EU AI Act: who's covered, what it requires

Who does EU AI Act apply to?

Out of scope only if: The system is not placed on the EU market, its output is not used in the EU, and you are not an EU-established deployer. The AI Act covers systems placed on the EU market or used in the EU, and deployers established in the EU (Art. 2(1)).

Not screened here: the Art. 2 exclusions for military, defence and national-security purposes (2(3)), scientific research and development (2(6)), research and testing before market placement other than real-world testing (2(8)), purely personal non-professional use (2(10)), and free and open-source AI systems that are not high-risk, prohibited or covered by Art. 50 (2(12)).

• What is your role for this AI system?
• Is the system offered to customers in the EU, or put into service in the EU?
• Is the system's output used in the EU?
• Is your organisation established or located in the EU?
• Does it interact directly with people?
• Does it generate or manipulate audio, image, video, or text?
• Was this system already on the EU market before 2 August 2026?
• Can image, audio, or video output realistically resemble real people, objects, places, entities, or events?
• Do you publish AI-generated or AI-manipulated text to inform the public on matters of public interest?
• Does it identify or infer people's emotions or intentions from biometric data?
• Is that emotion inference used in a workplace or education setting?
• Does it categorise people using biometric data?
• Does that categorisation deduce or infer race, political opinions, trade union membership, religious or philosophical beliefs, sex life, or sexual orientation?
• Does that categorisation infer other sensitive or protected characteristics, such as sex, age, disability or ethnic origin?
• Does it identify people at a distance from biometric data, for example by matching faces against a database?
• Is it used, or built to be used, for real-time identification in publicly accessible spaces for law enforcement?
• Does it score people or groups on social behaviour or personal characteristics, leading to unjustified or unrelated detrimental treatment?
• Does it create or expand facial-recognition databases through untargeted scraping of facial images from the internet or CCTV footage?
• Does it use subliminal or manipulative techniques, or exploit vulnerabilities (age, disability, social or economic situation), in a way likely to cause significant harm?
• Does it predict whether a person will commit a crime based solely on profiling or personality traits?
• Can it generate realistic intimate or sexually explicit images, video or audio of identifiable real people without their consent?
• Can it generate child sexual abuse material?
• Is it a safety component in managing or operating critical digital infrastructure, road traffic, or the supply of water, gas, heating or electricity?
• Does it decide who is admitted to, or assigned to, education or vocational training?
• Does it evaluate learning outcomes, including to steer a student's learning?
• Does it assess the level of education a person will receive or be able to access?
• Does it monitor or detect prohibited behaviour of students during tests?
• Is it used to recruit or select people: targeted job ads, screening or ranking applications, or evaluating candidates?
• Does it support decisions on promotion, termination or work terms, allocate tasks based on behaviour or personal traits, or monitor and evaluate workers' performance?
• Is it used by or for public authorities to decide eligibility for essential public benefits or services, including healthcare?
• Does it evaluate the creditworthiness of individuals or set their credit score? (Detecting financial fraud does not count.)
• Is it used for risk assessment and pricing of individuals in life or health insurance?
• Does it classify emergency calls, dispatch or prioritise emergency services, or triage patients in emergency healthcare?
• Is it used by or for law-enforcement authorities to assess victim or offending risk, as a polygraph, to evaluate evidence, or to profile people in criminal investigations?
• Is it used by or for migration, asylum or border authorities to assess risks, examine asylum, visa or residence applications, or identify people?
• Does it help a court or dispute-resolution body research or interpret facts and law, or apply the law to facts?
• Is it intended to influence the outcome of an election or referendum, or how people vote? (Tools that only organise campaigns administratively do not count.)
• Is it used in any of these Annex III areas? (legacy single-area question)
• Does the system profile people, that is, automatically evaluate personal aspects such as work performance, economic situation, health, preferences, behaviour or location?
• Does it only do one of these, without materially influencing the outcome of decisions: (a) a narrow procedural task, (b) improve the result of a completed human activity, (c) detect decision patterns or deviations for human review, or (d) a preparatory task?
• Is it a product, or a safety component of a product, covered by the EU laws in Annex I (such as medical devices, machinery, toys, lifts, radio equipment or vehicles) that needs third-party conformity assessment?
• Is the product a civil aircraft or aviation security equipment, a motor or agricultural vehicle, marine equipment, a rail system, or machinery? (Annex I, Section B)
• Are you a public body, or a private company providing public services?
• Do you provide a general-purpose AI model (e.g. a foundation model offered via API or weights)?
• Is that model released under a free and open-source licence with weights, architecture, and usage information publicly available?
• Does that model meet the systemic-risk threshold?
• Was the model placed on the market before 2 August 2025?

What are the duty tiers under EU AI Act?

Prohibited practice (Art. 5)

Citation: Art. 5 — https://artificialintelligenceact.eu/article/5/

Narrow statutory exceptions exist, for example emotion recognition for medical or safety reasons (Art. 5(1)(f)) or lawful labelling of biometric datasets (Art. 5(1)(g)). Relying on one needs a documented legal opinion. The two prohibitions added by Reg. (EU) 2026/1744, non-consensual intimate imagery and child sexual abuse material (Art. 5(1)(ba)-(bb)), apply from 2 December 2026.

High-risk (Annex III) (Art. 6(2), Annex III)

Citation: Art. 6(2), Annex III — https://artificialintelligenceact.eu/article/6/

The Annex III rules apply from 2 December 2027 (Reg. (EU) 2026/1744). Systems already on the market before then are caught only if their design changes significantly afterwards, but systems intended for use by public authorities must comply by 2 August 2030 (Art. 111(2)). The Art. 6(3) derogation for narrow or preparatory tasks is never available when the system profiles people.

High-risk (Annex I, Section B product) (Art. 2(2), Art. 6(1))

Citation: Art. 2(2), Art. 6(1) — https://artificialintelligenceact.eu/article/2/

For aviation, vehicles, marine equipment, rail and machinery (Annex I, Section B), only Art. 6(1), Art. 60a and Arts. 102-112 of the AI Act apply (Art. 2(2), as amended by Reg. (EU) 2026/1744). The high-risk requirements reach you through your sector's type-approval legislation, not as direct AI Act duties.

High-risk (Annex I) (Art. 6(1), Annex I)

Citation: Art. 6(1), Annex I — https://artificialintelligenceact.eu/article/6/

The Annex I rules apply from 2 August 2028 (Reg. (EU) 2026/1744). Systems already on the market before then are caught only if their design changes significantly afterwards (Art. 111(2)).

Potentially not high-risk (Art. 6(3)) (Art. 6(3)-(4), Art. 49(2))

Citation: Art. 6(3)-(4), Art. 49(2) — https://artificialintelligenceact.eu/article/6/

An Annex III system is not high-risk if it does not materially influence decisions and only performs a narrow procedural task, improves a completed human activity, detects decision patterns for human review, or performs a preparatory task, and it does not profile people. You must document that assessment before launch and register the system in the EU database (Arts. 6(4), 49(2)). A market-surveillance authority can challenge the assessment.

Transparency duties (Art. 50) (Art. 50)

Citation: Art. 50 — https://artificialintelligenceact.eu/article/50/

Art. 50 has applied since 2 August 2026. The only relief: providers of generative systems already on the market before that date have until 2 December 2026 for machine-readable marking (Art. 111(4)).

Minimal risk (Arts. 4, 95)

Citation: Arts. 4, 95 — https://artificialintelligenceact.eu/article/95/

No risk-tier duties apply to the system itself. AI literacy (Art. 4) still applies, and voluntary codes of conduct are encouraged (Art. 95).

What are the obligations and deadlines under EU AI Act?

What should you do about each obligation?

Support the AI literacy of staff operating the system (Art. 4 (as amended by Reg. (EU) 2026/1744))

Take and document measures that support the development of AI literacy for staff and others operating the system on your behalf — training, guidance, or role-specific briefings proportionate to their technical knowledge, experience, and the context of use. Since the Digital Omnibus (in force 27 July 2026) this is an obligation of effort: it does not require guaranteeing any specific literacy level for any individual.

Cease or do not launch the prohibited practice (Art. 5)

The flagged practice is banned outright, with the highest penalty tier. Stop, redesign, or obtain a documented legal opinion that a statutory exception applies before any EU exposure.

Tell people they are interacting with AI (Art. 50(1))

Design the system so users are informed they are interacting with AI at or before first interaction, unless obvious to a reasonably informed person. Law-enforcement carve-outs may apply.

Mark AI-generated content machine-readably (Art. 50(2))

Embed machine-readable markers in synthetic audio, image, video, and text. A visible label alone does not satisfy this duty. Exempt to the extent the system only performs an assistive standard-editing function or does not substantially alter the deployer's input or its meaning.

Disclose deep fakes you deploy (Art. 50(4))

Disclose that image, audio, or video content resembling real people, objects, places, entities, or events — where it could falsely appear authentic — has been artificially generated or manipulated. For evidently artistic, satirical, or fictional works, disclosure may be made in a manner that does not hamper display or enjoyment.

Disclose AI-generated text published to inform the public (Art. 50(4), second subparagraph)

Disclose that published text informing the public on matters of public interest was artificially generated or manipulated. Exempt where the text underwent human review or editorial control and a natural or legal person holds editorial responsibility for its publication.

Notify people subject to emotion recognition or biometric categorisation (Art. 50(3))

Inform exposed people that the system is operating, and process personal data under applicable EU data-protection law (GDPR, EUDPR, Law Enforcement Directive). Law-enforcement carve-outs may apply.

GPAI model documentation and downstream information (Art. 53(1)(a)-(b))

Maintain model technical documentation and provide information to downstream providers integrating the model. Free and open-source models with public weights are exempt from these two duties unless the model has systemic risk (Art. 53(2)).

Copyright policy and public training-content summary (Art. 53(1)(c)-(d))

Adopt a copyright-compliance policy and publish a sufficiently detailed summary of training content. These duties apply to open-source models too.

Systemic-risk model duties: evaluations, incident reporting, cybersecurity (Arts. 52, 55)

Notify the Commission within two weeks of meeting the systemic-risk threshold (Art. 52(1)); you may argue the model does not present systemic risk (Art. 52(2)). Perform model evaluations and adversarial testing, assess and mitigate systemic risks, report serious incidents, and ensure cybersecurity protection.

High-risk provider obligations (risk management, data governance, technical docs, logging, oversight, accuracy, QMS, conformity assessment, CE marking, registration, post-market monitoring) (Arts. 8-21, 43, 47-49, 72-73)

Build toward the full conformity package now: Arts. 9-15 systems and documentation, Art. 17 quality management system, Arts. 18-20 record-keeping and corrective actions, Art. 43 conformity assessment, Arts. 47-48 declaration of conformity and CE marking, Art. 49 EU database registration, and Arts. 72-73 post-market monitoring and serious-incident reporting.

High-risk deployer duties (use per instructions, human oversight, monitoring, logs, worker notification) (Art. 26)

Assign trained human oversight, use the system per the provider's instructions, ensure input data under your control is relevant and sufficiently representative, monitor operation, retain logs under your control for at least six months, inform workers and their representatives before workplace deployment, and inform affected natural persons where the system makes or helps make decisions about them.

Fundamental rights impact assessment (FRIA) (Art. 27)

Assess the impact on fundamental rights before first use: processes, categories of affected persons, risks, oversight, and mitigation. Triggered for public bodies and private providers of public services, and for any deployer using high-risk AI for creditworthiness or life/health insurance pricing (Annex III 5(b)-(c)); critical-infrastructure systems are excepted.

Importer duties for high-risk AI systems (Art. 23)

Before placing the system on the EU market, verify the provider completed conformity assessment, drew up the technical documentation, and affixed the CE marking with an EU declaration of conformity; add your name and contact details on the product or packaging; ensure storage and transport do not jeopardise compliance; keep documentation for 10 years; and inform the provider and market-surveillance authorities of any risk.

Distributor duties for high-risk AI systems (Art. 24)

Before making the system available, verify it bears the CE marking, is accompanied by the EU declaration of conformity and instructions, and that provider and importer complied with their obligations; ensure storage and transport do not jeopardise compliance; take corrective action or withdraw where non-conformity appears; and inform providers, importers, and authorities of any risk.

Appoint an authorised representative in the EU (Arts. 22, 54)

As a provider established outside the EU, appoint an authorised representative established in the EU by written mandate before making a high-risk system (Art. 22) or a general-purpose AI model (Art. 54) available in the EU. The mandate must let the representative keep documentation available to authorities and cooperate with them. Providers of free and open-source models without systemic risk are exempt from the Art. 54 duty (Art. 54(6)).

Document the Art. 6(3) assessment and register the system (Arts. 6(4), 49(2))

Before placing the system on the market, document why it is not high-risk under Art. 6(3) and give that assessment to national authorities on request (Art. 6(4)). Register yourself and the system in the EU database (Art. 49(2)); Reg. (EU) 2026/1744 kept this registration but reduced the information required in Annex VIII. Re-check the assessment whenever the system's purpose changes.

Follow the AI requirements in your sector's type-approval rules (Art. 2(2), Arts. 102-112)

Chapter III duties do not apply to you directly. Track the delegated and implementing acts adopted under your sector's legislation (civil aviation, vehicle type approval, marine equipment, rail, machinery), which must take the AI Act's high-risk requirements into account, and build them into your existing type-approval or conformity process.

Disclaimer

Compliance guidance, not legal advice. Confirm obligations with counsel. Content version 2026.09.16-1, verified 2026-09-16.

Frequently asked questions