Requirements & Obligations

Ensure staff and persons dealing with AI systems have sufficient AI literacy, considering technical knowledge, experience, and context of use

AI systems using subliminal techniques beyond consciousness or manipulative/deceptive techniques to materially distort behaviour causing significant harm

AI systems exploiting vulnerabilities due to age, disability, or social/economic situation to materially distort behaviour causing significant harm

AI systems evaluating or classifying persons based on social behaviour or personal characteristics, leading to detrimental or disproportionate treatment

AI systems assessing criminal risk based solely on profiling or personality traits (exceptions for human-assisted assessments based on objective facts)

AI systems creating or expanding facial recognition databases through untargeted scraping from the internet or CCTV footage

AI systems inferring emotions in workplace and education institutions (exceptions for medical or safety purposes)

Biometric categorisation systems deducing race, political opinions, trade union membership, religious beliefs, sex life, or sexual orientation

Real-time remote biometric identification in public spaces for law enforcement (narrow exceptions for victim search, imminent threats, serious crime suspects)

Establish, implement, document, and maintain a continuous iterative risk management system throughout the entire AI lifecycle

Training, validation, and testing datasets must meet quality criteria, be representative, free of errors, and subject to appropriate governance practices

Draw up comprehensive technical documentation per Annex IV before placing on market, demonstrating compliance with all requirements

Enable automatic recording of events (logs) over the system's lifetime for risk identification, post-market monitoring, and operation auditing

Design for sufficient operational transparency; provide instructions for use with performance characteristics, limitations, accuracy metrics, and oversight measures

Design for effective human oversight: ability to understand, detect anomalies, interpret output, override decisions, and intervene or stop the system

Achieve appropriate levels of accuracy, robustness against errors and attacks (data poisoning, adversarial examples, model evasion), and cybersecurity

Implement a documented QMS covering compliance strategy, design procedures, testing, data management, risk management, post-market monitoring, and accountability

Undergo conformity assessment before placing on market — internal control (Annex VI) or notified body assessment (Annex VII) for biometric systems

Draw up written EU declaration of conformity, affix CE marking on the system or documentation, and keep documentation for 10 years

Register the provider and each high-risk AI system in the EU database before placing on market or putting into service

Establish a proportionate post-market monitoring system to actively collect and analyse performance data throughout the system's lifetime

Report serious incidents to market surveillance authorities — within 15 days generally, 2 days for widespread infringements, 10 days for deaths

Take corrective actions for non-conforming systems (withdraw, disable, recall), inform all downstream operators, and cooperate with competent authorities

Keep technical documentation, QMS records, and conformity certificates for 10 years; retain automatically generated logs for at least 6 months

Use systems in accordance with instructions for use; ensure input data is relevant and sufficiently representative for the intended purpose

Assign human oversight to persons with necessary competence, training, authority, and support to effectively oversee the AI system

Monitor operation based on instructions; inform providers of risks; suspend use if risk identified; report serious incidents to authorities

Keep automatically generated logs for at least 6 months (or longer if required by law), to the extent logs are under deployer's control

Inform workers' representatives and affected workers before workplace deployment; inform natural persons that they are subject to AI system decisions

Public bodies and certain private deployers must assess impact on fundamental rights before deployment, covering processes, affected persons, risks, and oversight

Affected persons subject to AI-based decisions with legal effects have the right to clear and meaningful explanations of the AI system's role in the decision

Verify conformity assessment, technical documentation, CE marking, EU declaration, and authorised representative appointment before placing on market

Verify CE marking, EU declaration, instructions for use, and that provider and importer have met their obligations before making available on market

Any actor putting their name on a high-risk system, making substantial modifications, or changing its intended purpose becomes the provider with full obligations

AI systems interacting directly with persons must be designed so users are informed they are interacting with an AI system (unless obvious from context)

AI systems generating synthetic audio, image, video, or text must mark outputs in machine-readable format as artificially generated or manipulated

Deployers of emotion recognition or biometric categorisation systems must inform exposed persons and process data in accordance with GDPR

Deployers must disclose when content (image, audio, video, text on public interest matters) has been artificially generated or manipulated

Draw up and maintain technical documentation of the model including training/testing process and evaluation results per Annex XI

Provide documentation enabling downstream AI system providers to understand model capabilities and limitations per Annex XII

Implement copyright compliance policy respecting rights reservations; publish sufficiently detailed summary of training content per AI Office template

Perform model evaluation using standardised protocols, conduct adversarial testing, and assess/mitigate possible systemic risks at Union level

Track, document and report serious incidents to the AI Office without undue delay; ensure adequate cybersecurity for model and physical infrastructure