Preamble

Recitals

CHAPTER I GENERAL PROVISIONS

CHAPTER II PROHIBITED AI PRACTICES

CHAPTER III HIGH-RISK AI SYSTEMS

SECTION 1 Classification of AI systems as high-risk

SECTION 2 Requirements for high-risk AI systems

SECTION 3 Obligations of providers and deployers of high-risk AI systems and other parties

SECTION 4 Notifying authorities and notified bodies

SECTION 5 Standards, conformity assessment, certificates, registration

CHAPTER IV TRANSPARENCY OBLIGATIONS FOR PROVIDERS AND DEPLOYERS OF CERTAIN AI SYSTEMS

CHAPTER V GENERAL-PURPOSE AI MODELS

CHAPTER VI MEASURES IN SUPPORT OF INNOVATION

CHAPTER VII GOVERNANCE

CHAPTER VIII EU DATABASE FOR HIGH-RISK AI SYSTEMS

CHAPTER IX POST-MARKET MONITORING, INFORMATION SHARING AND MARKET SURVEILLANCE

SECTION 1 Post-market monitoring

SECTION 2 Sharing of information on serious incidents

SECTION 3 Enforcement

SECTION 4 Remedies

SECTION 5 Supervision, investigation, enforcement and monitoring in respect of providers of general-purpose AI models

CHAPTER X CODES OF CONDUCT AND GUIDELINES

CHAPTER XI DELEGATION OF POWER AND COMMITTEE PROCEDURE

CHAPTER XII PENALTIES

CHAPTER XIII FINAL PROVISIONS

ANNEX I List of Union harmonisation legislation

ANNEX II List of criminal offences referred to in Article 5(1), first subparagraph, point (h)(iii)

ANNEX III High-risk AI systems referred to in Article 6(2)

ANNEX IV Technical documentation referred to in Article 11(1)

ANNEX V EU declaration of conformity

ANNEX VI Conformity assessment procedure based on internal control

ANNEX VII Conformity based on an assessment of the quality management system and an assessment of the technical documentation

ANNEX VIII Information to be submitted upon the registration of high-risk AI systems in accordance with Article 49

ANNEX IX Information to be submitted upon the registration of high-risk AI systems listed in Annex III in relation to testing in real world conditions in accordance with Article 60

ANNEX X Union legislative acts on large-scale IT systems in the area of Freedom, Security and Justice

ANNEX XI Technical documentation referred to in Article 53(1), point (a) — technical documentation for providers of general-purpose AI models

ANNEX XII Transparency information referred to in Article 53(1), point (b)

ANNEX XIII Criteria for the designation of general-purpose AI models with systemic risk referred to in Article 51

Close

EU AI Act OJ version (DE)

As effective on 1 August 2024

Explore Plain Text

Exact match

Preamble

195 words

Recitals

35839 words

Recital 1

Recital 2

Recital 3

Recital 4

Recital 5

Recital 6

Recital 7

Recital 8

Recital 9

Recital 10

Recital 11

Recital 12

The notion of ‘AI system’ in this Regulation should be clearly defined and should be closely aligned with the work of international organisations working on AI to ensure legal certainty, facilitate international convergence and wide acceptance, while providing the flexibility to accommodate the rapid technological developments in this field. Moreover, the definition should be based on key characteristics of AI systems that distinguish it from simpler traditional software systems or programming approaches and should not cover systems that are based on the rules defined solely by natural persons to automatically execute operations. A key characteristic of AI systems is their capability to infer. This capability to infer refers to the process of obtaining the outputs, such as predictions, content, recommendations, or decisions, which can influence physical and virtual environments, and to a capability of AI systems to derive models or algorithms, or both, from inputs or data. The techniques that enable inference while building an AI system include machine learning approaches that learn from data how to achieve certain objectives, and logic- and knowledge-based approaches that infer from encoded knowledge or symbolic representation of the task to be solved. The capacity of an AI system to infer transcends basic data processing by enabling learning, reasoning or modelling. The term ‘machine-based’ refers to the fact that AI systems run on machines. The reference to explicit or implicit objectives underscores that AI systems can operate according to explicit defined objectives or to implicit objectives. The objectives of the AI system may be different from the intended purpose of the AI system in a specific context. For the purposes of this Regulation, environments should be understood to be the contexts in which the AI systems operate, whereas outputs generated by the AI system reflect different functions performed by AI systems and include predictions, content, recommendations or decisions. AI systems are designed to operate with varying levels of autonomy, meaning that they have some degree of independence of actions from human involvement and of capabilities to operate without human intervention. The adaptiveness that an AI system could exhibit after deployment, refers to self-learning capabilities, allowing the system to change while in use. AI systems can be used on a stand-alone basis or as a component of a product, irrespective of whether the system is physically integrated into the product (embedded) or serves the functionality of the product without being integrated therein (non-embedded).

Recital 13

Recital 14

Recital 15

Recital 16

Recital 17

The notion of ‘remote biometric identification system’ referred to in this Regulation should be defined functionally, as an AI system intended for the identification of natural persons without their active involvement, typically at a distance, through the comparison of a person’s biometric data with the biometric data contained in a reference database, irrespectively of the particular technology, processes or types of biometric data used. Such remote biometric identification systems are typically used to perceive multiple persons or their behaviour simultaneously in order to facilitate significantly the identification of natural persons without their active involvement. This excludes AI systems intended to be used for biometric verification, which includes authentication, the sole purpose of which is to confirm that a specific natural person is the person he or she claims to be and to confirm the identity of a natural person for the sole purpose of having access to a service, unlocking a device or having security access to premises. That exclusion is justified by the fact that such systems are likely to have a minor impact on fundamental rights of natural persons compared to the remote biometric identification systems which may be used for the processing of the biometric data of a large number of persons without their active involvement. In the case of ‘real-time’ systems, the capturing of the biometric data, the comparison and the identification occur all instantaneously, near-instantaneously or in any event without a significant delay. In this regard, there should be no scope for circumventing the rules of this Regulation on the ‘real-time’ use of the AI systems concerned by providing for minor delays. ‘Real-time’ systems involve the use of ‘live’ or ‘near-live’ material, such as video footage, generated by a camera or other device with similar functionality. In the case of ‘post’ systems, in contrast, the biometric data has already been captured and the comparison and identification occur only after a significant delay. This involves material, such as pictures or video footage generated by closed circuit television cameras or private devices, which has been generated before the use of the system in respect of the natural persons concerned.

Recital 18

Recital 19

For the purposes of this Regulation the notion of ‘publicly accessible space’ should be understood as referring to any physical space that is accessible to an undetermined number of natural persons, and irrespective of whether the space in question is privately or publicly owned, irrespective of the activity for which the space may be used, such as for commerce, for example, shops, restaurants, cafés; for services, for example, banks, professional activities, hospitality; for sport, for example, swimming pools, gyms, stadiums; for transport, for example, bus, metro and railway stations, airports, means of transport; for entertainment, for example, cinemas, theatres, museums, concert and conference halls; or for leisure or otherwise, for example, public roads and squares, parks, forests, playgrounds. A space should also be classified as being publicly accessible if, regardless of potential capacity or security restrictions, access is subject to certain predetermined conditions which can be fulfilled by an undetermined number of persons, such as the purchase of a ticket or title of transport, prior registration or having a certain age. In contrast, a space should not be considered to be publicly accessible if access is limited to specific and defined natural persons through either Union or national law directly related to public safety or security or through the clear manifestation of will by the person having the relevant authority over the space. The factual possibility of access alone, such as an unlocked door or an open gate in a fence, does not imply that the space is publicly accessible in the presence of indications or circumstances suggesting the contrary, such as. signs prohibiting or restricting access. Company and factory premises, as well as offices and workplaces that are intended to be accessed only by relevant employees and service providers, are spaces that are not publicly accessible. Publicly accessible spaces should not include prisons or border control. Some other spaces may comprise both publicly accessible and non-publicly accessible spaces, such as the hallway of a private residential building necessary to access a doctor’s office or an airport. Online spaces are not covered, as they are not physical spaces. Whether a given space is accessible to the public should however be determined on a case-by-case basis, having regard to the specificities of the individual situation at hand.

Recital 20

Recital 21

Recital 22

In light of their digital nature, certain AI systems should fall within the scope of this Regulation even when they are not placed on the market, put into service, or used in the Union. This is the case, for example, where an operator established in the Union contracts certain services to an operator established in a third country in relation to an activity to be performed by an AI system that would qualify as high-risk. In those circumstances, the AI system used in a third country by the operator could process data lawfully collected in and transferred from the Union, and provide to the contracting operator in the Union the output of that AI system resulting from that processing, without that AI system being placed on the market, put into service or used in the Union. To prevent the circumvention of this Regulation and to ensure an effective protection of natural persons located in the Union, this Regulation should also apply to providers and deployers of AI systems that are established in a third country, to the extent the output produced by those systems is intended to be used in the Union. Nonetheless, to take into account existing arrangements and special needs for future cooperation with foreign partners with whom information and evidence is exchanged, this Regulation should not apply to public authorities of a third country and international organisations when acting in the framework of cooperation or international agreements concluded at Union or national level for law enforcement and judicial cooperation with the Union or the Member States, provided that the relevant third country or international organisation provides adequate safeguards with respect to the protection of fundamental rights and freedoms of individuals. Where relevant, this may cover activities of entities entrusted by the third countries to carry out specific tasks in support of such law enforcement and judicial cooperation. Such framework for cooperation or agreements have been established bilaterally between Member States and third countries or between the European Union, Europol and other Union agencies and third countries and international organisations. The authorities competent for supervision of the law enforcement and judicial authorities under this Regulation should assess whether those frameworks for cooperation or international agreements include adequate safeguards with respect to the protection of fundamental rights and freedoms of individuals. Recipient national authorities and Union institutions, bodies, offices and agencies making use of such outputs in the Union remain accountable to ensure their use complies with Union law. When those international agreements are revised or new ones are concluded in the future, the contracting parties should make utmost efforts to align those agreements with the requirements of this Regulation.

Recital 23

Recital 24

If, and insofar as, AI systems are placed on the market, put into service, or used with or without modification of such systems for military, defence or national security purposes, those should be excluded from the scope of this Regulation regardless of which type of entity is carrying out those activities, such as whether it is a public or private entity. As regards military and defence purposes, such exclusion is justified both by Article 4(2) TEU and by the specificities of the Member States’ and the common Union defence policy covered by Chapter 2 of Title V TEU that are subject to public international law, which is therefore the more appropriate legal framework for the regulation of AI systems in the context of the use of lethal force and other AI systems in the context of military and defence activities. As regards national security purposes, the exclusion is justified both by the fact that national security remains the sole responsibility of Member States in accordance with Article 4(2) TEU and by the specific nature and operational needs of national security activities and specific national rules applicable to those activities. Nonetheless, if an AI system developed, placed on the market, put into service or used for military, defence or national security purposes is used outside those temporarily or permanently for other purposes, for example, civilian or humanitarian purposes, law enforcement or public security purposes, such a system would fall within the scope of this Regulation. In that case, the entity using the AI system for other than military, defence or national security purposes should ensure the compliance of the AI system with this Regulation, unless the system is already compliant with this Regulation. AI systems placed on the market or put into service for an excluded purpose, namely military, defence or national security, and one or more non-excluded purposes, such as civilian purposes or law enforcement, fall within the scope of this Regulation and providers of those systems should ensure compliance with this Regulation. In those cases, the fact that an AI system may fall within the scope of this Regulation should not affect the possibility of entities carrying out national security, defence and military activities, regardless of the type of entity carrying out those activities, to use AI systems for national security, military and defence purposes, the use of which is excluded from the scope of this Regulation. An AI system placed on the market for civilian or law enforcement purposes which is used with or without modification for military, defence or national security purposes should not fall within the scope of this Regulation, regardless of the type of entity carrying out those activities.

Recital 25

Recital 26

Recital 27

While the risk-based approach is the basis for a proportionate and effective set of binding rules, it is important to recall the 2019 Ethics guidelines for trustworthy AI developed by the independent AI HLEG appointed by the Commission. In those guidelines, the AI HLEG developed seven non-binding ethical principles for AI which are intended to help ensure that AI is trustworthy and ethically sound. The seven principles include human agency and oversight; technical robustness and safety; privacy and data governance; transparency; diversity, non-discrimination and fairness; societal and environmental well-being and accountability. Without prejudice to the legally binding requirements of this Regulation and any other applicable Union law, those guidelines contribute to the design of coherent, trustworthy and human-centric AI, in line with the Charter and with the values on which the Union is founded. According to the guidelines of the AI HLEG, human agency and oversight means that AI systems are developed and used as a tool that serves people, respects human dignity and personal autonomy, and that is functioning in a way that can be appropriately controlled and overseen by humans. Technical robustness and safety means that AI systems are developed and used in a way that allows robustness in the case of problems and resilience against attempts to alter the use or performance of the AI system so as to allow unlawful use by third parties, and minimise unintended harm. Privacy and data governance means that AI systems are developed and used in accordance with privacy and data protection rules, while processing data that meets high standards in terms of quality and integrity. Transparency means that AI systems are developed and used in a way that allows appropriate traceability and explainability, while making humans aware that they communicate or interact with an AI system, as well as duly informing deployers of the capabilities and limitations of that AI system and affected persons about their rights. Diversity, non-discrimination and fairness means that AI systems are developed and used in a way that includes diverse actors and promotes equal access, gender equality and cultural diversity, while avoiding discriminatory impacts and unfair biases that are prohibited by Union or national law. Social and environmental well-being means that AI systems are developed and used in a sustainable and environmentally friendly manner as well as in a way to benefit all human beings, while monitoring and assessing the long-term impacts on the individual, society and democracy. The application of those principles should be translated, when possible, in the design and use of AI models. They should in any case serve as a basis for the drafting of codes of conduct under this Regulation. All stakeholders, including industry, academia, civil society and standardisation organisations, are encouraged to take into account, as appropriate, the ethical principles for the development of voluntary best practices and standards.

Recital 28

Recital 29

Recital 30

Recital 31

Recital 32

Recital 33

Recital 34

Recital 35

Recital 36

Recital 37

Recital 38

Recital 39

Recital 40

Recital 41

Recital 42

Recital 43

Recital 44

Recital 45

Recital 46

Recital 47

Recital 48

Recital 49

Recital 50

Recital 51

Recital 52

Recital 53

It is also important to clarify that there may be specific cases in which AI systems referred to in pre-defined areas specified in this Regulation do not lead to a significant risk of harm to the legal interests protected under those areas because they do not materially influence the decision-making or do not harm those interests substantially. For the purposes of this Regulation, an AI system that does not materially influence the outcome of decision-making should be understood to be an AI system that does not have an impact on the substance, and thereby the outcome, of decision-making, whether human or automated. An AI system that does not materially influence the outcome of decision-making could include situations in which one or more of the following conditions are fulfilled. The first such condition should be that the AI system is intended to perform a narrow procedural task, such as an AI system that transforms unstructured data into structured data, an AI system that classifies incoming documents into categories or an AI system that is used to detect duplicates among a large number of applications. Those tasks are of such narrow and limited nature that they pose only limited risks which are not increased through the use of an AI system in a context that is listed as a high-risk use in an annex to this Regulation. The second condition should be that the task performed by the AI system is intended to improve the result of a previously completed human activity that may be relevant for the purposes of the high-risk uses listed in an annex to this Regulation. Considering those characteristics, the AI system provides only an additional layer to a human activity with consequently lowered risk. That condition would, for example, apply to AI systems that are intended to improve the language used in previously drafted documents, for example in relation to professional tone, academic style of language or by aligning text to a certain brand messaging. The third condition should be that the AI system is intended to detect decision-making patterns or deviations from prior decision-making patterns. The risk would be lowered because the use of the AI system follows a previously completed human assessment which it is not meant to replace or influence, without proper human review. Such AI systems include for instance those that, given a certain grading pattern of a teacher, can be used to check ex post whether the teacher may have deviated from the grading pattern so as to flag potential inconsistencies or anomalies. The fourth condition should be that the AI system is intended to perform a task that is only preparatory to an assessment relevant for the purposes of the AI systems listed in an annex to this Regulation, thus making the possible impact of the output of the system very low in terms of representing a risk for the assessment to follow. That condition covers, inter alia, smart solutions for file handling, which include various functions from indexing, searching, text and speech processing or linking data to other data sources, or AI systems used for translation of initial documents. In any case, AI systems used in high-risk use-cases listed in an annex to this Regulation should be considered to pose significant risks of harm to the health, safety or fundamental rights if the AI system implies profiling within the meaning of Article 4, point (4) of Regulation (EU) 2016/679 or Article 3, point (4) of Directive (EU) 2016/680 or Article 3, point (5) of Regulation (EU) 2018/1725. To ensure traceability and transparency, a provider who considers that an AI system is not high-risk on the basis of the conditions referred to above should draw up documentation of the assessment before that system is placed on the market or put into service and should provide that documentation to national competent authorities upon request. Such a provider should be obliged to register the AI system in the EU database established under this Regulation. With a view to providing further guidance for the practical implementation of the conditions under which the AI systems listed in an annex to this Regulation are, on an exceptional basis, non-high-risk, the Commission should, after consulting the Board, provide guidelines specifying that practical implementation, completed by a comprehensive list of practical examples of use cases of AI systems that are high-risk and use cases that are not.

Recital 54

Recital 55

Recital 56

Recital 57

Recital 58

Another area in which the use of AI systems deserves special consideration is the access to and enjoyment of certain essential private and public services and benefits necessary for people to fully participate in society or to improve one’s standard of living. In particular, natural persons applying for or receiving essential public assistance benefits and services from public authorities namely healthcare services, social security benefits, social services providing protection in cases such as maternity, illness, industrial accidents, dependency or old age and loss of employment and social and housing assistance, are typically dependent on those benefits and services and in a vulnerable position in relation to the responsible authorities. If AI systems are used for determining whether such benefits and services should be granted, denied, reduced, revoked or reclaimed by authorities, including whether beneficiaries are legitimately entitled to such benefits or services, those systems may have a significant impact on persons’ livelihood and may infringe their fundamental rights, such as the right to social protection, non-discrimination, human dignity or an effective remedy and should therefore be classified as high-risk. Nonetheless, this Regulation should not hamper the development and use of innovative approaches in the public administration, which would stand to benefit from a wider use of compliant and safe AI systems, provided that those systems do not entail a high risk to legal and natural persons. In addition, AI systems used to evaluate the credit score or creditworthiness of natural persons should be classified as high-risk AI systems, since they determine those persons’ access to financial resources or essential services such as housing, electricity, and telecommunication services. AI systems used for those purposes may lead to discrimination between persons or groups and may perpetuate historical patterns of discrimination, such as that based on racial or ethnic origins, gender, disabilities, age or sexual orientation, or may create new forms of discriminatory impacts. However, AI systems provided for by Union law for the purpose of detecting fraud in the offering of financial services and for prudential purposes to calculate credit institutions’ and insurance undertakings’ capital requirements should not be considered to be high-risk under this Regulation. Moreover, AI systems intended to be used for risk assessment and pricing in relation to natural persons for health and life insurance can also have a significant impact on persons’ livelihood and if not duly designed, developed and used, can infringe their fundamental rights and can lead to serious consequences for people’s life and health, including financial exclusion and discrimination. Finally, AI systems used to evaluate and classify emergency calls by natural persons or to dispatch or establish priority in the dispatching of emergency first response services, including by police, firefighters and medical aid, as well as of emergency healthcare patient triage systems, should also be classified as high-risk since they make decisions in very critical situations for the life and health of persons and their property.

Recital 59

Given their role and responsibility, actions by law enforcement authorities involving certain uses of AI systems are characterised by a significant degree of power imbalance and may lead to surveillance, arrest or deprivation of a natural person’s liberty as well as other adverse impacts on fundamental rights guaranteed in the Charter. In particular, if the AI system is not trained with high-quality data, does not meet adequate requirements in terms of its performance, its accuracy or robustness, or is not properly designed and tested before being put on the market or otherwise put into service, it may single out people in a discriminatory or otherwise incorrect or unjust manner. Furthermore, the exercise of important procedural fundamental rights, such as the right to an effective remedy and to a fair trial as well as the right of defence and the presumption of innocence, could be hampered, in particular, where such AI systems are not sufficiently transparent, explainable and documented. It is therefore appropriate to classify as high-risk, insofar as their use is permitted under relevant Union and national law, a number of AI systems intended to be used in the law enforcement context where accuracy, reliability and transparency is particularly important to avoid adverse impacts, retain public trust and ensure accountability and effective redress. In view of the nature of the activities and the risks relating thereto, those high-risk AI systems should include in particular AI systems intended to be used by or on behalf of law enforcement authorities or by Union institutions, bodies, offices, or agencies in support of law enforcement authorities for assessing the risk of a natural person to become a victim of criminal offences, as polygraphs and similar tools, for the evaluation of the reliability of evidence in in the course of investigation or prosecution of criminal offences, and, insofar as not prohibited under this Regulation, for assessing the risk of a natural person offending or reoffending not solely on the basis of the profiling of natural persons or the assessment of personality traits and characteristics or the past criminal behaviour of natural persons or groups, for profiling in the course of detection, investigation or prosecution of criminal offences. AI systems specifically intended to be used for administrative proceedings by tax and customs authorities as well as by financial intelligence units carrying out administrative tasks analysing information pursuant to Union anti-money laundering law should not be classified as high-risk AI systems used by law enforcement authorities for the purpose of prevention, detection, investigation and prosecution of criminal offences. The use of AI tools by law enforcement and other relevant authorities should not become a factor of inequality, or exclusion. The impact of the use of AI tools on the defence rights of suspects should not be ignored, in particular the difficulty in obtaining meaningful information on the functioning of those systems and the resulting difficulty in challenging their results in court, in particular by natural persons under investigation.

Recital 60

Recital 61

Recital 62

Recital 63

Recital 64

To mitigate the risks from high-risk AI systems placed on the market or put into service and to ensure a high level of trustworthiness, certain mandatory requirements should apply to high-risk AI systems, taking into account the intended purpose and the context of use of the AI system and according to the risk-management system to be established by the provider. The measures adopted by the providers to comply with the mandatory requirements of this Regulation should take into account the generally acknowledged state of the art on AI, be proportionate and effective to meet the objectives of this Regulation. Based on the New Legislative Framework, as clarified in Commission notice ‘The “Blue Guide” on the implementation of EU product rules 2022’, the general rule is that more than one legal act of Union harmonisation legislation may be applicable to one product, since the making available or putting into service can take place only when the product complies with all applicable Union harmonisation legislation. The hazards of AI systems covered by the requirements of this Regulation concern different aspects than the existing Union harmonisation legislation and therefore the requirements of this Regulation would complement the existing body of the Union harmonisation legislation. For example, machinery or medical devices products incorporating an AI system might present risks not addressed by the essential health and safety requirements set out in the relevant Union harmonised legislation, as that sectoral law does not deal with risks specific to AI systems. This calls for a simultaneous and complementary application of the various legislative acts. To ensure consistency and to avoid an unnecessary administrative burden and unnecessary costs, providers of a product that contains one or more high-risk AI system, to which the requirements of this Regulation and of the Union harmonisation legislation based on the New Legislative Framework and listed in an annex to this Regulation apply, should have flexibility with regard to operational decisions on how to ensure compliance of a product that contains one or more AI systems with all the applicable requirements of that Union harmonised legislation in an optimal manner. That flexibility could mean, for example a decision by the provider to integrate a part of the necessary testing and reporting processes, information and documentation required under this Regulation into already existing documentation and procedures required under existing Union harmonisation legislation based on the New Legislative Framework and listed in an annex to this Regulation. This should not, in any way, undermine the obligation of the provider to comply with all the applicable requirements.

Recital 65

The risk-management system should consist of a continuous, iterative process that is planned and run throughout the entire lifecycle of a high-risk AI system. That process should be aimed at identifying and mitigating the relevant risks of AI systems on health, safety and fundamental rights. The risk-management system should be regularly reviewed and updated to ensure its continuing effectiveness, as well as justification and documentation of any significant decisions and actions taken subject to this Regulation. This process should ensure that the provider identifies risks or adverse impacts and implements mitigation measures for the known and reasonably foreseeable risks of AI systems to the health, safety and fundamental rights in light of their intended purpose and reasonably foreseeable misuse, including the possible risks arising from the interaction between the AI system and the environment within which it operates. The risk-management system should adopt the most appropriate risk-management measures in light of the state of the art in AI. When identifying the most appropriate risk-management measures, the provider should document and explain the choices made and, when relevant, involve experts and external stakeholders. In identifying the reasonably foreseeable misuse of high-risk AI systems, the provider should cover uses of AI systems which, while not directly covered by the intended purpose and provided for in the instruction for use may nevertheless be reasonably expected to result from readily predictable human behaviour in the context of the specific characteristics and use of a particular AI system. Any known or foreseeable circumstances related to the use of the high-risk AI system in accordance with its intended purpose or under conditions of reasonably foreseeable misuse, which may lead to risks to the health and safety or fundamental rights should be included in the instructions for use that are provided by the provider. This is to ensure that the deployer is aware and takes them into account when using the high-risk AI system. Identifying and implementing risk mitigation measures for foreseeable misuse under this Regulation should not require specific additional training for the high-risk AI system by the provider to address foreseeable misuse. The providers however are encouraged to consider such additional training measures to mitigate reasonable foreseeable misuses as necessary and appropriate.

Recital 66

Recital 67

High-quality data and access to high-quality data plays a vital role in providing structure and in ensuring the performance of many AI systems, especially when techniques involving the training of models are used, with a view to ensure that the high-risk AI system performs as intended and safely and it does not become a source of discrimination prohibited by Union law. High-quality data sets for training, validation and testing require the implementation of appropriate data governance and management practices. Data sets for training, validation and testing, including the labels, should be relevant, sufficiently representative, and to the best extent possible free of errors and complete in view of the intended purpose of the system. In order to facilitate compliance with Union data protection law, such as Regulation (EU) 2016/679, data governance and management practices should include, in the case of personal data, transparency about the original purpose of the data collection. The data sets should also have the appropriate statistical properties, including as regards the persons or groups of persons in relation to whom the high-risk AI system is intended to be used, with specific attention to the mitigation of possible biases in the data sets, that are likely to affect the health and safety of persons, have a negative impact on fundamental rights or lead to discrimination prohibited under Union law, especially where data outputs influence inputs for future operations (feedback loops). Biases can for example be inherent in underlying data sets, especially when historical data is being used, or generated when the systems are implemented in real world settings. Results provided by AI systems could be influenced by such inherent biases that are inclined to gradually increase and thereby perpetuate and amplify existing discrimination, in particular for persons belonging to certain vulnerable groups, including racial or ethnic groups. The requirement for the data sets to be to the best extent possible complete and free of errors should not affect the use of privacy-preserving techniques in the context of the development and testing of AI systems. In particular, data sets should take into account, to the extent required by their intended purpose, the features, characteristics or elements that are particular to the specific geographical, contextual, behavioural or functional setting which the AI system is intended to be used. The requirements related to data governance can be complied with by having recourse to third parties that offer certified compliance services including verification of data governance, data set integrity, and data training, validation and testing practices, as far as compliance with the data requirements of this Regulation are ensured.

Recital 68

Recital 69

Recital 70

Recital 71

Recital 72

To address concerns related to opacity and complexity of certain AI systems and help deployers to fulfil their obligations under this Regulation, transparency should be required for high-risk AI systems before they are placed on the market or put it into service. High-risk AI systems should be designed in a manner to enable deployers to understand how the AI system works, evaluate its functionality, and comprehend its strengths and limitations. High-risk AI systems should be accompanied by appropriate information in the form of instructions of use. Such information should include the characteristics, capabilities and limitations of performance of the AI system. Those would cover information on possible known and foreseeable circumstances related to the use of the high-risk AI system, including deployer action that may influence system behaviour and performance, under which the AI system can lead to risks to health, safety, and fundamental rights, on the changes that have been pre-determined and assessed for conformity by the provider and on the relevant human oversight measures, including the measures to facilitate the interpretation of the outputs of the AI system by the deployers. Transparency, including the accompanying instructions for use, should assist deployers in the use of the system and support informed decision making by them. Deployers should, inter alia, be in a better position to make the correct choice of the system that they intend to use in light of the obligations applicable to them, be educated about the intended and precluded uses, and use the AI system correctly and as appropriate. In order to enhance legibility and accessibility of the information included in the instructions of use, where appropriate, illustrative examples, for instance on the limitations and on the intended and precluded uses of the AI system, should be included. Providers should ensure that all documentation, including the instructions for use, contains meaningful, comprehensive, accessible and understandable information, taking into account the needs and foreseeable knowledge of the target deployers. Instructions for use should be made available in a language which can be easily understood by target deployers, as determined by the Member State concerned.

Recital 73

High-risk AI systems should be designed and developed in such a way that natural persons can oversee their functioning, ensure that they are used as intended and that their impacts are addressed over the system’s lifecycle. To that end, appropriate human oversight measures should be identified by the provider of the system before its placing on the market or putting into service. In particular, where appropriate, such measures should guarantee that the system is subject to in-built operational constraints that cannot be overridden by the system itself and is responsive to the human operator, and that the natural persons to whom human oversight has been assigned have the necessary competence, training and authority to carry out that role. It is also essential, as appropriate, to ensure that high-risk AI systems include mechanisms to guide and inform a natural person to whom human oversight has been assigned to make informed decisions if, when and how to intervene in order to avoid negative consequences or risks, or stop the system if it does not perform as intended. Considering the significant consequences for persons in the case of an incorrect match by certain biometric identification systems, it is appropriate to provide for an enhanced human oversight requirement for those systems so that no action or decision may be taken by the deployer on the basis of the identification resulting from the system unless this has been separately verified and confirmed by at least two natural persons. Those persons could be from one or more entities and include the person operating or using the system. This requirement should not pose unnecessary burden or delays and it could be sufficient that the separate verifications by the different persons are automatically recorded in the logs generated by the system. Given the specificities of the areas of law enforcement, migration, border control and asylum, this requirement should not apply where Union or national law considers the application of that requirement to be disproportionate.

Recital 74

Recital 75

Recital 76

Recital 77

Recital 78

Recital 79

Recital 80

Recital 81

Recital 82

Recital 83

Recital 84

Recital 85

Recital 86

Recital 87

Recital 88

Recital 89

Recital 90

Recital 91

Recital 92

Recital 93

Recital 94

Recital 95

Recital 96

In order to efficiently ensure that fundamental rights are protected, deployers of high-risk AI systems that are bodies governed by public law, or private entities providing public services and deployers of certain high-risk AI systems listed in an annex to this Regulation, such as banking or insurance entities, should carry out a fundamental rights impact assessment prior to putting it into use. Services important for individuals that are of public nature may also be provided by private entities. Private entities providing such public services are linked to tasks in the public interest such as in the areas of education, healthcare, social services, housing, administration of justice. The aim of the fundamental rights impact assessment is for the deployer to identify the specific risks to the rights of individuals or groups of individuals likely to be affected, identify measures to be taken in the case of a materialisation of those risks. The impact assessment should be performed prior to deploying the high-risk AI system, and should be updated when the deployer considers that any of the relevant factors have changed. The impact assessment should identify the deployer’s relevant processes in which the high-risk AI system will be used in line with its intended purpose, and should include a description of the period of time and frequency in which the system is intended to be used as well as of specific categories of natural persons and groups who are likely to be affected in the specific context of use. The assessment should also include the identification of specific risks of harm likely to have an impact on the fundamental rights of those persons or groups. While performing this assessment, the deployer should take into account information relevant to a proper assessment of the impact, including but not limited to the information given by the provider of the high-risk AI system in the instructions for use. In light of the risks identified, deployers should determine measures to be taken in the case of a materialisation of those risks, including for example governance arrangements in that specific context of use, such as arrangements for human oversight according to the instructions of use or, complaint handling and redress procedures, as they could be instrumental in mitigating risks to fundamental rights in concrete use-cases. After performing that impact assessment, the deployer should notify the relevant market surveillance authority. Where appropriate, to collect relevant information necessary to perform the impact assessment, deployers of high-risk AI system, in particular when AI systems are used in the public sector, could involve relevant stakeholders, including the representatives of groups of persons likely to be affected by the AI system, independent experts, and civil society organisations in conducting such impact assessments and designing measures to be taken in the case of materialisation of the risks. The European Artificial Intelligence Office (AI Office) should develop a template for a questionnaire in order to facilitate compliance and reduce the administrative burden for deployers.

Recital 97

The notion of general-purpose AI models should be clearly defined and set apart from the notion of AI systems to enable legal certainty. The definition should be based on the key functional characteristics of a general-purpose AI model, in particular the generality and the capability to competently perform a wide range of distinct tasks. These models are typically trained on large amounts of data, through various methods, such as self-supervised, unsupervised or reinforcement learning. General-purpose AI models may be placed on the market in various ways, including through libraries, application programming interfaces (APIs), as direct download, or as physical copy. These models may be further modified or fine-tuned into new models. Although AI models are essential components of AI systems, they do not constitute AI systems on their own. AI models require the addition of further components, such as for example a user interface, to become AI systems. AI models are typically integrated into and form part of AI systems. This Regulation provides specific rules for general-purpose AI models and for general-purpose AI models that pose systemic risks, which should apply also when these models are integrated or form part of an AI system. It should be understood that the obligations for the providers of general-purpose AI models should apply once the general-purpose AI models are placed on the market. When the provider of a general-purpose AI model integrates an own model into its own AI system that is made available on the market or put into service, that model should be considered to be placed on the market and, therefore, the obligations in this Regulation for models should continue to apply in addition to those for AI systems. The obligations laid down for models should in any case not apply when an own model is used for purely internal processes that are not essential for providing a product or a service to third parties and the rights of natural persons are not affected. Considering their potential significantly negative effects, the general-purpose AI models with systemic risk should always be subject to the relevant obligations under this Regulation. The definition should not cover AI models used before their placing on the market for the sole purpose of research, development and prototyping activities. This is without prejudice to the obligation to comply with this Regulation when, following such activities, a model is placed on the market.

Recital 98

Recital 99

Recital 100

Recital 101

Recital 102

Recital 103

Recital 104

Recital 105

Recital 106

Recital 107

Recital 108

Recital 109

Recital 110

General-purpose AI models could pose systemic risks which include, but are not limited to, any actual or reasonably foreseeable negative effects in relation to major accidents, disruptions of critical sectors and serious consequences to public health and safety; any actual or reasonably foreseeable negative effects on democratic processes, public and economic security; the dissemination of illegal, false, or discriminatory content. Systemic risks should be understood to increase with model capabilities and model reach, can arise along the entire lifecycle of the model, and are influenced by conditions of misuse, model reliability, model fairness and model security, the level of autonomy of the model, its access to tools, novel or combined modalities, release and distribution strategies, the potential to remove guardrails and other factors. In particular, international approaches have so far identified the need to pay attention to risks from potential intentional misuse or unintended issues of control relating to alignment with human intent; chemical, biological, radiological, and nuclear risks, such as the ways in which barriers to entry can be lowered, including for weapons development, design acquisition, or use; offensive cyber capabilities, such as the ways in vulnerability discovery, exploitation, or operational use can be enabled; the effects of interaction and tool use, including for example the capacity to control physical systems and interfere with critical infrastructure; risks from models of making copies of themselves or ‘self-replicating’ or training other models; the ways in which models can give rise to harmful bias and discrimination with risks to individuals, communities or societies; the facilitation of disinformation or harming privacy with threats to democratic values and human rights; risk that a particular event could lead to a chain reaction with considerable negative effects that could affect up to an entire city, an entire domain activity or an entire community.

Recital 111

It is appropriate to establish a methodology for the classification of general-purpose AI models as general-purpose AI model with systemic risks. Since systemic risks result from particularly high capabilities, a general-purpose AI model should be considered to present systemic risks if it has high-impact capabilities, evaluated on the basis of appropriate technical tools and methodologies, or significant impact on the internal market due to its reach. High-impact capabilities in general-purpose AI models means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. The full range of capabilities in a model could be better understood after its placing on the market or when deployers interact with the model. According to the state of the art at the time of entry into force of this Regulation, the cumulative amount of computation used for the training of the general-purpose AI model measured in floating point operations is one of the relevant approximations for model capabilities. The cumulative amount of computation used for training includes the computation used across the activities and methods that are intended to enhance the capabilities of the model prior to deployment, such as pre-training, synthetic data generation and fine-tuning. Therefore, an initial threshold of floating point operations should be set, which, if met by a general-purpose AI model, leads to a presumption that the model is a general-purpose AI model with systemic risks. This threshold should be adjusted over time to reflect technological and industrial changes, such as algorithmic improvements or increased hardware efficiency, and should be supplemented with benchmarks and indicators for model capability. To inform this, the AI Office should engage with the scientific community, industry, civil society and other experts. Thresholds, as well as tools and benchmarks for the assessment of high-impact capabilities, should be strong predictors of generality, its capabilities and associated systemic risk of general-purpose AI models, and could take into account the way the model will be placed on the market or the number of users it may affect. To complement this system, there should be a possibility for the Commission to take individual decisions designating a general-purpose AI model as a general-purpose AI model with systemic risk if it is found that such model has capabilities or an impact equivalent to those captured by the set threshold. That decision should be taken on the basis of an overall assessment of the criteria for the designation of a general-purpose AI model with systemic risk set out in an annex to this Regulation, such as quality or size of the training data set, number of business and end users, its input and output modalities, its level of autonomy and scalability, or the tools it has access to. Upon a reasoned request of a provider whose model has been designated as a general-purpose AI model with systemic risk, the Commission should take the request into account and may decide to reassess whether the general-purpose AI model can still be considered to present systemic risks.

Recital 112

Recital 113

Recital 114

Recital 115

Recital 116

Recital 117

Recital 118

Recital 119

Recital 120

Recital 121

Standardisation should play a key role to provide technical solutions to providers to ensure compliance with this Regulation, in line with the state of the art, to promote innovation as well as competitiveness and growth in the single market. Compliance with harmonised standards as defined in Article 2, point (1)(c), of Regulation (EU) No 1025/2012 of the European Parliament and of the Council41, which are normally expected to reflect the state of the art, should be a means for providers to demonstrate conformity with the requirements of this Regulation. A balanced representation of interests involving all relevant stakeholders in the development of standards, in particular SMEs, consumer organisations and environmental and social stakeholders in accordance with Articles 5 and 6 of Regulation (EU) No 1025/2012 should therefore be encouraged. In order to facilitate compliance, the standardisation requests should be issued by the Commission without undue delay. When preparing the standardisation request, the Commission should consult the advisory forum and the Board in order to collect relevant expertise. However, in the absence of relevant references to harmonised standards, the Commission should be able to establish, via implementing acts, and after consultation of the advisory forum, common specifications for certain requirements under this Regulation. The common specification should be an exceptional fall back solution to facilitate the provider’s obligation to comply with the requirements of this Regulation, when the standardisation request has not been accepted by any of the European standardisation organisations, or when the relevant harmonised standards insufficiently address fundamental rights concerns, or when the harmonised standards do not comply with the request, or when there are delays in the adoption of an appropriate harmonised standard. Where such a delay in the adoption of a harmonised standard is due to the technical complexity of that standard, this should be considered by the Commission before contemplating the establishment of common specifications. When developing common specifications, the Commission is encouraged to cooperate with international partners and international standardisation bodies.

Recital 122

Recital 123

Recital 124

Recital 125

Recital 126

Recital 127

Recital 128

Recital 129

Recital 130

Recital 131

In order to facilitate the work of the Commission and the Member States in the AI field as well as to increase the transparency towards the public, providers of high-risk AI systems other than those related to products falling within the scope of relevant existing Union harmonisation legislation, as well as providers who consider that an AI system listed in the high-risk use cases in an annex to this Regulation is not high-risk on the basis of a derogation, should be required to register themselves and information about their AI system in an EU database, to be established and managed by the Commission. Before using an AI system listed in the high-risk use cases in an annex to this Regulation, deployers of high-risk AI systems that are public authorities, agencies or bodies, should register themselves in such database and select the system that they envisage to use. Other deployers should be entitled to do so voluntarily. This section of the EU database should be publicly accessible, free of charge, the information should be easily navigable, understandable and machine-readable. The EU database should also be user-friendly, for example by providing search functionalities, including through keywords, allowing the general public to find relevant information to be submitted upon the registration of high-risk AI systems and on the use case of high-risk AI systems, set out in an annex to this Regulation, to which the high-risk AI systems correspond. Any substantial modification of high-risk AI systems should also be registered in the EU database. For high-risk AI systems in the area of law enforcement, migration, asylum and border control management, the registration obligations should be fulfilled in a secure non-public section of the EU database. Access to the secure non-public section should be strictly limited to the Commission as well as to market surveillance authorities with regard to their national section of that database. High-risk AI systems in the area of critical infrastructure should only be registered at national level. The Commission should be the controller of the EU database, in accordance with Regulation (EU) 2018/1725. In order to ensure the full functionality of the EU database, when deployed, the procedure for setting the database should include the development of functional specifications by the Commission and an independent audit report. The Commission should take into account cybersecurity risks when carrying out its tasks as data controller on the EU database. In order to maximise the availability and use of the EU database by the public, the EU database, including the information made available through it, should comply with requirements under the Directive (EU) 2019/882.

Recital 132

Recital 133

A variety of AI systems can generate large quantities of synthetic content that becomes increasingly hard for humans to distinguish from human-generated and authentic content. The wide availability and increasing capabilities of those systems have a significant impact on the integrity and trust in the information ecosystem, raising new risks of misinformation and manipulation at scale, fraud, impersonation and consumer deception. In light of those impacts, the fast technological pace and the need for new methods and techniques to trace origin of information, it is appropriate to require providers of those systems to embed technical solutions that enable marking in a machine readable format and detection that the output has been generated or manipulated by an AI system and not a human. Such techniques and methods should be sufficiently reliable, interoperable, effective and robust as far as this is technically feasible, taking into account available techniques or a combination of such techniques, such as watermarks, metadata identifications, cryptographic methods for proving provenance and authenticity of content, logging methods, fingerprints or other techniques, as may be appropriate. When implementing this obligation, providers should also take into account the specificities and the limitations of the different types of content and the relevant technological and market developments in the field, as reflected in the generally acknowledged state of the art. Such techniques and methods can be implemented at the level of the AI system or at the level of the AI model, including general-purpose AI models generating content, thereby facilitating fulfilment of this obligation by the downstream provider of the AI system. To remain proportionate, it is appropriate to envisage that this marking obligation should not cover AI systems performing primarily an assistive function for standard editing or AI systems not substantially altering the input data provided by the deployer or the semantics thereof.

Recital 134

Recital 135

Recital 136

Recital 137

Recital 138

Recital 139

The objectives of the AI regulatory sandboxes should be to foster AI innovation by establishing a controlled experimentation and testing environment in the development and pre-marketing phase with a view to ensuring compliance of the innovative AI systems with this Regulation and other relevant Union and national law. Moreover, the AI regulatory sandboxes should aim to enhance legal certainty for innovators and the competent authorities’ oversight and understanding of the opportunities, emerging risks and the impacts of AI use, to facilitate regulatory learning for authorities and undertakings, including with a view to future adaptions of the legal framework, to support cooperation and the sharing of best practices with the authorities involved in the AI regulatory sandbox, and to accelerate access to markets, including by removing barriers for SMEs, including start-ups. AI regulatory sandboxes should be widely available throughout the Union, and particular attention should be given to their accessibility for SMEs, including start-ups. The participation in the AI regulatory sandbox should focus on issues that raise legal uncertainty for providers and prospective providers to innovate, experiment with AI in the Union and contribute to evidence-based regulatory learning. The supervision of the AI systems in the AI regulatory sandbox should therefore cover their development, training, testing and validation before the systems are placed on the market or put into service, as well as the notion and occurrence of substantial modification that may require a new conformity assessment procedure. Any significant risks identified during the development and testing of such AI systems should result in adequate mitigation and, failing that, in the suspension of the development and testing process. Where appropriate, national competent authorities establishing AI regulatory sandboxes should cooperate with other relevant authorities, including those supervising the protection of fundamental rights, and could allow for the involvement of other actors within the AI ecosystem such as national or European standardisation organisations, notified bodies, testing and experimentation facilities, research and experimentation labs, European Digital Innovation Hubs and relevant stakeholder and civil society organisations. To ensure uniform implementation across the Union and economies of scale, it is appropriate to establish common rules for the AI regulatory sandboxes’ implementation and a framework for cooperation between the relevant authorities involved in the supervision of the sandboxes. AI regulatory sandboxes established under this Regulation should be without prejudice to other law allowing for the establishment of other sandboxes aiming to ensure compliance with law other than this Regulation. Where appropriate, relevant competent authorities in charge of those other regulatory sandboxes should consider the benefits of using those sandboxes also for the purpose of ensuring compliance of AI systems with this Regulation. Upon agreement between the national competent authorities and the participants in the AI regulatory sandbox, testing in real world conditions may also be operated and supervised in the framework of the AI regulatory sandbox.

Recital 140

Recital 141

Recital 142

Recital 143

Recital 144

Recital 145

Recital 146

Recital 147

Recital 148

Recital 149

In order to facilitate a smooth, effective and harmonised implementation of this Regulation a Board should be established. The Board should reflect the various interests of the AI eco-system and be composed of representatives of the Member States. The Board should be responsible for a number of advisory tasks, including issuing opinions, recommendations, advice or contributing to guidance on matters related to the implementation of this Regulation, including on enforcement matters, technical specifications or existing standards regarding the requirements established in this Regulation and providing advice to the Commission and the Member States and their national competent authorities on specific questions related to AI. In order to give some flexibility to Member States in the designation of their representatives in the Board, such representatives may be any persons belonging to public entities who should have the relevant competences and powers to facilitate coordination at national level and contribute to the achievement of the Board’s tasks. The Board should establish two standing sub-groups to provide a platform for cooperation and exchange among market surveillance authorities and notifying authorities on issues related, respectively, to market surveillance and notified bodies. The standing subgroup for market surveillance should act as the administrative cooperation group (ADCO) for this Regulation within the meaning of Article 30 of Regulation (EU) 2019/1020. In accordance with Article 33 of that Regulation, the Commission should support the activities of the standing subgroup for market surveillance by undertaking market evaluations or studies, in particular with a view to identifying aspects of this Regulation requiring specific and urgent coordination among market surveillance authorities. The Board may establish other standing or temporary sub-groups as appropriate for the purpose of examining specific issues. The Board should also cooperate, as appropriate, with relevant Union bodies, experts groups and networks active in the context of relevant Union law, including in particular those active under relevant Union law on data, digital products and services.

Recital 150

Recital 151

Recital 152

Recital 153

Recital 154

Recital 155

Recital 156

Recital 157

Recital 158