Artificial Intelligence Act

Article 1 Subject matter

The regulation aims to improve the internal market and promote trustworthy AI while protecting health, safety, fundamental rights, democracy, rule of law and environmental protection against harmful AI systems. It establishes harmonised rules on placing AI systems on market and use, prohibits certain practices, sets requirements for high-risk systems, mandates transparency for certain systems, covers general-purpose AI models, and includes market monitoring and innovation support provisions.

Article 2 Scope

This Regulation applies to providers of AI systems and general-purpose models in the Union regardless of location, Union deployers, and parties whose output is used in the Union, including importers, distributors, manufacturers, and authorized representatives; for high-risk systems related to products in Annex I Section B, only specified articles apply. Exemptions include activities outside Union law, military/defence/national security uses, AI systems not placed on market where output is used only for security purposes, third-country authorities in international cooperation, and research and development activities prior to market placement, though real-world testing activities are covered. Personal data rules under Regulation 2016/679, 2018/1725, Directive 2002/58/EC and Regulation 2016/680 apply, and natural persons using AI in purely personal non-professional activity are exempt.

Article 3 Definitions

Article 3 defines key terms including AI systems designed for autonomy that may adapt and infer outputs; roles such as provider, deployer, authorized representative, importer and distributor; placing on market as first Union availability and putting into service as supply for intended first use; recall and withdrawal of systems; conformity assessment performed by bodies and notified bodies; training, validation and testing datasets, input data, and substantial modification; biometric data, biometric identification and verification; sensitive operational data on crime-related law enforcement; law enforcement activities, the AI Office, national competent authorities, and serious incidents including death, critical infrastructure disruption, fundamental rights infringement or environmental harm; personal data, profiling, testing frameworks and AI literacy; informed consent, deepfakes, widespread infringement across Member States, general-purpose AI models with significant generality and high-impact capabilities, systemic risk, and downstream providers integrating AI models.

Article 4 AI literacy

Providers and deployers must ensure their staff and persons handling AI systems have sufficient AI literacy appropriate to their technical knowledge, experience, education and the systems' intended use and target populations.

Article 5 Prohibited AI practices

Prohibits AI systems deploying subliminal, manipulative or deceptive techniques causing material behaviour distortion and significant harm; exploiting age, disability or socio-economic vulnerabilities; creating social scores with detrimental treatment in unrelated contexts; assessing criminal risk based solely on profiling; scraping facial images for recognition databases; or inferring workplace or education emotions except for medical or safety purposes. Prohibits biometric categorisation systems inferring ethnicity, political views, union status, faith, sexuality or sexual identity from biometric data, except for lawfully obtained dataset labelling, filtering, or law enforcement categorisation. Restricts real-time remote biometric identification in public spaces for law enforcement to targeted searches for abduction/trafficking victims or missing persons, preventing imminent threats to life, or investigating specific serious crimes; requires fundamental rights impact assessment and EU database registration before use, though urgent cases may start without registration if completed promptly; requires prior judicial or independent administrative authority authorisation based on necessity and proportionality, except emergencies where authorisation must be sought within 24 hours.

Article 6 Classification rules for high-risk AI systems

AI systems are high-risk if they are safety components of or products under Union harmonisation laws requiring third-party conformity assessment, or if listed in Annex III; however, Annex III systems are not high-risk where they pose no significant harm risk and meet conditions such as performing narrow procedural tasks, improving prior human work, or detecting decision-making deviations without replacing human assessment, except that systems performing profiling are always high-risk. Providers must document assessments and register systems, and the Commission shall issue implementation guidelines by February 2026 and may adopt delegated acts to add, modify or delete conditions based on evidence of risk to health, safety or fundamental rights, provided amendments maintain overall protection levels and consistency with other delegated acts.

Article 7 Amendments to Annex III

The Commission may add or modify high-risk AI system use-cases in Annex III when they pose risks to health, safety or fundamental rights equivalent to or greater than existing listed systems, considering the system's purpose, extent of use, data nature and quantity, autonomy level, past harm or documented concerns, potential harm scope and intensity, persons' dependency on outcomes, power imbalances or vulnerability, outcome reversibility, and deployment benefits. The Commission may remove high-risk AI systems from Annex III when they no longer pose material risks to fundamental rights, health or safety, and their removal does not reduce the overall level of protection provided by Union law.

Article 8 Compliance with the requirements

Providers must ensure high-risk AI systems comply with section requirements based on intended purpose and current state of the art, with a risk management system in place. For products combining AI systems with requirements under Union harmonisation legislation, providers must ensure full compliance with all applicable requirements and may integrate testing, reporting, information and documentation processes to avoid duplication.

Article 9 Risk management system

Providers must establish a documented risk management system for high-risk AI systems covering the entire lifecycle as a continuous iterative process, identifying known and foreseeable risks to health, safety and fundamental rights, assessing risks under intended use and reasonably foreseeable misuse, evaluating emerging risks from post-market monitoring, and adopting targeted mitigation measures. High-risk AI systems must be tested throughout development and before placing on market or service to ensure consistent performance and compliance with predefined metrics and probabilistic thresholds, considering impacts on persons under 18 and other vulnerable groups, with testing potentially occurring in real-world conditions. Risk management aspects may be combined with internal risk management required under other Union law.

Article 10 Data and data governance

High-risk AI systems must use training, validation and testing data that are relevant, representative, error-free, complete, statistically appropriate and suitable for their geographical, contextual, behavioural or functional setting, with data governance practices addressing design choices, data origins, preparation, assumptions, availability, suitability, bias assessment especially for health, safety, fundamental rights and discrimination, bias mitigation, and data gaps. Providers can process special categories of personal data solely for bias detection and correction where bias detection cannot be achieved otherwise, subject to technical re-use limitations, security measures, controlled and documented access, non-transmission to third parties, deletion once bias is corrected or retention period ends, and documented processing reasons; for non-machine learning systems these rules apply only to testing data.

Article 11 Technical documentation

Providers must prepare technical documentation before placing high-risk AI systems on the market or into service, keeping it current. The documentation must demonstrate compliance with Section 2 requirements and provide information for authorities and notified bodies, including all elements in Annex IV; SMEs may use a simplified form. For systems related to harmonised product legislation, a single set of documentation covers both AI and product requirements.

Article 12 Record-keeping

High-risk AI systems must technically enable automatic event logging throughout their lifetime, recording events relevant to identifying risks or substantial modifications, post-market monitoring, and operation monitoring. For systems in Annex III point 1(a), logging must record the period of each use, the reference database checked, matching input data, and identification of natural persons verifying results.

Article 13 Transparency and provision of information to deployers

Providers must design high-risk AI systems with sufficient transparency for deployers to interpret outputs appropriately, and supply instructions for use in digital or other format containing provider identity, system characteristics and limitations including intended purpose, accuracy and robustness metrics, known risks to health, safety or rights, performance for specific persons or groups, input data specifications, human oversight measures, computational resources needed, system lifetime, maintenance requirements, and log mechanisms.

Article 14 Human oversight

High-risk AI systems must enable effective human supervision through appropriate interfaces and built-in or deployer-implemented measures proportionate to risks and autonomy levels. Supervisors must understand system capacities and limits, monitor for malfunctions, recognise automation bias, interpret outputs correctly, and retain ability to override or stop the system. For biometric identification systems, decisions must be separately verified and confirmed by two competent persons, unless law enforcement, migration or asylum contexts make this disproportionate.

Article 15 Accuracy, robustness and cybersecurity

High-risk AI systems must achieve and maintain appropriate accuracy, robustness and cybersecurity throughout their lifecycle, with accuracy levels and metrics declared in instructions for use. Systems must be resilient against errors and system or environmental faults through technical and organisational measures, eliminate or reduce biased outputs from feedback loops in systems that continue learning after placement, and be resilient against unauthorised third-party manipulation through cybersecurity addressing data poisoning, model poisoning, adversarial examples, confidentiality attacks and model flaws.

Article 16 Obligations of providers of high-risk AI systems

Providers of high-risk AI systems must ensure compliance with Section 2 requirements, display their name and contact details, operate a quality management system, keep documentation, maintain automatic logs under their control, undergo conformity assessment before placement, draw up EU declarations of conformity, affix CE marking, register systems, take corrective actions, demonstrate conformity to authorities upon request, and ensure accessibility compliance.

Article 17 Quality management system

Providers of high-risk AI systems must establish a documented quality management system covering compliance strategy, design and development procedures, testing and validation, technical specifications, data management, risk management, post-market monitoring, incident reporting, authority communications, record-keeping, supply security, and management accountability, with implementation proportionate to organisational size. Providers subject to quality management obligations under other EU sectoral law may integrate these requirements into existing systems, and financial institutions complying with EU financial services law on internal governance may consider that obligation fulfilled except for risk management, post-market monitoring and incident reporting, provided harmonised standards are taken into account.

Article 18 Documentation keeping

Providers must keep technical documentation, quality management system records, notified body approvals and decisions, and EU conformity declarations available to national authorities for 10 years after market placement or deployment. Member States set conditions for documentation retention when providers go bankrupt or cease activity. Financial institutions may maintain technical documentation under applicable financial services law.

Article 19 Automatically generated logs

Providers must retain automatically generated system logs under their control for at least six months or longer if appropriate to the system's intended purpose, subject to applicable EU or national law, particularly data protection rules. Financial institutions shall maintain logs under their financial services law documentation requirements.

Article 20 Corrective actions and duty of information

Providers who consider or suspect non-compliance must immediately take corrective action, withdraw, disable or recall the system, and inform distributors, deployers, authorised representatives and importers. If the system presents a risk, providers must immediately investigate causes with deployers and inform market surveillance authorities and relevant notified bodies about non-compliance and corrective actions.

Article 21 Cooperation with competent authorities

Upon reasoned request, providers must supply competent authorities all information and documentation demonstrating conformity in easily understood language, and provide access to automatically generated logs under their control. Information obtained is subject to confidentiality obligations.

Article 22 Authorised representatives of providers of high-risk AI systems

Providers established outside the Union must appoint a written authorised representative in the Union before marketing high-risk systems. The representative must verify conformity declaration and technical documentation, keep these and provider details available for 10 years, provide authorities access to logs and information upon request, cooperate in compliance actions, and handle registration obligations. Representatives may terminate mandates if the provider breaches obligations, informing authorities and notified bodies.

Article 23 Obligations of importers

Before marketing high-risk systems, importers must verify the provider conducted conformity assessment, prepared technical documentation, applied CE marking and provided conformity declaration and use instructions, and appointed an authorised representative. Importers must refuse non-conforming or falsified systems and inform providers and authorities of non-compliance risks. They must display contact details, maintain storage conditions preserving compliance, keep 10-year copies of notified body certificates and conformity declarations, and cooperate with authorities.

Article 24 Obligations of distributors

Before distributing high-risk AI systems, distributors must verify CE marking, conformity declarations, instructions, and provider/importer compliance. Distributors must prevent non-conforming systems from reaching the market, maintain proper storage and transport conditions, take corrective actions or withdrawals if issues arise, provide documentation to authorities upon request, and cooperate with market surveillance.

Article 25 Responsibilities along the AI value chain

Distributors, importers, deployers and third parties become high-risk AI system providers if they brand an existing system, make substantial modifications retaining high-risk status, or change intended purpose making a non-high-risk system become high-risk; the original provider must cease being a provider but shall closely cooperate and provide necessary information and technical assistance unless it explicitly specified the system was not to become high-risk. High-risk AI system providers and third-party suppliers of tools, services, components or processes must specify by written agreement the information, capabilities and access needed for full compliance; free and open-source tools and components are exempt, and the AI Office may develop voluntary model contract terms for such agreements and publish them free of charge.

Article 26 Obligations of deployers of high-risk AI systems

Deployers of high-risk AI systems must use them per instructions, assign human oversight, ensure input data relevance and representativeness where controlled, monitor operation, keep logs for at least six months, inform workers before workplace deployment, and report serious incidents or foreseeable risks to authorities; financial institutions may comply through existing financial services law. Public authorities and deployers must register high-risk systems and conduct data protection impact assessments; post-remote biometric identification in criminal investigations requires prior court or binding administrative authorization except for initial suspect identification, and untargeted biometric use is prohibited. Deployers must document each post-remote biometric use in police files, disclose to market surveillance and data protection authorities, provide annual biometric use reports, and inform individuals of system use except where law enforcement directives apply.

Article 27 Fundamental rights impact assessment for high-risk AI systems

Public bodies and private entities providing public services deploying certain high-risk AI systems must assess impacts on fundamental rights before first use, including describing deployment context, affected persons, identified risks, human oversight measures, and risk mitigation arrangements. They may reuse earlier assessments in similar cases and must notify the market surveillance authority and update information if elements change.

Article 28 Notifying authorities

Each Member State must designate at least one notifying authority to assess and designate conformity assessment bodies. These authorities must be independent from assessment bodies, make notification decisions separately from assessments, refrain from providing commercial services, maintain confidentiality, and employ competent personnel with expertise in AI, IT and law.

Article 29 Application of a conformity assessment body for notification

Conformity assessment bodies apply to notifying authorities in their Member State with documentation describing their assessment activities, modules and AI system types covered, along with accreditation certificates or equivalent evidence. They must update documentation when relevant changes occur.

Article 30 Notification procedure

Notifying authorities notify the Commission and other Member States using an electronic tool. A conformity assessment body may operate as notified only if no objections arise within two weeks (with accreditation certificate) or two months (with other evidence). The Commission decides on objections after consulting relevant parties.

Article 31 Requirements relating to notified bodies

Notified bodies must be legally established, independent of high-risk AI system providers and competitors, have organisational structures safeguarding impartiality, maintain confidentiality, exclude personnel from involvement in design or marketing of assessed systems, and maintain documented procedures protecting professional secrecy except as required by law. Notified bodies must carry liability insurance (unless assumed by the Member State), possess high professional integrity and relevant competence, maintain sufficient internal staff with expertise in AI systems and relevant requirements, and participate in coordination activities and European standardisation organisations.

Article 32 Presumption of conformity with requirements relating to notified bodies

A conformity assessment body demonstrating compliance with relevant harmonised standards published in the Official Journal is presumed to meet the requirements for notified bodies to the extent those standards cover them.

Article 33 Subsidiaries of notified bodies and subcontracting

Notified bodies must ensure subcontractors and subsidiaries meet requirements laid down in Article 31, inform the notifying authority, and take full responsibility for their work. Activities may only be subcontracted with provider agreement, and notified bodies must publicly list their subsidiaries. Documentation on qualification assessment and work performed must be kept available to the notifying authority for five years after subcontracting ends.

Article 34 Operational obligations of notified bodies

Notified bodies must verify high-risk AI system conformity according to Article 43 procedures, avoiding unnecessary burdens on providers while accounting for their size, sector, structure and system complexity to minimise costs for micro and small enterprises. They must maintain required rigour and protection levels, and provide all relevant documentation to notifying authorities upon request.

Article 35 Identification numbers and lists of notified bodies

The Commission assigns each notified body a single identification number even if notified under multiple Union acts. The Commission must publicly maintain a current list of notified bodies including their identification numbers and activities they are notified for.

Article 36 Changes to notifications

Notifying authorities must notify the Commission and other Member States of relevant notification changes via electronic tool, with scope extensions following Articles 29 and 30 procedures. Notified bodies ceasing activities must inform the notifying authority and providers at least one year before; their certificates may remain valid for nine months if another notified body confirms it will assume responsibilities and complete full assessment within that period. Within 10 days of suspension, restriction or withdrawal, notified bodies must inform affected providers; authorities must keep notified body files available to other authorities and market surveillance, assess certificate impact within three months, and report to the Commission and Member States with suspension or withdrawal of unduly issued certificates.

Article 37 Challenge to the competence of notified bodies

The Commission may investigate cases where a notified body's competence or continued meeting of Article 31 requirements is doubted. Notifying authorities must provide relevant information on request, and the Commission must keep sensitive information confidential under Article 78. If the Commission finds a notified body unqualified, it requests corrective measures from the Member State or may suspend, restrict or withdraw designation via implementing act.

Article 38 Coordination of notified bodies

The Commission must ensure coordination and cooperation between notified bodies through a sectoral group. Each notifying authority must ensure its bodies participate directly or through representatives. The Commission must facilitate knowledge and best practices exchange between notifying authorities.

Article 39 Conformity assessment bodies of third countries

Third country bodies established under law of countries with which the Union has agreements may be authorised as notified bodies if they meet Article 31 requirements or ensure equivalent compliance.

Article 40 Harmonised standards and standardisation deliverables

High-risk and general-purpose AI systems complying with published harmonised standards are presumed conformant with applicable regulatory requirements. The Commission must request standardisation covering all relevant requirements, including resource efficiency and energy-efficient development, consulting the Board and stakeholders. Standardisation participants shall promote investment, innovation, legal certainty, market competitiveness, global cooperation and balanced multi-stakeholder governance consistent with Union values and fundamental rights.

Article 41 Common specifications

The Commission may establish implementing acts with technical specifications for requirements in Chapters IV and V if European standardisation organisations fail to deliver, delay, or produce inadequate harmonised standards, and no standards are published; systems complying with these common specifications are presumed to meet applicable requirements. Providers not conforming to common specifications must justify that their alternative technical solutions achieve equivalent compliance. Member States finding common specifications inadequate must inform the Commission with detailed explanation, and the Commission shall assess and amend the implementing act if appropriate.

Article 42 Presumption of conformity with certain requirements

High-risk AI systems trained and tested on data reflecting their specific geographical, behavioural, contextual or functional setting are presumed compliant with requirements on data quality. High-risk AI systems certified or conforming to statements under the EU cybersecurity scheme with published references are presumed compliant with cybersecurity requirements.

Article 43 Conformity assessment

For high-risk AI systems in Annex III point 1, providers may choose internal control or notified body assessment if harmonised standards or common specifications exist and are applied, but must use notified body assessment where standards do not exist, are not applied, or were published with restrictions; other Annex III systems use internal control only. High-risk systems under existing Union harmonisation legislation must follow those procedures with AI requirements incorporated, and notified bodies competent under those laws may assess AI conformity if their compliance was assessed during notification. Manufacturers opting out of third-party assessment must apply all harmonised standards or common specifications covering AI requirements.

Article 44 Certificates

Notified body certificates must be in a language understood by relevant national authorities. Validity periods are five years maximum for Annex I systems and four years for Annex III systems, extendable by similar periods upon re-assessment. Notified bodies must suspend, withdraw or restrict non-compliant systems unless corrective action is taken, providing reasons; an appeal procedure must be available.

Article 45 Information obligations of notified bodies

Notified bodies must inform their notifying authority of issued technical documentation assessments and quality management approvals, refusals or suspensions, changes affecting scope, information requests from market surveillance authorities, and on request their conformity assessment activities. They must inform other notified bodies of refused or withdrawn approvals and certificates, and provide relevant information on conformity assessment results.

Article 46 Derogation from conformity assessment procedure

Market surveillance authorities may exceptionally authorize placing or putting into service specific high-risk AI systems without full conformity assessment for public security, health, environmental or infrastructure protection. Law-enforcement or civil protection authorities may use such systems urgently without prior authorization if requested during or after use; refusal requires immediate cessation and discarding results. Authorization requires compliance with core requirements and must be notified to the Commission and other Member States; no objection within 15 days deems it justified.

Article 47 EU declaration of conformity

Providers must create a written, machine-readable EU declaration for each systemic-risk system, stating compliance with Section 2 requirements, including information per Annex V, and retain it for 10 years after market placement. Where other EU law also requires such a declaration, a single combined declaration applies.

Article 48 CE marking

High-risk AI systems must bear CE marking following principles in Regulation (EC) No 765/2008. Digital systems must have digital CE marking accessible via interface, machine-readable code or electronic means. For physical systems, marking shall be visible, legible and indelible; if impractical, it shall be affixed to packaging or documentation.

Article 49 Registration

Providers or authorized representatives must register high-risk AI systems listed in Annex III (except point 2) in the EU database before market placement or putting into service. Providers concluding systems are not high-risk per Article 6(3) must also register. Public authorities must register use before deployment.

Article 50 Transparency obligations for providers and deployers of certain AI systems

Providers must inform users of AI interaction unless obvious, and must mark artificially generated audio, images, videos or text in machine-readable format with effective, interoperable and robust solutions where technically feasible. Providers of emotion recognition or biometric categorisation systems and deployers must inform exposed individuals and comply with data protection rules. Deployers of deepfake-generating systems must disclose artificial generation or manipulation except for law enforcement and artistic, creative, satirical or fictional works with appropriate disclosure; deployers of systems producing artificially-generated text for public information must disclose its artificial nature unless authorized by law or the content has undergone human editorial review.

Article 51 Classification of general-purpose AI models as general-purpose AI models with systemic risk

A general-purpose AI model is classified as systemic risk if it has high-impact capabilities assessed through technical tools and benchmarks, or meets equivalent criteria determined by the Commission; a presumption of high-impact capabilities applies when training computation exceeds 10^25 floating point operations. The Commission may update these thresholds and benchmarks by delegated acts reflecting technological developments.

Article 52 Procedure

Providers must notify the Commission within two weeks of meeting systemic risk conditions, including substantiating information; providers may argue their model lacks systemic risks due to specific characteristics. The Commission may designate models as systemic risk ex officio or following qualified alert; designated providers may request reassessment at least six months after designation decision, providing new objective reasons. The Commission shall maintain and publish an updated list of systemic-risk models.

Article 53 Obligations for providers of general-purpose AI models

Providers must maintain technical documentation including training, testing and evaluation results; provide information and documentation to downstream AI system providers enabling understanding of capabilities and compliance; implement copyright compliance policies including identifying reserved rights; and publish training content summaries using AI Office template. Obligations on documentation and information need not apply to free and open-source models with publicly available parameters, except systemic-risk models.

Article 54 Authorised representatives of providers of general-purpose AI models

Third-country providers must appoint a written-mandated EU-established authorised representative before market placement. Representatives must verify technical documentation and Article 53-55 compliance, maintain documentation copies for ten years, provide information to the AI Office upon request, and cooperate with authorities; they may terminate the mandate if the provider acts contrary to obligations and must notify the AI Office. The requirement does not apply to free and open-source models with publicly available parameters, unless they present systemic risks.

Article 55 Obligations of providers of general-purpose AI models with systemic risk

Providers must conduct model evaluation using standardized protocols including adversarial testing, assess and mitigate systemic risks, document and report serious incidents to the AI Office, and maintain adequate cybersecurity. They may rely on approved codes of practice or comply with harmonised standards to demonstrate compliance; those not doing so must demonstrate alternative adequate compliance means.

Article 56 Codes of practice

The AI Office shall encourage development of codes of practice covering articles 53 and 55 obligations, addressing training data transparency and systemic risk identification, with participants reporting regularly on implementation. Codes must set clear objectives with commitments and key performance indicators and be ready by 2 May 2025; non-systemic risk providers may adhere to article 53 obligations only unless opting in fully. If finalisation fails or proves inadequate by 2 August 2025, the Commission may establish common rules via implementing acts under the examination procedure.

Article 57 AI regulatory sandboxes

Member States must ensure competent authorities establish at least one operational AI regulatory sandbox by 2 August 2026, which may be joint or part of existing sandboxes; the European Data Protection Supervisor may establish one for Union institutions. Competent authorities shall provide guidance on regulatory expectations and written proof of success; exit reports may demonstrate compliance and are positively considered by market surveillance authorities and notified bodies for conformity assessment. Data protection and other national authorities must participate in supervision where personal data is processed or within their remit; competent authorities retain full supervisory powers, may suspend testing if significant risks to health, safety or fundamental rights cannot be mitigated, and must inform the AI Office.

Article 58 Detailed arrangements for, and functioning of, AI regulatory sandboxes

The Commission shall adopt implementing acts establishing common principles on AI regulatory sandbox eligibility, application and participation, requiring transparent selection criteria and decisions within three months, with free access for SMEs and start-ups except proportionate exceptional costs. Sandboxes must facilitate testing and benchmarking of AI system dimensions such as accuracy, robustness and cybersecurity, and measures to mitigate risks to fundamental rights, directing prospective providers especially SMEs and start-ups to pre-deployment services. Authorities may authorise real-world testing with agreed safeguards protecting fundamental rights, health and safety, and shall cooperate where appropriate for consistent practices.

Article 59 Further processing of personal data for developing certain AI systems in the public interest in the AI regulatory sandbox

Personal data obtained for other purposes may be used to develop, train and test AI systems in regulatory sandboxes when developed for substantial public interest in areas such as public safety, health, environment, energy, transport, infrastructure or administration, on conditions that the data is necessary, kept in isolated environments with limited access, not shared outside the sandbox, not affecting data subjects' rights, technically protected and deleted at sandbox exit with logs kept throughout. Sandbox processing must include effective risk monitoring and mitigation mechanisms, cannot share personal data outside the sandbox except per data protection law, and AI project summaries must be published on competent authority websites except for sensitive law enforcement, border control or immigration operational data.

Article 60 Testing of high-risk AI systems in real world conditions outside AI regulatory sandboxes

High-risk AI system testing in real-world conditions outside regulatory sandboxes requires market surveillance authority approval or tacit approval within 30 days, a Union-wide unique identification number, duration not exceeding six months with optional extension, informed consent from test subjects, vulnerability protections, qualified person oversight, and reversibility of predictions or recommendations. Market surveillance authorities may inspect testing and demand information, and providers must immediately report serious incidents to national authorities and mitigate or suspend testing. Providers and prospective providers are liable for damages under applicable law and must notify authorities of suspension, termination or final outcomes.

Article 61 Informed consent to participate in testing in real world conditions outside AI regulatory sandboxes

Test subjects must give freely-given informed consent after receiving clear information about testing objectives, conditions, duration, rights to refuse or withdraw without detriment, how to request reversal of AI decisions, and testing identification number and provider contact details. Consent must be dated, documented and copied to subjects.

Article 62 Measures for providers and deployers, in particular SMEs, including start-ups

Member States must give SMEs and start-ups priority access to AI regulatory sandboxes, provide tailored training on the Regulation, establish communication channels, and reduce conformity assessment fees proportionately. The AI Office must provide templates, maintain an information platform, conduct awareness campaigns, and promote convergence of public procurement practices.

Article 63 Derogations for specific operators

Microenterprises may comply with simplified quality management system elements under Article 17 without partner or linked enterprises. The Commission shall develop guidelines on simplified compliance, without reducing protection or high-risk AI system requirements, and microenterprises remain bound by all other Regulation obligations.

Article 64 AI Office

The Commission shall establish Union-wide expertise and capacities in artificial intelligence via the AI Office. Member States shall support the Office's assigned responsibilities under this Regulation.

Article 65 Establishment and structure of the European Artificial Intelligence Board

Establishes an EU AI Board composed of one representative per Member State, with the European Data Protection Supervisor as observer and the AI Office attending without voting. Representatives serve three-year terms renewable once and must have relevant competences and powers to contribute to Board tasks, coordinate national authorities, and facilitate stakeholder engagement.

Article 66 Tasks of the Board

The Board advises the Commission and Member States on consistent application of this Regulation. Its functions include coordinating national authorities, sharing expertise and best practices, advising on implementation (especially general-purpose AI), harmonising administrative practices, issuing recommendations and opinions on various matters including codes of conduct, delegated acts, standards, AI literacy, common criteria, and cooperation with Union bodies and third countries. The Board shall help competent authorities and the Commission build capacity for implementation, support sandbox development and inter-sandbox information exchange, provide guidance, advise on international AI matters, assess qualified alerts on general-purpose AI models, and receive Member State views on alerts and national monitoring practices.

Article 67 Advisory forum

Establishes a technical advisory body to advise the Board and Commission. Membership, appointed by the Commission for two-year terms renewably extended up to four years, balances industry, startups, SMEs, civil society, academia, and commercial versus non-commercial interests. The Fundamental Rights Agency, ENISA, CEN, CENELEC and ETSI are permanent members.

Article 68 Scientific panel of independent experts

The Commission establishes a scientific panel of experts selected based on AI expertise, independence from AI system providers, and objective capability. The panel advises the AI Office on enforcing this Regulation regarding general-purpose AI models, including identifying systemic risks, developing evaluation tools and methodologies, classifying models with systemic risk, and supporting market surveillance and Union safeguard procedures.

Article 69 Access to the pool of experts by the Member States

Member States may request scientific panel experts to help enforce this Regulation. States may pay fees set in implementing rules, considering adequate implementation, cost-efficiency and fair access for all. The Commission facilitates timely expert access and coordinates it with Union AI testing support for best added value.

Article 70 Designation of national competent authorities and single points of contact

Member States must establish or designate at least one notifying authority and one market surveillance authority exercising powers independently and impartially, report their identities and tasks to the Commission, make contact information publicly available by 2 August 2025, designate a market surveillance authority as single point of contact, and ensure authorities have adequate technical, financial, human resources and personnel with expertise in AI, data protection, cybersecurity and fundamental rights. National competent authorities may provide implementation guidance to SMEs and start-ups, and must consult the competent authorities under other Union law when providing guidance on AI systems falling under that law. The European Data Protection Supervisor acts as competent authority for Union institutions, bodies, offices and agencies within this Regulation's scope.

Article 71 EU database for high-risk AI systems listed in Annex III

The Commission shall establish and maintain an EU database containing information on high-risk AI systems registered under Articles 49 and 60, and certain non-high-risk systems registered under Article 6(4) and 49. Providers or authorised representatives enter data from Sections A and B of Annex VIII; public authority deployers enter data from Section C. Information from Article 49 is publicly accessible unless from Article 60, which is restricted to authorities unless the provider consents to public access.

Article 72 Post-market monitoring by providers and post-market monitoring plan for high-risk AI systems

Suppliers must set up proportionate post-market monitoring systems that actively gather and analyse data on high-risk AI system performance throughout their operational life to verify ongoing compliance with Chapter III Section 2 requirements. Monitoring may include analysis of interactions with other AI systems but excludes law-enforcement deployer sensitive operational data. Monitoring must follow a documented plan included in technical documentation; the Commission shall issue a template by 2 February 2026.

Article 73 Reporting of serious incidents

Providers of high-risk AI systems must report serious incidents to market surveillance authorities where the incident occurred immediately after establishing a causal link or within 15 days of becoming aware. For widespread infringements or specific serious incidents, reporting must be within 2 days. For deaths, reporting must be within 10 days or immediately upon suspicion.

Article 74 Market surveillance and control of AI systems in the Union market

Regulation (EU) 2019/1020 applies to AI systems; market surveillance authorities report annually on competition-relevant information and prohibited practices, with sectoral authorities supervising high-risk systems related to Union harmonised products. For high-risk systems used in law enforcement, border management, justice, democracy or specified purposes (Annex III points 6, 7, 8), Member States designate data protection authorities or equivalently independent authorities to supervise, with the European Data Protection Supervisor supervising Union institutions' systems. Market surveillance authorities have full access to high-risk system documentation, training, validation and testing data through appropriate technical means; source code access requires a reasoned request demonstrating necessity to assess Chapter III compliance and exhaustion of testing and documentation-based verification, with all information obtained subject to confidentiality obligations.

Article 75 Mutual assistance, market surveillance and control of general-purpose AI systems

The AI Office supervises high-risk systems based on general-purpose models developed by the same provider, with full market surveillance authority powers. Market surveillance authorities cooperate with the AI Office where general-purpose AI systems can serve high-risk purposes and are considered non-compliant. If unable to access information, authorities may request the AI Office to enforce access within 30 days.

Article 76 Supervision of testing in real world conditions by market surveillance authorities

Market surveillance authorities ensure testing in real world conditions complies with the Regulation. They verify compliance for sandbox testing and may allow derogations from certain conditions. They may suspend or terminate testing or require modifications where serious incidents occur or conditions are unmet.

Article 77 Powers of authorities protecting fundamental rights

National authorities protecting fundamental rights including non-discrimination may request high-risk AI system documentation in accessible formats when necessary for their mandates. Member States must identify and publicly list such authorities by 2 November 2024. If documentation is insufficient to assess compliance, authorities may request market surveillance authorities organise testing through technical means.

Article 78 Confidentiality

The Commission, market surveillance authorities, notified bodies and persons applying this Regulation must protect confidential business information, trade secrets, source code, security interests, criminal or administrative proceedings, and classified information, request only necessary data, implement adequate cybersecurity measures, and delete data when no longer needed; for high-risk AI systems used by law enforcement or immigration authorities, information must not be disclosed without consulting the originating authority and deployer if this would jeopardise security interests. These confidentiality requirements do not restrict the Commission, Member States and notified bodies from exchanging information, issuing warnings in cross-border cooperation, or providing information under criminal law, nor prevent the Commission and Member States from exchanging confidential information with third-country authorities under bilateral or multilateral confidentiality arrangements guaranteeing adequate protection.

Article 79 Procedure at national level for dealing with AI systems presenting a risk

Market surveillance authorities must evaluate AI systems presenting risks to health, safety or fundamental rights, particularly those threatening vulnerable groups, and require corrective actions, withdrawal or recall within 15 working days if non-compliance is found; when non-compliance is not limited to their territory, they shall notify the Commission and other Member States. Notifications must detail non-compliance reasons including violations of Article 5 prohibitions, failures to meet Chapter III Section 2 requirements, shortcomings in harmonised standards, or Article 50 breaches. A measure is deemed justified if no objection is raised within three months (or 30 days for Article 5 violations).

Article 80 Procedure for dealing with AI systems classified by the provider as non-high-risk in application of Annex III

Market surveillance authorities may evaluate AI systems providers classify as non-high-risk to determine if they are actually high-risk. If high-risk is established, providers must comply with all requirements within a prescribed period or face fines. Providers must correct all affected systems on the Union market.

Article 81 Union safeguard procedure

If objections to national measures are raised within three months (30 days for Article 5 violations), the Commission shall consult affected parties and decide within six months (60 days for Article 5 violations) whether the measure is justified. If justified, all Member States shall withdraw the system from their markets; if unjustified, the Member State shall withdraw the measure. Where unjustified measures result from shortcomings in harmonised standards, the Commission applies the procedure in Regulation (EU) No 1025/2012.

Article 82 Compliant AI systems which present a risk

Market surveillance authorities must require operators to address risks to health, safety, fundamental rights or public interest in compliant high-risk AI systems without undue delay on a timeline they prescribe. Operators must take corrective action across all affected EU market systems; Member States must inform the Commission and other Member States; the Commission then evaluates and decides whether measures are justified.

Article 83 Formal non-compliance

Market surveillance authorities must require providers to end non-compliance including missing or improper CE marking, EU declaration of conformity, database registration, authorised representative appointment, or technical documentation within a prescribed period. If non-compliance persists, authorities must restrict, prohibit or recall the high-risk AI system without delay.

Article 84 Union AI testing support structures

The Commission designates one or more Union AI testing support structures to perform AI-related testing tasks under EU Regulation 2019/1020. They also provide independent technical or scientific advice to the Board, Commission, and market surveillance authorities upon request.

Article 85 Right to lodge a complaint with a market surveillance authority

Any natural or legal person believing there has been an infringement of the Regulation may lodge complaints with the relevant market surveillance authority. Such complaints are taken into account for market surveillance activities and handled according to procedures established by market surveillance authorities.

Article 86 Right to explanation of individual decision-making

People subject to deployer decisions based on most high-risk AI systems listed in Annex III that produce legal effects or significantly adverse impacts on health, safety or fundamental rights have the right to clear, meaningful explanations of the AI system's role and main decision elements. This does not apply where EU or national law provides exceptions or restrictions in compliance with Union law.

Article 87 Reporting of infringements and protection of reporting persons

EU Directive 2019/1937 applies to reporting infringements of this Regulation and protecting persons who report them.

Article 88 Enforcement of the obligations of providers of general-purpose AI models

The Commission has exclusive power to supervise and enforce Chapter V obligations of general-purpose AI model providers, delegating implementation to the AI Office while maintaining procedural guarantees. Market surveillance authorities may request Commission action where necessary and proportionate to assist with their tasks.

Article 89 Monitoring actions

The AI Office may monitor general-purpose AI model providers' compliance with this Regulation and approved codes of practice. Downstream providers may lodge reasoned complaints indicating the provider contact, relevant facts, provisions concerned, and why infringement occurred.

Article 90 Alerts of systemic risks by the scientific panel

The scientific panel may issue a qualified alert to the AI Office when it suspects a general-purpose AI model poses a concrete, identifiable Union-level risk or meets conditions in Article 51. Upon alert, the Commission through the AI Office may exercise powers to assess the matter after informing the Board. Alerts must be reasoned and include the provider contact, relevant facts, and any additional information the panel considers relevant.

Article 91 Power to request documentation and information

The Commission may request providers of general-purpose AI models to supply documentation per Articles 53 and 55 or additional information needed to assess compliance. The AI Office may initiate structured dialogue before requesting information. Upon substantiated scientific panel request, the Commission may also issue information requests necessary for the panel's tasks.

Article 92 Power to conduct evaluations

After consulting the Board, the AI Office may evaluate general-purpose AI models to assess compliance with the Regulation or investigate systemic risks, including following a qualified alert. The Commission may appoint independent experts, including from the scientific panel, request access through APIs or technical means including source code, and prior to access may initiate structured dialogue with the provider about testing and safeguards.

Article 93 Power to request measures

The Commission may request providers to take appropriate measures to comply with Articles 53 and 54, implement mitigation measures where evaluation reveals serious systemic Union-level risk, or restrict market availability, withdraw or recall the model. Before requesting measures, the AI Office may initiate structured dialogue. If the provider offers binding commitments to address systemic risk during dialogue, the Commission may declare no further action is needed.

Article 94 Procedural rights of economic operators of the general-purpose AI model

Providers of general-purpose AI models have procedural rights under Article 18 of Regulation (EU) 2019/1020, adapted as appropriate, without prejudice to more specific procedural rights in this Regulation.

Article 95 Codes of conduct for voluntary application of specific requirements

The AI Office and Member States shall encourage codes of conduct for voluntary application of Chapter III Section 2 requirements to non-high-risk AI systems, and other specific requirements to all AI systems based on clear objectives and performance indicators. Codes may address ethical guidelines, environmental sustainability, AI literacy, inclusive design, and impact on vulnerable persons or groups. Individual providers, organisations representing them, stakeholders and civil society may draw up codes covering one or more systems.

Article 96 Guidelines from the Commission on the implementation of this Regulation

The Commission shall develop implementation guidelines on Articles 8-15 and 25 requirements, Article 5 prohibited practices, substantial modification, Article 50 transparency, consistency with Union harmonisation law listed in Annex I, and Article 3 AI system definition. Guidelines shall address SME, start-up, local public authority and affected sector needs, account for state-of-the-art AI knowledge and harmonised standards or specifications. The Commission shall update guidelines when requested or deemed necessary.

Article 97 Exercise of the delegation

The Commission holds delegated power for Articles 6(6-7), 7(1,3), 11(3), 43(5-6), 47(5), 51(3), 52(4), and 53(5-6) for five years from 1 August 2024, with automatic extension unless Parliament or Council objects three months before expiry. The Commission must report nine months before period end. Either institution may revoke the power at any time.

Article 98 Committee procedure

The Commission receives support from a committee operating as defined in Regulation (EU) No 182/2011, with Article 5 of that Regulation applying to prescribe the examination procedure.

Article 99 Penalties

Member States shall establish effective, proportionate and dissuasive penalty rules applicable to operator infringements, considering SME viability: breaching Article 5 prohibition attracts fines up to EUR 35 million or 7% of worldwide annual turnover (whichever is higher); breaching other operator or notified body obligations attracts fines up to EUR 15 million or 3% (whichever is higher); providing incorrect information attracts fines up to EUR 7.5 million or 1% (whichever is higher); for SMEs, fines use the lower percentage or amount. When determining whether to impose a fine and its amount, authorities shall consider factors such as infringement nature, gravity and duration, prior fines, operator size and turnover, cooperation with authorities, technical safeguards implemented, how the infringement was discovered, whether it was intentional or negligent, and mitigating actions. Member States may decide how fines apply to public authorities, and fines may be imposed by courts or other competent bodies.

Article 100 Administrative fines on Union institutions, bodies, offices and agencies

The European Data Protection Supervisor may impose fines on Union institutions and agencies for Regulation infringements, up to EUR 1,500,000 for prohibited AI practices and up to EUR 750,000 for other non-compliance. Fines must consider factors such as infringement nature and gravity, responsibility degree, cooperation with the Supervisor, and the institution's budget, and recipients must have an opportunity to be heard.

Article 101 Fines for providers of general-purpose AI models

The Commission may impose fines up to 3% of annual worldwide turnover or EUR 15 million (whichever is higher) on general-purpose AI model providers who intentionally or negligently infringe the regulation, fail to comply with information requests, fail to comply with Article 93 measures, or fail to provide model access for evaluation. The Court of Justice may cancel, reduce or increase imposed fines. Implementing acts with procedural safeguards shall be adopted.

Article 102 Amendment to Regulation (EC) No 300/2008

When adopting detailed measures for security equipment technical specifications and procedures concerning AI systems, the requirements of Chapter III, Section 2, of the AI Regulation must be taken into account.

Article 103 Amendment to Regulation (EU) No 167/2013

When adopting delegated acts concerning AI systems that are safety components, the requirements of Chapter III, Section 2, of the AI Regulation must be taken into account.

Article 104 Amendment to Regulation (EU) No 168/2013

When adopting delegated acts concerning AI systems that are safety components, the requirements of Chapter III, Section 2, of the AI Regulation must be taken into account.

Article 105 Amendment to Directive 2014/90/EU

When carrying out activities and adopting technical specifications and testing standards for AI systems that are safety components, the Commission must take into account the requirements of Chapter III, Section 2, of the AI Regulation.

Article 106 Amendment to Directive (EU) 2016/797

When adopting delegated and implementing acts under Directive (EU) 2016/797 concerning AI safety components, the requirements in Chapter III, Section 2 of the AI Act must be taken into account.

Article 107 Amendment to Regulation (EU) 2018/858

When adopting delegated acts concerning AI systems that are safety components, the requirements of Chapter III, Section 2, of the AI Regulation must be taken into account.

Article 108 Amendments to Regulation (EU) 2018/1139

Six amendments to Regulation (EU) 2018/1139 require that when adopting implementing, delegated, or other acts concerning AI safety components, the Commission and other bodies must take into account the requirements from Chapter III, Section 2 of the AI Act.

Article 109 Amendment to Regulation (EU) 2019/2144

When adopting implementing acts for AI systems that serve as safety components under this regulation, the Commission must consider the specifications in the section on high-risk AI system requirements.

Article 110 Amendment to Directive (EU) 2020/1828

The AI Act is added to Annex I of Directive (EU) 2020/1828 as point (68).

Article 111 AI systems already placed on the market or put into service and general-purpose AI models already placed on the marked

AI components of large-scale IT systems placed on market or in service before 2 August 2027 must comply by 31 December 2030. High-risk AI systems placed on market or in service before 2 August 2026 must comply only if they undergo significant design changes, unless intended for public authorities (compliance by 2 August 2030). General-purpose AI models placed on market before 2 August 2025 must comply by 2 August 2027.

Article 112 Evaluation and review

The Commission must assess the prohibited practices list and Annex III annually until delegated powers expire, then on 2 August 2028 and every four years for amendments and governance; on 2 August 2029 and every four years for comprehensive review; on 2 August 2028 for AI Office evaluation; on 2 August 2028 and every four years for energy-efficient standards; and on 2 August 2028 and every three years for voluntary codes. The Board, Member States and authorities must provide requested information promptly, and the Commission must consider positions from the Board, Parliament, Council and other sources when evaluating. The AI Office must develop objective evaluation methodology for inclusion in high-risk lists and prohibited practices.

Article 113 Entry into force and application

The regulation enters into force twenty days after publication in the Official Journal and applies generally from August 2, 2026, but certain chapters and articles apply earlier: Chapters I and II from February 2025, other specified provisions from August 2025, and requirements for general-purpose AI models from August 2027.