the field of civil aviation. Article 108 of Regulation (EU) 2024/1689 amends Regulation (EU) 2018/1139 to ensure that the Commission, when adopting any relevant delegated or implementing acts pursuant to Regulation (EU) 2018/1139, takes into account, on the basis of the technical and regulatory specificities of the civil aviation sector, and without interfering with existing governance, conformity assessment and enforcement mechanisms and authorities established therein, the mandatory requirements for high-risk AI systems laid down in Regulation (EU) 2024/1689. A technical correction to amend additional Articles of Regulation (EU) 2018/1139 is necessary to ensure that the mandatory requirements for high-risk AI systems laid down in Regulation (EU) 2024/1689 are fully covered when adopting relevant delegated or implementing acts pursuant to Regulation (EU) 2018/1139.

(45)In order to amend certain non-essential elements of Regulations (EU) 2024/1689 and (EU) 2023/1230, the power to adopt acts in accordance with Article 290 of the Treaty on the Functioning of the European Union should be delegated to the Commission in respect of:

—limiting the application of specific requirements or obligations laid down in Regulation (EU) 2024/1689 where Union harmonisation legislation listed in Section A of Annex I of that Regulation lays down requirements providing for an equivalent level of protection,

—amending the list of codes, categories, and corresponding types of AI systems in Annex XIV of Regulation (EU) 2024/1689, and

—amending Annex III of Regulation (EU) 2023/1230 to reflect the relevant requirements of Regulation (EU) 2024/1689.

It is of particular importance that the Commission carry out appropriate consultations during its preparatory work, including at expert level, and that those consultations be conducted in accordance with the principles laid down in the Interinstitutional Agreement of 13 April 2016 on Better Law-Making. In particular, to ensure equal participation in the preparation of delegated acts, the European Parliament and the Council receive all documents at the same time as Member States’ experts, and their experts systematically have access to meetings of Commission expert groups dealing with the preparation of delegated acts.

(46)In order to ensure legal certainty without delay, with a view to the imminent general application of Regulation (EU) 2024/1689, this Regulation should enter into force as a matter of urgency on the third day following that of its publication in the Official Journal of the European Union.

(47)The European Data Protection Supervisor and the European Data Protection Board were consulted in accordance with Article 42(1) and (2) of Regulation (EU) 2018/1725 and delivered their joint opinion on 20 January 2026,

HAVE ADOPTED THIS REGULATION: