{
  "id": "eu-ai-act",
  "title": "EU AI Act",
  "lang": "en",
  "source_lang": "en",
  "summary": "The EU AI Act establishes harmonized rules for artificial intelligence, combining prohibited practices, risk-based system requirements, general-purpose model duties, and regulatory oversight.",
  "category": "society",
  "subcategory": "law",
  "type": "work",
  "aliases": [
    "European Union Artificial Intelligence Act",
    "Artificial Intelligence Act",
    "AI Act",
    "Regulation (EU) 2024/1689"
  ],
  "generated_at": "2026-11-02",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/eu-ai-act/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/eu-ai-act/",
      "json": "https://aiwiki.page/zh/eu-ai-act.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "The EU AI Act, formally Regulation (EU) 2024/1689, is a European Union regulation establishing harmonized rules for artificial intelligence. It regulates the development, marketing, deployment, and use of AI through a combination of prohibited practices, requirements for high-risk systems, transparency obligations, and rules for general-purpose AI models. Its stated objectives include improving the functioning of the internal market, promoting human-centric and trustworthy AI, and protecting health, safety, and fundamental rights while supporting innovation.\n\nThe Act is not a general prohibition on AI, nor does it subject every AI application to the same approval process. Its obligations depend on the system’s intended purpose, the context in which it operates, and the role of the organization supplying or using it. A recruitment-screening system, a medical-device component, a conversational assistant, and a general-purpose language model can therefore fall under different provisions, sometimes simultaneously.\n\nAs a regulation, the Act is directly applicable in member states, although national authorities, procedural rules, and implementing measures remain important to its operation. It forms part of a wider legal environment that includes data protection, product safety, consumer protection, employment law, and fundamental rights. Understanding its architecture requires distinguishing AI models from AI systems, prohibited uses from regulated uses, and the obligations of providers from those of organizations deploying AI."
    },
    {
      "heading": "Legislative origins, adoption, and phased application",
      "text": "The European Commission presented its legislative proposal on 21 April 2021. It followed earlier EU work on AI governance, including the Commission’s 2020 White Paper on artificial intelligence and the 2019 ethics guidelines prepared by its High-Level Expert Group on AI. The proposal combined an internal-market approach with protections against risks associated with automated decision-making, biometric identification, and safety-critical applications. Its principal legal bases are Articles 16 and 114 of the Treaty on the Functioning of the European Union, concerning personal data protection and internal-market harmonization.\n\nThe European Parliament and the Council of the European Union negotiated the legislation through the ordinary legislative procedure. A provisional political agreement was reached in December 2023. Parliament approved the agreed text on 13 March 2024, and the Council gave its final approval on 21 May 2024. The regulation is dated 13 June 2024, was published in the EU’s Official Journal on 12 July, and entered into force on 1 August 2024. These dates distinguish political agreement, formal adoption, publication, and legal entry into force.\n\nThe final text differs substantially from the original proposal. In particular, it contains a separate framework for general-purpose AI models, reflecting the growing significance of systems built on broadly capable models. It also specifies additional prohibited practices, regulates certain uses of emotion recognition and biometric categorization, and establishes an EU-level supervisory role for general-purpose AI. The resulting legislation regulates both particular applications and parts of the upstream model-supply chain.\n\nThe application schedule established in the 2024 regulation is staggered. Its opening provisions, including AI literacy duties, and its prohibitions apply from 2 February 2025. Most provisions governing general-purpose AI models and the governance framework apply from 2 August 2025. The general application date is 2 August 2026, while the high-risk classification and corresponding obligations for systems covered by Article 6(1), principally certain regulated-product systems, apply from 2 August 2027.\n\nTransitional provisions further distinguish systems and models already on the market. Providers of general-purpose AI models placed on the market before 2 August 2025 receive until 2 August 2027 to comply with the relevant obligations. Other provisions address existing high-risk systems and large-scale EU information systems. Consequently, entry into force does not mean that every substantive duty becomes enforceable on the same day; the applicable provision, product history, and transitional rule must be identified separately."
    },
    {
      "heading": "Definitions, territorial reach, and regulated actors",
      "text": "The Act defines an AI system as a machine-based system designed to operate with varying levels of autonomy and potentially capable of adaptation after deployment. For explicit or implicit objectives, it infers from inputs how to generate outputs—such as predictions, content, recommendations, or decisions—that can influence physical or virtual environments. This formulation focuses on inference rather than on a particular software architecture. It encompasses many machine-learning applications without treating all conventional software as AI merely because it automates a task.\n\nAn AI model and an AI system are related but distinct legal objects. A model may be an underlying component that requires additional software, interfaces, retrieval mechanisms, or other elements before becoming a usable system. A large language model, for example, can support a general conversational service, a recruitment application, or an educational assessment tool. The model’s general-purpose obligations and the application’s use-specific obligations must therefore be assessed separately. A system built on a general-purpose model is not automatically high-risk merely because its model is powerful or broadly capable.\n\nThe main actors include providers, deployers, importers, distributors, and authorized representatives. A provider develops an AI system or general-purpose AI model, or has one developed, and places it on the market under its own name or trademark; for systems, putting it into service can also establish provider status. A deployer uses an AI system under its authority, excluding personal, non-professional activity. These are functional roles rather than permanent labels: one company can be a provider for one product and a deployer for another.\n\nThe regulation’s territorial reach extends beyond organizations established in the EU. It covers providers placing AI systems or general-purpose models on the Union market regardless of their establishment, and deployers established or located in the Union. It also reaches certain providers and deployers outside the EU where an AI system’s output is used in the Union. Importers and distributors have their own verification and cooperation duties, while specified non-EU providers must appoint an authorized representative.\n\nSeveral exclusions limit this reach. Systems used exclusively for military, defence, or national-security purposes fall outside the Act. There are also exclusions for qualifying scientific research and development, and for research, testing, or development before market placement or putting into service, subject to distinctions concerning real-world testing. Free and open-source software receives a qualified exemption, not blanket immunity: prohibited practices and specified high-risk or transparency rules can still apply. Separately, Article 4 requires providers and deployers to take measures to ensure sufficient AI literacy among relevant personnel, considering their knowledge, experience, training, and the context of use."
    },
    {
      "heading": "Prohibited practices and their exceptions",
      "text": "Article 5 prohibits specified practices considered incompatible with the Act’s protection of safety and fundamental rights. These prohibitions are not simply the highest tier of a licensing scheme: where a practice meets the legal conditions, ordinary compliance documentation does not make it permissible. Conversely, the wording and exceptions matter. Describing an application as manipulative, biometric, or predictive does not by itself establish that it falls within a prohibition.\n\nThe manipulation provisions address subliminal techniques beyond a person’s consciousness and purposefully manipulative or deceptive techniques. The prohibition requires material distortion of behaviour through appreciable impairment of informed decision-making, leading to decisions that cause or are reasonably likely to cause significant harm. A related provision addresses exploitation of vulnerabilities associated with age, disability, or a specific social or economic situation. The harm thresholds and causal conditions distinguish these provisions from a prohibition on every persuasive interface or personalized recommendation.\n\nThe Act also prohibits specified forms of social scoring. The restriction concerns evaluating or classifying people over time on the basis of social behaviour or known, inferred, or predicted personal characteristics when the resulting score produces certain detrimental or unfavourable treatment. Relevant circumstances include treatment in unrelated social contexts or treatment that is unjustified or disproportionate. Another prohibition covers individual criminal-offence risk assessments based solely on profiling or personality traits and characteristics. It does not exclude AI support for human assessments already based on objective, verifiable facts directly linked to criminal activity.\n\nBiometric restrictions operate at several levels. The Act prohibits creating or expanding facial-recognition databases through untargeted scraping of facial images from the internet or closed-circuit television footage. It prohibits certain biometric categorization used to infer sensitive characteristics, including political opinions, religious beliefs, or sexual orientation, subject to the provision’s specified exclusions. It also prohibits inferring emotions in workplaces and educational institutions except where the system is intended for medical or safety reasons. This is distinct from a universal prohibition on all software that analyzes facial expressions.\n\nThe use of “real-time” remote biometric identification in publicly accessible spaces for law enforcement is prohibited in principle, but narrowly defined exceptions cover certain searches for victims or missing persons, prevention of specified serious threats, and locating or identifying suspects in qualifying serious offences. Use must meet necessity and proportionality conditions and is generally subject to prior judicial or binding independent administrative authorization, with a limited urgency procedure. Member states must provide the relevant national legal basis to make use of the exceptions. The Act therefore establishes restrictive conditions rather than an unrestricted EU-wide authorization for police facial recognition."
    },
    {
      "heading": "High-risk classification and the role of intended purpose",
      "text": "The high-risk framework has two principal routes. Under Article 6(1), an AI system is high-risk where it is a safety component of a product, or itself a product, covered by specified EU harmonization legislation in Annex I, and the relevant product must undergo third-party conformity assessment before market placement or putting into service. Medical devices, machinery, and other regulated products illustrate this route, although classification depends on the particular sectoral rules rather than on the product’s name alone.\n\nThe second route concerns uses listed in Annex III. These include specified biometric applications; safety components in certain critical infrastructure; education and vocational training; employment and worker management; access to essential private and public services; law enforcement; migration, asylum, and border control; and the administration of justice and certain democratic processes. The list identifies contexts where system outputs can substantially affect safety, opportunities, or rights. It does not classify every AI use by an organization operating in one of these sectors as high-risk.\n\nIn education, covered uses include determining admission or access, evaluating learning outcomes in specified circumstances, assessing an appropriate educational level, and monitoring prohibited behaviour during tests. In employment, the list includes recruitment and candidate selection, decisions affecting employment relationships, allocation of tasks based on individual characteristics or behaviour, and worker monitoring or evaluation. A spelling tool used by a university and a system used to rank applicants for admission therefore present different classification questions.\n\nAnnex III also covers specified systems used to evaluate creditworthiness or establish credit scores, with an exception for systems used to detect financial fraud. Risk assessment and pricing in life and health insurance are included, as are certain public-benefit eligibility decisions and emergency-response applications. In the justice domain, the list addresses AI intended to assist judicial authorities in interpreting facts and law and applying the law to particular cases, while distinguishing purely ancillary administrative activities. Electoral provisions concern systems intended to influence election or referendum outcomes or voting behaviour, with an exclusion for specified organizational tools.\n\nArticle 6(3) provides a limited qualification for Annex III systems that do not pose a significant risk of harm to health, safety, or fundamental rights, including by not materially influencing decision outcomes. Listed circumstances include narrow procedural tasks, improving the result of a previously completed human activity, detecting decision patterns without replacing the assessment, and preparatory tasks. A system that performs profiling of natural persons remains high-risk under this framework. Providers relying on the qualification must document their assessment and fulfil the applicable registration obligation. Classification thus depends on intended purpose, actual decision-making function, and the statutory criteria, not solely on a supplier’s marketing description."
    },
    {
      "heading": "Duties across the high-risk system lifecycle",
      "text": "High-risk systems must satisfy requirements addressing risk management, data governance, technical documentation, record-keeping, transparency, human oversight, accuracy, robustness, and cybersecurity. These requirements operate together. A technically accurate model can still fail to comply if its limitations are not communicated, its records are inadequate, or its deployment arrangements make meaningful oversight impossible. Compliance is organized around the system’s intended purpose and reasonably foreseeable misuse throughout its lifecycle.\n\nThe risk-management system is a continuous, iterative process. Providers must identify and analyze known and reasonably foreseeable risks, estimate and evaluate relevant risks, adopt targeted management measures, and test the system. For systems involving model training, the data-governance provisions address training, validation, and testing datasets, including collection, preparation, assumptions, relevant biases, and suitability for the intended context. The requirement for appropriate training data does not imply that every dataset must represent every population equally; it requires attention to the system’s purpose and the characteristics of the persons or groups on whom it is intended to be used.\n\nTechnical documentation must enable assessment of compliance, while automatic logging supports traceability. Instructions for use must communicate relevant capabilities, limitations, expected performance, and oversight measures to deployers. Human oversight must enable designated people to understand limitations, recognize anomalies, interpret outputs appropriately, and intervene or stop operation where necessary. The Act specifically addresses over-reliance on automated outputs, often described as automation bias. Oversight is therefore not satisfied merely by placing a human name in an organizational chart.\n\nProviders must operate a quality-management system and complete the applicable conformity assessment before placing a high-risk system on the market or putting it into service. Depending on the system and legal route, assessment may involve internal controls, a notified body, or procedures integrated into sectoral product legislation. Compliant systems require an EU declaration of conformity and, where applicable, CE marking and registration. The Act does not require an independent external audit for every high-risk system. Providers must also conduct post-market monitoring and report qualifying serious incidents.\n\nDeployers have separate operational duties, including following instructions, assigning suitably competent oversight personnel, monitoring operation, and retaining logs under their control for the required period. Where they control input data, deployers must ensure that those data are relevant and sufficiently representative for the intended purpose. Employers must inform affected workers and their representatives before workplace use of a high-risk system. Obligations can shift along the supply chain: rebranding a high-risk system, substantially modifying it, or changing an originally non-high-risk system’s intended purpose so that it becomes high-risk can make another actor a provider. Contractual arrangements support information exchange but cannot simply remove statutory responsibilities."
    },
    {
      "heading": "General-purpose AI models and systemic risk",
      "text": "A general-purpose AI model is characterized by significant generality and the ability to competently perform a wide range of distinct tasks, regardless of how it is placed on the market, and by its capacity for integration into different downstream systems or applications. The category is broader than a chatbot and is not confined to text generation. It addresses models whose versatility makes them reusable across many contexts, including applications that their original developers do not directly operate.\n\nProviders of these models must prepare and maintain technical documentation describing the model and its development, including training, testing, and evaluation. They must also supply information and documentation that enable downstream system providers to understand the model’s capabilities and limitations and comply with their own obligations. This creates an information chain between an upstream model developer and a downstream organization that may build a recruitment, educational, or other regulated system around the model.\n\nTwo additional duties concern copyright and training transparency. Providers must establish a policy to comply with Union copyright law, including identifying and complying with rights reservations relevant to text and data mining under the applicable legislation. They must publish a sufficiently detailed summary of the content used to train the model, following the template provided by the AI Office. This summary is not equivalent to publication of the entire training dataset, nor does compliance itself establish that every use of copyrighted material was lawful. The Act adds obligations without replacing the underlying copyright framework.\n\nSome models are classified as presenting systemic risk. One route concerns high-impact capabilities evaluated using appropriate technical tools, indicators, and benchmarks; another allows a Commission designation based on equivalent capabilities or impact. A model is presumed to have high-impact capabilities when the cumulative computation used for training exceeds \\(10^{25}\\) floating-point operations. The threshold is a legal presumption rather than a universal scientific boundary between safe and dangerous models. The framework includes notification procedures and allows providers to submit substantiated arguments concerning classification.\n\nProviders of general-purpose models with systemic risk face additional obligations: model evaluations, including adversarial testing; assessment and mitigation of systemic risks at Union level; documentation and reporting of serious incidents; and adequate cybersecurity protection for the model and its physical infrastructure. Certain freely licensed, open-source models benefit from exemptions from specified documentation obligations, but those exemptions do not remove the copyright-policy and training-summary duties, and they do not apply in the same way to systemic-risk models. Codes of practice can help providers demonstrate compliance pending harmonized standards; they are implementation instruments, not substitutes for the regulation’s binding requirements."
    },
    {
      "heading": "Transparency, affected persons, and impact assessments",
      "text": "Article 50 establishes transparency duties for several categories of systems regardless of whether they are classified as high-risk. Providers must ensure that people interacting directly with an AI system are informed that they are dealing with AI, unless this is obvious to a reasonably well-informed, observant, and circumspect person in the circumstances. The rule includes a qualified exception for certain legally authorized criminal-law uses. Its purpose is disclosure of the nature of the interaction, not mandatory publication of source code or internal model reasoning.\n\nProviders of systems generating synthetic audio, images, video, or text must ensure that outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. The technical measures must be effective, interoperable, robust, and reliable as far as technically feasible, taking account of the content type, implementation costs, and relevant technical limitations. Exceptions address, among other things, standard editing assistance and cases where input data or its meaning are not substantially altered. Machine-readable marking and a visible notice serve different functions and should not be treated as interchangeable.\n\nDeployers have additional disclosure duties. Those using emotion-recognition or biometric-categorization systems must inform the people exposed to their operation. Deployers of systems producing or manipulating image, audio, or video content constituting a deepfake must disclose its artificial origin. For evidently artistic, creative, satirical, fictional, or analogous works, disclosure must be appropriate without hampering display or enjoyment. AI-generated or manipulated text published to inform the public on matters of public interest is also subject to disclosure requirements, with exceptions including qualifying human review or editorial control and identified editorial responsibility.\n\nThe Act provides a right to obtain clear and meaningful explanations in specified cases. It concerns decisions taken by deployers on the basis of outputs from certain Annex III high-risk systems where the decision produces legal effects or similarly significantly affects a person in a way considered adverse to health, safety, or fundamental rights. The explanation concerns the AI system’s role in the decision-making procedure and the main elements of the decision. This is not a general right to obtain model weights, proprietary source code, or an account of every internal computation. Persons who have grounds to consider that the regulation has been infringed may also lodge complaints with the relevant market-surveillance authority.\n\nCertain deployers must carry out a fundamental rights impact assessment before deploying specified Annex III systems. The requirement covers public-law bodies, private entities providing public services, and deployers of specified creditworthiness and life- or health-insurance systems, subject to the regulation’s scope and exceptions. The assessment addresses affected groups, risks, oversight arrangements, and measures to respond if risks materialize. Where a General Data Protection Regulation data protection impact assessment is also required, the AI Act assessment complements it rather than automatically replacing it."
    },
    {
      "heading": "Supervisory institutions, enforcement, and penalties",
      "text": "The Act combines national supervision with centralized responsibilities for general-purpose AI. Member states must designate competent authorities, including at least one notifying authority and one market-surveillance authority, and identify a single point of contact. Notifying authorities oversee the designation and monitoring of conformity-assessment bodies. Market-surveillance authorities investigate compliance, obtain information, and take corrective action within the applicable legal framework. In some regulated sectors, existing supervisory authorities perform relevant functions.\n\nAt EU level, the European AI Office, established within the European Commission, supports implementation and has a central role in supervising general-purpose AI obligations. The European Artificial Intelligence Board brings together member-state representatives to support coordination and consistent application. An advisory forum supplies stakeholder and technical input, while a scientific panel of independent experts provides expertise, particularly concerning general-purpose models and systemic risks. These bodies have distinct roles; the Board is not a substitute for national enforcement authorities or the Commission.\n\nAuthorities can require information, inspect compliance, and seek corrective measures. Depending on the circumstances, responses can include requiring a system or model to be brought into compliance, restricting its availability, or requiring withdrawal or recall. The general-purpose model framework gives the Commission powers concerning documentation, evaluations, and enforcement. These powers operate alongside safeguards for confidential information, trade secrets, and procedural rights. Confidentiality does not make regulated information inaccessible to an authority entitled to obtain it.\n\nThe principal maximum administrative fine for prohibited AI practices is €35 million or, for an undertaking, 7 percent of total worldwide annual turnover for the preceding financial year, whichever is higher. Specified other infringements can attract up to €15 million or 3 percent, while supplying incorrect, incomplete, or misleading information to competent bodies can attract up to €7.5 million or 1 percent. For small and medium-sized enterprises, including start-ups, the relevant rule uses the lower of the specified monetary or percentage maxima. Separate provisions permit Commission fines on general-purpose model providers of up to 3 percent of worldwide annual turnover for listed infringements.\n\nThese figures are statutory ceilings, not automatic penalties. Authorities must consider factors such as the nature, gravity, and duration of an infringement; affected persons and damage; the operator’s size; cooperation; and measures taken to mitigate harm. Member states determine whether, and to what extent, their public authorities are subject to administrative fines. Decisions remain subject to relevant procedural protections and judicial remedies. The Court of Justice of the European Union has its ordinary role in interpreting EU law, while national courts adjudicate matters within their jurisdiction."
    },
    {
      "heading": "Interaction with other laws, standards, and innovation measures",
      "text": "The AI Act does not displace other laws governing the same activity. A high-risk AI system that satisfies its technical requirements may still process personal data unlawfully, discriminate contrary to applicable law, or infringe intellectual property. The GDPR continues to govern matters such as lawful processing, special categories of personal data, information duties, and restrictions on certain solely automated decisions. The AI Act does not itself provide a general legal basis for processing personal data. Its requirements and data-protection obligations must therefore be assessed together.\n\nThe same cumulative approach applies to sectoral legislation. An AI-enabled medical device may fall under medical-device rules and the AI Act, with assessment procedures coordinated where the legislation provides for this. Employment, equality, migration, financial-services, and consumer laws continue to govern decisions made using AI. The Act also differs from the Digital Services Act, which principally addresses intermediary services and online platforms, and the Digital Markets Act, which imposes obligations on designated digital gatekeepers. A single business can fall under all three instruments for different aspects of its activities.\n\nTechnical standards have a major implementation role. Compliance with harmonized standards whose references are published in the Official Journal can create a presumption of conformity with the requirements those standards cover. This mechanism translates broad statutory requirements—such as risk management, robustness, or record-keeping—into more detailed technical practices. It does not mean that a standards organization can rewrite the regulation. The Commission can adopt common specifications under defined conditions, and legal conformity remains tied to the Act’s requirements and the system’s actual characteristics.\n\nThe regulation also establishes regulatory sandboxes: controlled frameworks in which providers or prospective providers can develop, train, test, and validate innovative AI under regulatory supervision. Under the timetable in the adopted text, each member state must ensure that at least one national sandbox is operational by 2 August 2026, with possibilities for joint arrangements. Sandboxes are intended to support learning about compliance and risk before wider deployment. Participation does not generally remove liability under other law or turn otherwise prohibited conduct into an unrestricted experiment.\n\nSeparate provisions address real-world testing of certain high-risk systems outside sandboxes, including testing plans, supervision, time limits, and protections for participants. Measures for smaller enterprises include priority access to sandboxes where eligibility conditions are met, information and training support, and proportionate treatment in specified procedures. These mechanisms operate alongside a continuing need to define testing methods and interpret concepts such as substantial modification, significant risk, and adequate human oversight. In practice, those determinations connect the regulation’s legal categories to the changing capabilities, deployment environments, and supply chains of AI systems."
    }
  ],
  "links": [
    "european-union",
    "artificial-intelligence",
    "data-privacy",
    "consumer-protection",
    "human-rights",
    "european-parliament",
    "council-of-the-european-union",
    "machine-learning",
    "large-language-model",
    "disability",
    "internet",
    "infrastructure",
    "cybersecurity",
    "training-data",
    "copyright",
    "general-data-protection-regulation",
    "trade-secret",
    "court-of-justice-of-the-european-union",
    "intellectual-property",
    "digital-markets-act"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}