AI Act

A dual purpose: a functioning internal market and trustworthy AI that protects health, safety and fundamental rights. The interpretive anchor reached for when reading every other provision teleologically.

Extraterritorial reach. It catches providers placing a system on the market in the EU regardless of where they are established, and providers and deployers in third countries where the system’s output is used in the EU. Outside the regime sit national security and defence, scientific research and development, pre-market activity, purely personal non-professional use, and in part open-source models. Member States may keep stronger protection for employees.

The 68 definitions the whole regulation stands on. An AI system is recognised by its inference of outputs from inputs, a varying degree of autonomy and possible adaptiveness, which decides whether a given piece of software is in the regime at all. The load-bearing pair of roles is the provider (develops it, or has it developed, and places it on the market) and the deployer (uses it in the course of a professional activity). Most obligations hang on the provider.

Providers and deployers alike ensure a sufficient level of AI literacy among the staff working with AI, having regard to the context and the persons affected. It has applied since 2 February 2025 and is the most far-reaching obligation in the regulation, reaching even those who merely use AI. In practice it is met through training, internal rules and documentation of both. The AI Act’s harmonised fining tiers do not list Article 4 (Article 4 is absent from the list in Article 99(4)), so no direct fine from those tiers threatens for breaching it. National enforcement and non-monetary measures may still be relevant. The open AI-literacy course (in Czech) on this site serves to meet the obligation.

Prohibitions (a) to (h). Subliminal and manipulative techniques that materially distort decision-making; exploitation of vulnerabilities (age, disability, a social or economic situation); social scoring with unjustified consequences outside the original context; predicting criminality purely from personality profiling; the untargeted scraping of facial images from the internet and CCTV footage; inferring emotions in the workplace and in education (outside medical and safety reasons); biometric categorisation revealing sensitive traits; and real-time remote biometric identification in publicly accessible spaces for law enforcement. For the last of these, paragraphs 2 to 7 allow exhaustive exceptions (victims, an imminent threat to life, offenders in listed crimes) with the prior authorisation of a court or an independent authority. The practical test before deploying anything with AI is to run through those eight points. Regulation (EU) 2026/1744 added two more, points (ba) and (bb), applicable from 2 December 2026: non-consensual intimate material depicting an identifiable person, and child sexual abuse material. New paragraphs 1a and 1b spell out what counts as placing on the market and what technical safeguards a provider is expected to have.

Two routes into high risk. The system is a safety component of a regulated product under Annex I with third-party conformity assessment (paragraph 1), or it falls into the areas of Annex III (paragraph 2), namely biometrics, critical infrastructure, education, employment, essential services and credit, law enforcement, migration, the administration of justice. Paragraph 3 provides a filter for systems with only a preparatory or ancillary role without significant risk, but the exception has to be documented and the system registered. Profiling of natural persons is always high-risk. The step-by-step decision chart is on the AI Act card.

The Commission may by delegated acts both add to and narrow the list of high-risk uses according to criteria of the severity and probability of harm. This is the mechanism by which Annex III will keep changing; tracking it is a continuing compliance obligation in itself.

The system meets the requirements of the section having regard to its intended purpose and the generally acknowledged state of the art. For products under Annex I, compliance is integrated into the sectoral procedures so that nothing is assessed twice.

A continuous, iterative process across the whole life cycle. Identifying foreseeable risks including reasonably foreseeable misuse, design measures, testing against metrics, and particular regard for persons under 18 and otherwise vulnerable groups.

Training, validation and testing sets must be relevant, sufficiently representative and, to the best extent possible, free of errors, with data-governance practices from collection through labelling to the detection and mitigation of bias. Processing special categories of personal data to detect bias is allowed only under strict cumulative conditions. In practice the most expensive article in the regulation.

Kept up to date before the system is placed on the market and thereafter, with content per Annex IV. Small and medium-sized enterprises may use the Commission’s simplified form.

The system automatically logs events across the whole life cycle so that a risky situation, a substantial modification and the basis for post-market monitoring can be traced back.

Instructions for use setting out the system’s capabilities and limits, performance metrics, known risks, and the requirements for human oversight and maintenance. Without this article the deployer could not carry its own obligations under Article 26.

Oversight must be effective: the overseeing person understands the capabilities and limits, can interpret the output correctly, resist automation bias, and intervene or stop the system. For remote biometric identification, no action may be taken on the basis of the output alone without verification by at least two competent natural persons.

Declared levels of accuracy in the instructions, robustness against errors and against attempts at misuse (data poisoning, adversarial inputs), and the handling of feedback loops in systems that learn in operation.

A summary catalogue. Ensure compliance with Section 2, have a quality management system, keep documentation and logs, undergo conformity assessment, draw up the declaration, affix the CE marking, register, take corrective action, cooperate with the authorities, and meet accessibility requirements under Directives (EU) 2016/2102 and 2019/882.

A written quality management system covering the compliance strategy, design and development, testing, data governance, risk management, post-market monitoring, incident reporting and communication with the authorities. The counterpart of the QMS from the world of medical devices.

The technical documentation, the QMS documentation, the declaration of conformity and the certificates kept available to the authorities for ten years after the system is placed on the market.

Logs under the provider’s control are kept for a period appropriate to the purpose, at least six months.

Bring a non-conforming system into compliance, withdraw it from the market, disable it or recall it, and inform distributors, deployers, the authorised representative and importers.

On a reasoned request, demonstrate compliance and make the documentation and logs available in a language the authority understands.

A provider from a third country must have a written-mandated representative established in the EU who holds the documentation and is the authorities’ point of contact. The same logic as Article 27 GDPR.

Before placing a system on the market, verify that conformity assessment has been carried out, that documentation exists, that the system bears the CE marking and that the provider has a representative. The importer puts its name on the packaging and must not place a system it knows to be non-conforming.

Verification of markings and documents, storage and transport that do not jeopardise conformity, and a duty to withdraw or report a system presenting a risk.

The most underestimated article in the regulation. Anyone who puts their own name on a third party’s high-risk system, substantially modifies it, or turns a general-purpose system’s purpose into a high-risk one, becomes a provider with all the obligations. The original provider hands over the documentation, and component suppliers conclude written cooperation agreements. This is exactly the way integrators building on other people’s models slip into the regime.

Use it in accordance with the instructions, entrust oversight to competent persons, ensure relevant input data, monitor operation, report incidents, keep logs for at least six months, inform workers and their representatives before deployment in the workplace, and inform the persons about whom the system helps to decide.

Public-body deployers, private providers of public services, and deployers in credit scoring and life insurance describe, before first use, the processes, the groups affected, the risks of harm, human oversight and corrective measures, and notify this to the supervisory authority. It builds on the DPIA from the GDPR and, where one already exists, is added to it.

Each State designates an authority to assess, notify and monitor conformity assessment bodies, with a guarantee of impartiality towards them.

A conformity assessment body demonstrates its competence by accreditation or, where appropriate, other documentation.

Notification to the Commission and the States is done electronically, with windows for objections; only then may the body act as notified.

Independence from providers and competitors, confidentiality, expert staff, liability insurance, and internal processes proportionate to the size of the client.

Accreditation under harmonised standards raises a presumption that the requirements of Article 31 are met.

Subcontracting of assessment only with the client’s consent and with the notified body retaining full responsibility.

Assess proportionately, do not burden small providers needlessly, and keep the documentation available to the notifying authority.

The Commission assigns numbers and maintains a public list of notified bodies.

The procedure on loss of competence, suspension or withdrawal, and the fate of certificates already issued, including the transfer of files to another body.

The Commission may investigate competence and require the State to take remedial action, including withdrawing the notification.

Sectoral groups for sharing practice, so that assessment does not diverge between States.

On equivalent conditions and on the basis of agreements, bodies established outside the EU may also carry out the activity.

Conformity with a harmonised standard raises a presumption of conformity with the requirements the standard covers. The practical route to compliance; standards for the AI Act are being drawn up in CEN/CENELEC at the Commission’s request.

Where standards are missing or insufficient, the Commission may by implementing act issue its own specifications with the same presumption effect.

Two special presumptions. Training on data reflecting the deployment setting, for representativeness, and cybersecurity certification, for Article 15.

For biometrics under Annex III, point 1, the provider chooses between internal control under Annex VI and assessment with a notified body under Annex VII only where it has used harmonised standards or common specifications. Where such standards do not exist or it did not apply them in full, it must go the Annex VII route with a notified body. For the other areas of Annex III, internal control under Annex VI is enough. Systems in products under Annex I run under the sectoral procedures. A substantial modification of the system means a new assessment.

Issued by notified bodies for a limited period, with the possibility of extension, suspension and withdrawal, against which an appeal procedure exists.

Reporting of certificates issued, refused, suspended and withdrawn to the notifying authority and sharing with the other bodies.

On exceptional grounds of public security, the protection of life and health, the environment or key assets, the supervisory authority may authorise placing on the market without a completed assessment, temporarily and under control.

Through it the provider assumes responsibility for conformity for ten years, with content per Annex V.

Visibly, legibly and indelibly, digitally for digital systems, with the number of the notified body where one was involved.

The provider registers in the EU database under Article 71 before placing the system on the market; systems exempted by the Article 6(3) filter and public-body deployers register too. Sensitive areas (law enforcement, migration) go into the non-public part.

The obligations split by role. The provider ensures that a person can tell they are dealing with AI (unless it is obvious from the context) and that synthetic audio, image, video and text carry a machine-readable marking. The deployer informs persons exposed to emotion recognition or biometric categorisation, discloses the artificial origin of a deepfake (for manifestly artistic and satirical works an unobtrusive acknowledgement is enough), and labels AI text published to inform the public on matters of public interest, unless the text has undergone human editorial review and someone bears editorial responsibility for it. All of it clearly and distinguishably, at the latest at the first interaction, not buried in the terms. It is a duty to inform, not a requirement on the quality of the output; why that is not enough for legal advice is discussed in hypothesis H3 on the research map.

A model has systemic risk when its high-impact capabilities match the training-compute threshold of more than 10^25 floating-point operations, or when the Commission designates it as such of its own motion under the criteria in Annex XIII.

The provider notifies the Commission of meeting the threshold within two weeks and may argue that the model nonetheless does not present systemic risk. The Commission maintains and publishes a list of such models.

Technical documentation per Annex XI, information for downstream providers per Annex XII, a policy to comply with Union copyright law including respect for text-and-data-mining reservations, and a public summary of the training content per the AI Office template (24 July 2025, also in Czech). Open-source models get relief from the documentation, not from copyright and the summary.

Providers of general-purpose models outside the Union appoint a representative in the EU, the counterpart of Article 22.

In addition, model evaluation including adversarial testing, the assessment and mitigation of systemic risks at EU level, the reporting of serious incidents to the AI Office, and cyber protection of the model and its infrastructure.

The AI Office facilitates codes that make meeting the requirements of Chapter V practical; they were due by 2 May 2025. Adhering to a code is a way to demonstrate compliance until harmonised standards exist. The GPAI Code of Practice of 10 July 2025 (chapters on transparency, copyright, and safety and security) was endorsed by the Commission and the AI Board as a suitable voluntary tool and was signed by, among others, the large model providers.

Each State sets up at least one sandbox, after the amendment by Regulation (EU) 2026/1744 by 2 August 2027, whether alone or jointly with other States. A controlled environment for development and testing under the authorities’ guidance, with a written report and evidence as the output, which count towards conformity assessment. Participation does not remove liability, but acting in good faith under the guidance protects against fines.

Uniform criteria and procedures are laid down by Commission implementing acts; for small and medium-sized enterprises access is to be free of charge in principle.

Lawfully obtained personal data may be further processed in the sandbox to develop systems in a substantial public interest (health, the environment, energy, transport, public administration), in a separated environment, with deletion once it ends.

High-risk systems from Annex III may be tested outside the sandbox before being placed on the market, with a testing plan, registration, the informed consent of the subjects, and a cap of six months extendable once by a further six.

The free, specific and documented consent of testing participants, with the right to withdraw at any time.

Priority free access to the sandboxes, training, communication channels, and conformity-assessment fees proportionate to size.

Micro-enterprises may meet selected elements of the quality management system in a simplified way.

The Commission concentrates expertise and enforcement through the AI Office, which is at the same time the direct supervisor of general-purpose models.

One representative from each State, a coordinating body between the Commission and the national authorities.

Uniform application of the regulation, opinions, recommendations and support for aligned administrative practice, including the sandboxes.

Industry, small and medium-sized enterprises, academia and civil society advise the Board and the Commission, in a balanced composition.

An expert backbone for the Office; among other things it may issue a qualified alert about the systemic risk of a general-purpose model.

States may draw on the same pool of experts for support of their own supervision.

Each State designated, by 2 August 2025, at least one notifying authority and one market surveillance authority and a single point of contact, with a guarantee of resources and competences.

The Commission operates a public, machine-readable database of registrations under Article 49 with content per Annex VIII. Law enforcement and migration have a non-public section accessible only to the authorities.

The provider actively collects and evaluates operating experience under a plan that is part of the technical documentation; the Commission issues the template.

The provider reports a serious incident to the authority of the State where it occurred, at the latest within 15 days, within 10 days in the event of death, and within 2 days for a widespread infringement or a disruption of critical infrastructure. Investigation and a report on the remedy follow.

The market surveillance regime of Regulation (EU) 2019/1020 applies. The authorities have access to documentation, data and (under conditions) source code; for financial institutions the sectoral authorities supervise, and for biometrics for law enforcement the data protection authorities do.

Where a system builds on a general-purpose model from the same provider, the AI Office cooperates with the national supervisor so that competences do not overlap.

The authorities oversee testing under Article 60 and may suspend or end it.

Listed national fundamental-rights authorities receive access to documentation where they need it for their remit.

Protection of intellectual property, trade secrets and source code in the exercise of supervision; the exchange of information between authorities only to the extent necessary.

The national procedure for evaluation and measures (bringing into compliance, withdrawal, recall) with notification to the Commission and the other States.

The authority may review a system that escaped classification through the Article 6(3) filter and reclassify it as high-risk, with a fine for circumvention.

Disputes between States over measures are decided by the Commission.

Even a formally compliant system may present a risk to health, safety or fundamental rights; measures are then ordered as well.

A missing CE marking, declaration, registration or representative is remedied, otherwise restriction or withdrawal from the market follows.

Horizontal technical support for supervision (Union testing capacity).

Anyone may lodge a complaint with the market surveillance authority, without prejudice to other remedies.

A person affected by a decision with legal or similarly significant effects that was taken on the basis of the output of a high-risk system from Annex III has the right to a clear and meaningful explanation of the system’s role and the main elements of the decision. The quiet relative of Article 22 GDPR, but aimed at the decision-making process rather than at data processing.

Reporting of infringements of the regulation is subject to the Whistleblowing Directive (EU) 2019/1937, in the Czech Republic therefore the Whistleblower Protection Act (zákon o ochraně oznamovatelů).

An exclusive competence of the Commission, exercised through the AI Office.

The Office monitors compliance, and downstream providers may complain of infringements committed against them.

A qualified alert about systemic risk may trigger a model evaluation by the Commission.

The Commission requests documentation and further information from the model provider.

The Commission may evaluate the model, including through independent experts, with access to the model via an API or other appropriate means.

From mitigating risks, through restricting availability, to withdrawing the model from the market.

For model providers, procedural safeguards analogous to the market surveillance regime apply.

Providers of systems outside high risk may voluntarily undertake selected requirements of Chapter III, plus elements such as sustainability, literacy, inclusive design or the participation of the groups affected.

The Commission issues practical guidelines, among other things on the requirements of Articles 8 to 15, on the Article 5 prohibitions, on Article 50, on the relationship to other legislation, and on the definition of an AI system, with regard to small and medium-sized enterprises. In practice they will often matter more than the text of the regulation itself.

Delegation to the Commission (among other things to amend the annexes) for five years with tacit extension; the EP and the Council may veto an act.

Standard comitology for implementing acts.

Breach of the Article 5 prohibitions up to EUR 35 million or 7% of worldwide turnover. Breach of the other obligations (among them Articles 16, 25, 26, 50) up to EUR 15 million or 3%. False information to the authorities up to EUR 7.5 million or 1%. For small and medium-sized enterprises the lower of the two figures always applies, otherwise the higher. The States add their own enforcement regime.

For Union institutions and bodies, fines are imposed by the European Data Protection Supervisor, up to EUR 1.5 million for prohibited practices and up to EUR 750,000 for other breaches.

Imposed by the Commission, up to EUR 15 million or 3% of turnover, among other things for breaching Chapter V or failing to comply with measures under Article 93.

An amendment to Regulation (EC) No 300/2008 on the protection of civil aviation. Implementing rules on AI-enabled security equipment are to take account of the requirements of Chapter III, Section 2.

An amendment to Regulation (EU) No 167/2013, the approval of agricultural and forestry vehicles, on the same logic of taking the requirements into account.

An amendment to Regulation (EU) No 168/2013, two- and three-wheel vehicles and quadricycles.

An amendment to Directive 2014/90/EU on marine equipment.

An amendment to Directive (EU) 2016/797 on the interoperability of the rail system.

An amendment to Regulation (EU) 2018/858 on the approval of motor vehicles.

Amendments to Regulation (EU) 2018/1139 on civil aviation (the EASA framework).

An amendment to Regulation (EU) 2019/2144 on vehicle safety requirements.

A breach of the AI Act is added to the annex of Directive (EU) 2020/1828, so consumer organisations may bring representative actions for breaches of this regulation too. A key and overlooked track for private enforcement.

A high-risk system placed on the market before the obligations took effect is brought into line only on a significant change to its design. Public authorities always bring theirs into line, at the latest by 2 August 2030. Large-scale EU information systems under Annex X placed before 2 August 2027 by 31 December 2030. General-purpose models placed before 2 August 2025 by 2 August 2027. A new paragraph 4, inserted by Regulation (EU) 2026/1744, gives generative systems placed on the market before 2 August 2026 until 2 December 2026 to comply with Article 50(2).

The Commission regularly assesses the need to amend the Annex III list and the prohibitions, and every four years evaluates the whole regulation and reports, including on the effectiveness of the penalties and the energy consumption of general-purpose models.

The staggered phase-in of the obligations; see the timeline on the AI Act card. The third paragraph was amended by Regulation (EU) 2026/1744, Article 1(40): Annex III from 2 December 2027, Annex I from 2 August 2028, the new Article 5 prohibitions from 2 December 2026 and Articles 102 to 110 from 27 July 2026.