Scope: study of the original 2024 regulation. The timeline and examples below are retained study material. This visual refresh does not validate their current legal applicability; check the official text, subsequent amendments and Commission implementation guidance.
Regulation (EU) 2024/1689 establishes the European Union's harmonized, risk-based legal framework governing the development, placing on the market, putting into service, and use of artificial intelligence systems. This study guide distills your AI EU ACT notebook into an actionable engineering and legal reference centered on High-Risk AI obligations (Articles 6–27 & 43–49).
Administrative Fine Tiers (Article 99)
Applies to non-compliance with Prohibited AI Practices under Article 5 (e.g., social scoring, untargeted facial scraping, workplace emotion recognition).
Applies to violations of High-Risk AI obligations (Articles 8–27), Authorized Representative/Importer/Distributor rules, Transparency (Art. 50), and GPAI provider duties (Arts. 51–54).
Applies to supplying incorrect, incomplete, or misleading information to Notified Bodies or National Competent Authorities in response to a request.
Four-Tier Risk Hierarchy & Article 6(3) Checker
The EU AI Act calibrates regulatory intensity proportionally to the severity and probability of harm to health, safety, and fundamental rights. Select a risk tier in the pyramid below to inspect statutory triggers and obligations:
Unacceptable Risk — Prohibited AI Practices
AI systems considered a clear threat to Union values and fundamental rights are banned outright across the EU:
- Social Scoring: Evaluating or classifying individuals over time based on social behavior or personal traits leading to detrimental or disproportionate treatment.
- Untargeted Facial Scraping: Untargeted scraping of facial images from the internet or CCTV footage to create or expand facial recognition databases.
- Workplace & Education Emotion Recognition: Inferring emotions in workplaces or educational institutions (except strictly for medical or safety reasons, such as pilot drowsiness detection).
- Subliminal Manipulation & Vulnerability Exploitation: Deceptive or subliminal techniques distorting behavior to cause significant harm, or exploiting vulnerabilities due to age, disability, or socio-economic status.
- Predictive Policing via Profiling: Assessing the risk of a natural person committing a criminal offense based solely on profiling or personality traits.
- Sensitive Biometric Categorization: Inferring race, political opinions, trade union membership, religious/philosophical beliefs, sex life, or sexual orientation.
- Real-Time Remote Biometric Identification (RBI) in Public Spaces: Prohibited for law enforcement except in narrowly defined, judicially authorized scenarios (targeted search for kidnapping/trafficking victims, imminent terrorist threat, or locating suspects of serious enumerated crimes).
High-Risk AI Systems — Strict Pre-Market & Lifecycle Controls
An AI system is classified as High-Risk under two distinct pathways in Article 6:
- Pathway A — Annex I Safety Components (Art. 6(1)): The AI system is intended to be used as a safety component of a product (or is itself a product) covered by EU harmonization legislation listed in Annex I (e.g., medical devices, in-vitro diagnostics, machinery, aviation, motor vehicles, toys, lifts, radio equipment) and is subject to third-party conformity assessment.
- Pathway B — Annex III Standalone High-Risk Domains (Art. 6(2)): AI systems deployed in 8 sensitive socio-economic and public-interest areas:
Limited Risk — Transparency Obligations
Systems that pose specific risks of impersonation, deception, or covert interaction must satisfy targeted transparency rules under Article 50:
- Human-AI Interaction Disclosure (Art. 50(1)): Providers of conversational agents/chatbots must design the system so natural persons are informed they are interacting with an AI system, unless obvious from the circumstances and context of use.
- Synthetic Content Watermarking (Art. 50(2)): Providers of generative AI systems creating synthetic audio, image, video, or text content must ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated.
- Emotion Recognition & Biometric Categorization (Art. 50(3)): Deployers must inform natural persons exposed to the system of its operation and process personal data in compliance with GDPR.
- Deepfakes & Public-Interest Text (Art. 50(4)): Deployers generating or manipulating image, audio, or video content constituting a "deepfake" (or AI-generated text published to inform the public on matters of public interest) must disclose that the content has been artificially generated or manipulated.
Minimal or No Risk — Voluntary Codes of Conduct
The vast majority of everyday AI applications fall into this category and can be developed and deployed without mandatory EU AI Act requirements:
- Representative Examples: AI-enabled email spam filters, video game NPC behavior, inventory and supply-chain forecasting, basic search/recommendation engines, and spell-checkers.
- Voluntary Governance (Art. 95): The AI Office and Member States encourage providers and deployers of non-high-risk AI systems to voluntarily adopt Codes of Conduct applying elements of the High-Risk governance framework (such as environmental sustainability, accessibility, and stakeholder diversity).
Article 6(3) Annex III Exemption / Derogation Self-Check
Under Article 6(3), an AI system listed in Annex III is not considered High-Risk if it does not pose a significant risk of harm to health, safety, or fundamental rights (including not materially influencing the outcome of decision-making) AND satisfies one or more of the four narrow conditions below:
Interactive Compliance Timeline (2024–2028)
Regulation (EU) 2024/1689 enters into application in staggered waves following its Official Journal publication on July 12, 2024. Click any milestone below to review its enforceable obligations and compliance checklist:
August 1, 2024 — Entry into Force (Day 0)
Article 113The EU AI Act officially entered into force 20 days after publication in the Official Journal of the European Union, triggering the statutory 6-, 12-, 24-, and 36-month compliance countdowns.
- Establish internal AI inventory and map all systems against Article 5 bans and Annex I / Annex III High-Risk domains.
- Assign internal governance ownership across Legal, Security, Data Science, and Product teams.
February 2, 2025 — Chapters I & II Apply (6 Months)
Active & EnforceableGeneral provisions, definitions, Prohibited AI Practices (Article 5), and the AI Literacy mandate (Article 4) become legally binding.
- Decommission or block any AI use case falling under Article 5 prohibitions (social scoring, workplace emotion recognition, untargeted facial scraping).
- Implement role-based AI Literacy programs (Art. 4) ensuring staff and operators understand AI capabilities, limitations, risks, and oversight duties.
August 2, 2025 — GPAI, Governance & Penalties Apply (12 Months)
Active & EnforceableChapter V rules for General-Purpose AI (GPAI) models (Articles 51–56), Notifying Authorities, the European AI Board, confidentiality rules, and Article 99 administrative penalties take effect.
- GPAI providers must publish training content summaries, enforce EU copyright compliance policies, and supply Annex XI/XII documentation to downstream integrators.
- Providers of GPAI models with systemic risk (> 1025 FLOPs) must perform adversarial evaluations and report serious incidents to the European AI Office.
August 2, 2026 — Full Applicability & Annex III High-Risk (24 Months)
Core High-Risk DateThe core operational regime of the EU AI Act applies in full, including all requirements for Annex III standalone High-Risk AI systems (Articles 8–27 & 43–49), Article 50 transparency duties, and national AI regulatory sandboxes.
- Complete Articles 9–15 engineering controls, Annex IV technical documentation, Conformity Assessment, CE marking, and EU Database registration before placing Annex III systems on the market.
- Deployers in public services, credit scoring, and life/health insurance must complete Article 27 Fundamental Rights Impact Assessments (FRIAs) prior to deployment.
August 2, 2027 (and Transition Window to Aug 2028) — Annex I Safety Components
36 Months + Harmonized Standards BufferHigh-Risk AI systems embedded as safety components in products regulated under Annex I EU harmonization legislation (Article 6(1)) become subject to full High-Risk obligations, alongside Digital Omnibus / harmonized standards transition windows extending up to August 2028 for legacy systems placed on the market before August 2026 unless they undergo significant modifications.
- Coordinate EU AI Act conformity assessments with sectoral Notified Bodies (e.g., Medical Device Regulation MDR, Machinery Regulation, Aviation, and Automotive type-approval).
- Monitor CEN-CENELEC harmonized standards adoption and legacy system "significant modification" triggers through August 2028.
High-Risk Technical Obligations (Articles 8–15)
Under Article 8, High-Risk AI systems must comply with the mandatory requirements laid out in Articles 9 through 15, taking into account their intended purpose and the generally acknowledged state of the art:
Risk Management System
Establish, implement, document, and maintain a continuous, iterative lifecycle process that identifies and analyzes known and reasonably foreseeable risks to health, safety, and fundamental rights, estimates/evaluates risks emerging during intended use and reasonably foreseeable misuse, and adopts targeted risk-management measures so residual risks are judged acceptable.
Data & Data Governance
Training, validation, and testing datasets must be subject to rigorous governance practices covering design choices, data collection origin, preparation operations (annotation, labeling, cleaning), assumption formulation, and examination for possible biases likely to affect health, safety, or fundamental rights. Datasets must be relevant, sufficiently representative, and to the best extent possible, free of errors and complete.
Technical Documentation
Draw up comprehensive technical documentation before the system is placed on the market or put into service, and keep it continuously updated. Must satisfy Annex IV: general system description, architecture and computational resources, training methodologies and data provenance, validation/testing metrics, risk management outcomes, and cybersecurity controls.
Record-Keeping & Automatic Logging
Technically enable the automatic recording of events ("logs") over the lifetime of the system to ensure traceability of operation. Logging capabilities must record the period of each use (start/end date and time), the reference database against which input data was checked, input data that led to a match, and the identification of the natural persons involved in verification under Article 14(5).
Transparency & Instructions for Use
Ensure operation is sufficiently transparent to enable deployers to interpret the system's output and use it appropriately. Accompany the system with clear, concise Instructions for Use detailing provider identity, intended purpose, accuracy/robustness/cybersecurity levels, known limitations or circumstances leading to risks, input data specifications, and human oversight measures.
Human Oversight
Design systems with appropriate human-machine interface tools so they can be effectively overseen by natural persons during use. Controls must prevent automation bias (over-reliance on AI outputs) and empower operators to understand capacities/limitations, correctly interpret outputs, decide not to use the system, disregard/override/reverse outputs, or interrupt operation via a safe "stop" button.
Accuracy, Robustness & Cybersecurity
High-Risk AI systems must achieve an appropriate level of accuracy, robustness, and cybersecurity consistently throughout their lifecycle. Accuracy metrics must be declared in the Instructions for Use. Systems must be resilient against errors, faults, or inconsistencies (via technical redundancy/fail-safe plans) and protected against unauthorized third-party attempts to alter their use, outputs, or performance—specifically guarding against data poisoning, model poisoning, adversarial examples, confidentiality attacks, and model flaws.
{
"regulation": "Regulation (EU) 2024/1689 (EU AI Act)",
"manifest_version": "1.0.0",
"system_metadata": {
"system_id": "hr-talent-screener-eu-v2",
"provider": "NordicAI Solutions AB",
"intended_purpose": "Automated candidate CV ranking and interview shortlisting for enterprise recruitment",
"lifecycle_status": "pre_market_conformity_assessment"
},
"classification": {
"risk_tier": "HIGH_RISK",
"legal_basis": "Article 6(2) & Annex III, Point 4(a) (Employment & Recruitment)",
"art6_3_derogation_claimed": false,
"performs_natural_person_profiling": true
},
"art10_data_governance": {
"dataset_lineage_id": "ds-hr-eu-2026-q3",
"representativeness_verified": true,
"bias_audit_metrics": {
"demographic_parity_difference": 0.028,
"equalized_odds_ratio": 0.96,
"special_category_data_art10_5_safeguards": "strict_access_control_and_deletion_post_mitigation"
}
},
"art12_record_keeping": {
"automatic_event_logging": true,
"logged_fields": ["session_start_utc", "session_end_utc", "model_version_hash", "input_record_id", "verifier_operator_id"],
"minimum_retention_months": 12
},
"art14_human_oversight": {
"oversight_mode": "HUMAN_IN_THE_LOOP",
"automation_bias_safeguards": ["confidence_interval_display", "counterfactual_feature_attribution", "mandatory_recruiter_signoff"],
"override_and_stop_mechanism": "immediate_ui_override_and_api_circuit_breaker"
},
"art15_robustness_and_cybersecurity": {
"declared_accuracy_benchmark": { "f1_score": 0.942, "false_negative_rate_max": 0.035 },
"adversarial_hardening": ["prompt_injection_sanitization", "input_schema_validation", "model_weights_cryptographic_signing"],
"fail_safe_fallback": "route_to_manual_hr_review_queue"
}
}
Provider vs. Deployer & GPAI Matrix
The EU AI Act allocates distinct compliance responsibilities across the AI value chain—primarily between Providers (who develop or place the AI system on the market under their own name/trademark) and Deployers (who use the AI system under their authority in a professional capacity):
| Compliance Dimension | Providers (Arts. 16–22, 43–49) | Deployers (Arts. 26–27) |
|---|---|---|
| Core Governance & QMS | Establish a comprehensive Quality Management System (Art. 17) covering regulatory compliance strategy, design/testing controls, data management, and post-market monitoring. | Operate the system strictly in accordance with the provider's Instructions for Use (Art. 26(1)) and implement organizational controls. |
| Conformity Assessment (Art. 43) | Conduct pre-market conformity assessment: Internal Control (Annex VI) for most Annex III systems, or Notified Body assessment (Annex VII) for biometric identification/categorization if harmonized standards are absent. | Verify prior to deployment that the High-Risk AI system bears valid CE marking and is accompanied by Instructions for Use. |
| CE Marking & EU Database (Arts. 47–49) | Draw up an EU Declaration of Conformity (Art. 47), affix visible/digital CE Marking (Art. 48), and register the system in the EU Database (Art. 49) before market placement. | Public authorities and EU institutions deploying Annex III systems must also register their use in the EU Database (Art. 49(3)). |
| Human Oversight & Input Data | Build technical human-oversight interfaces and capabilities into the system architecture (Art. 14). | Assign competent, trained, and empowered natural persons to execute human oversight (Art. 26(2)) and ensure input data under deployer control is relevant and sufficiently representative (Art. 26(4)). |
| Logging & Incident Reporting | Design automatic event logging (Art. 12) and report serious incidents or widespread infringements to market surveillance authorities (Art. 73). | Monitor operation, report serious incidents/risks to the provider and authorities (Art. 26(5)), and retain automatically generated logs for at least 6 months (Art. 26(6)). |
| Transparency & Rights Notices | Provide clear Article 13 Instructions for Use and technical transparency disclosures. | Inform workers and their representatives before deploying high-risk AI in the workplace (Art. 26(7)), inform affected individuals subject to Annex III decisions (Art. 26(11)), and provide meaningful explanations upon request (Art. 86). |
| Fundamental Rights Impact Assessment (Art. 27) | Assess foreseeable fundamental rights risks within the Article 9 Risk Management System. | Mandatory FRIA (Art. 27) prior to first use for deployers that are bodies governed by public law, private entities providing public services, and deployers of Annex III systems for credit scoring (5b) or life/health insurance pricing (5c). |
General-Purpose AI (GPAI) & Systemic Risk Models
GPAI models (including large foundation models capable of performing a wide range of distinct tasks) are governed under a two-tier framework effective August 2, 2025:
- All GPAI Model Providers (Art. 53): Must maintain up-to-date technical documentation (Annex XI), provide technical integration documentation to downstream AI system providers (Annex XII), establish a policy to comply with EU Copyright Law (respecting text-and-data-mining opt-outs), and publish a sufficiently detailed public summary of training content. (Open-source GPAI models without systemic risk receive targeted documentation exemptions under Art. 53(2).)
- GPAI Models with Systemic Risk (Arts. 51 & 55): A GPAI model is presumed to pose systemic risk if the cumulative amount of compute used for its training exceeds 1025 floating-point operations (FLOPs), or by European Commission designation. Providers must additionally conduct standardized model evaluations, perform adversarial red-teaming, track and report serious incidents to the European AI Office, and ensure an adequate level of cybersecurity protection.
High-Risk Study Q&As + Custom Note Builder
5-Question High-Risk Compliance Self-Check Quiz
A fintech company builds an Annex III AI model to screen personal loan applicants. The model performs automated behavioral profiling to estimate default risk, though a human loan officer reviews the score before final approval. Can the provider claim an Article 6(3) "preparatory task" exemption to avoid High-Risk obligations?
During a pre-market audit of a High-Risk diagnostic triage AI system, auditors notice clinicians routinely accept the model’s top recommendation within 2 seconds without inspecting confidence intervals or alternative differentials. Which EU AI Act requirement is directly violated?
A private commercial bank plans to deploy a third-party High-Risk AI system (CE-marked and registered in the EU Database) to evaluate consumer creditworthiness under Annex III, Point 5(b). What mandatory assessment must the bank complete before first use?
To detect and mitigate demographic bias in an Annex III employment screening model, a provider needs to process protected ethnic origin and disability attributes from validation datasets. How does Article 10(5) regulate this?
An AI lab trains a new General-Purpose AI (GPAI) foundation model using a cumulative training compute of 3.2 × 10^25 floating-point operations (FLOPs). What legal presumption applies under Article 51(2)?
References and source links
- European Union (2024). Regulation (EU) 2024/1689, original published text. Check the current legal text and amendments before applying this study material.https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- European Commission. AI Act policy and implementation information.https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
Reference list added during the visual refresh on 14 Nov 2026. It is a reading path, not a completed claim-by-claim audit of this imported notebook.