Regulation (EU) 2024/1689 [EU AI Act] establishes a uniform, risk-based legal framework to ensure trustworthy AI throughout the European Union while safeguarding fundamental rights.
The AI Act strictly prohibits unacceptable practices, such as subliminal manipulation, social scoring, untargeted scraping of facial images, and workplace emotion recognition.
High-risk systems, such as those used in critical infrastructure, employment, and essential services, must meet stringent requirements for quality management, logging, human oversight, and cybersecurity.
Baseline transparency and copyright rules apply to general-purpose AI models, whereas high-impact models face additional systemic risk obligations.
The Digital Omnibus on AI streamlines implementation by setting revised application dates for high-risk systems in December 2027 and August 2028, and by imposing penalties of up to €35 million or 7% of global turnover for severe violations.
EU AI Act Interactive Explorer
The two referenced sources contain a combined total of 200+ pages. Yes, it takes a considerable amount of coffee to digest them. I’ve made it easy for you to grasp the big picture with this interactive Explorer.
Of course, should you need to understand the intricate details, you’ll have to consult the sources themselves.
DISCLAIMER: The Explorer was created with AI support. I’ve done my best to quality-assure the Explorer, but AI can make mistakes. This is not legal or technical advice.
EU AI Act Explorer
Comprehensive Guide to Regulation (EU) 2024/1689
Sources:
- Regulation (EU) 2024/1689 (EU AI Act - Official Journal)
- Digital Omnibus on AI (COM(2025) 836 / Trilogue Agreement May 2026)
Scope & Risk Approach
The AI Act regulates the placing on the market, putting into service, and use of AI systems in the EU. It applies universally, protecting the fundamental rights enshrined in the EU Charter, democracy, the rule of law, and the environment. It does not apply to AI used exclusively for military, defence, national security, or sole scientific R&D purposes.
The framework follows a strict, proportionate risk-based approach, classifying systems into four categories. (Source: AI Act Recitals 1-12)
The AI Act Risk Pyramid
Risk
Hover or click a tier to explore
Select a risk tier from the pyramid
to view its detailed definitions.
Prohibited AI Practices
Article 5 details systems posing unacceptable risks to fundamental rights. These are banned from being placed on the market or put into service in the EU. Click each section to expand details.
1. Cognitive & Subliminal Manipulation
Deploying AI systems that use subliminal, manipulative, or deceptive techniques to distort human behaviour in a way that impairs informed decision-making, causing significant harm.
- Exploitation of Vulnerabilities: This includes exploiting individuals or groups due to their age, disability, or specific social/economic situations.
- Harm Requirement: The prohibition applies when the manipulation causes, or is reasonably likely to cause, significant physical, psychological, or financial harm.
- Medical Exemption: Lawful practices in the context of medical treatment (e.g., psychological therapy or physical rehabilitation) with explicit consent are not considered prohibited manipulation, as clarified in Recital 29.
2. Social Scoring
AI systems used by public or private actors for the evaluation or classification of natural persons based on their social behaviour over time or inferred personality traits.
- Prohibited Outcomes: The scoring is banned if it leads to detrimental or unfavourable treatment in social contexts that are unrelated to where the data was originally generated, or if the treatment is unjustified or disproportionate.
- Context: This prevents mass surveillance and citizen scoring systems that evaluate individuals' general "trustworthiness" or societal worth.
3. Real-Time Remote Biometrics (Law Enforcement)
The use of 'real-time' remote biometric identification systems in publicly accessible spaces for the purpose of law enforcement is heavily prohibited.
- Targeted search for victims of abduction, trafficking, or missing persons.
- Prevention of a specific, substantial, and imminent threat to life, physical safety, or a terrorist attack.
- Localisation/identification of a suspect for specific serious criminal offences (e.g., murder, kidnapping, terrorism) punishable by a custodial sentence of at least four years.
4. Untargeted Scraping of Facial Images
The use of AI systems that create or expand facial recognition databases through the untargeted scraping of facial images from the internet or from CCTV footage.
- Privacy Protection: This ban is designed to prevent mass surveillance capabilities and protect the fundamental right to privacy from tools like Clearview AI.
5. Emotion Recognition at Work & School
Using AI to infer the emotions of natural persons in the areas of the workplace and educational institutions is strictly prohibited due to the power imbalance in these environments.
- Unscientific Basis: The regulation acknowledges that emotion recognition is often unscientific, unreliable, and prone to cultural bias.
- Safety Exception: It is not prohibited if the system is intended to be put in place strictly for medical or safety reasons (e.g., systems detecting state of fatigue in professional pilots or truck drivers to prevent accidents).
6. Biometric Categorisation (Sensitive Traits)
Categorising individuals based on biometric data to deduce or infer sensitive personal attributes.
- Prohibited Traits: Inferring race, political opinions, trade union membership, religious or philosophical beliefs, sex life, or sexual orientation.
- Law Enforcement Exemption: This prohibition does not cover the lawful labelling or filtering of lawfully acquired biometric datasets by law enforcement (e.g., sorting suspects by hair or eye color from acquired footage).
7. CSAM & Illegal Sexual Content
Generating or disseminating Child Sexual Abuse Material (CSAM) and non-consensual illegal sexual content (deepfakes) via AI systems.
- Strict Enforcement: Introduced via the Digital Omnibus, this dedicated ban takes effect in December 2026.
- Severe Penalties: Violations fall under the Article 5 penalty structure, carrying the highest possible fines (up to €35M or 7% of global turnover).
High-Risk AI (Articles 6-49)
Systems that negatively affect safety or fundamental rights must comply with strict requirements.
- Maintain Assessment Evidence for Art. 6(3) Derogations: While the Digital Omnibus removed the mandatory EU Database registration for systems deemed non-high-risk via narrow procedural derogations, technical documentation justifying this classification must still be retained internally for market surveillance authorities.
- Track Technical Standards: Because the enforcement of High-Risk rules is directly tied to the publication of CEN/CENELEC harmonized standards, compliance officers should monitor the EU Commission's formal decisions confirming standard availability.
- Establish Data Pseudonymisation Logs: When leveraging Article 4a for sensitive data bias testing, maintain explicit processing logs detailing why synthetic data was insufficient, as required by GDPR Art. 30 / AI Act Art. 4a(1)(f).
If a deployer modifies a high-risk system in a way that affects its compliance with the AI Act, or modifies the intended purpose of a system such that it becomes high-risk, this constitutes a "substantial modification".
The Consequence: The deployer legally assumes the role of the Provider, becomes subject to all 7 mandatory requirements below, and the system must undergo a new conformity assessment.
7 Mandatory Requirements for Providers
1. Risk Management System
A continuous, iterative process run throughout the entire lifecycle of a high-risk AI system, requiring regular systematic review and updating.
- Identify & Evaluate: Providers must identify known and reasonably foreseeable risks to health, safety, or fundamental rights when the system is used as intended or under conditions of reasonably foreseeable misuse.
- Mitigation Measures: Providers must adopt targeted measures to eliminate or reduce risks as far as technically feasible. If risks cannot be eliminated, adequate mitigation and control measures must be implemented.
- Pre-Market Testing: High-risk systems must be tested against prior defined metrics to ensure they perform consistently and comply with all requirements before being placed on the market.
2. Data and Data Governance
Models trained with data must be developed using training, validation, and testing datasets that meet strict quality criteria.
- Quality Standards: Datasets must be relevant, sufficiently representative, and to the best extent possible, free of errors and complete.
- Bias Mitigation: Data governance practices must include examination for possible biases that could affect health, safety, or lead to discrimination prohibited by EU law.
- Sensitive Data Exception: Source: Digital Omnibus on AI: Article 4a was introduced to explicitly permit the processing of sensitive personal data (e.g., race, health) strictly for bias detection and correction across all AI systems, subject to state-of-the-art security (like pseudonymisation) and strict access controls.
3. Technical Documentation
Must be drawn up before the system is placed on the market to demonstrate compliance and provide authorities with the necessary information to assess the system.
- Contents: Must include system architecture, algorithm logic, training methodologies, metrics used for accuracy/robustness, and the risk management system description.
- SME & SMC Provisions: Source: Digital Omnibus on AI: Simplified documentation provisions and templates originally meant for SMEs and start-ups have been extended to Small Mid-Caps (SMCs).
4. Record-Keeping (Logs)
Systems must technically allow for the automatic recording of events ('logs') over their lifetime.
- Traceability: Logging must be appropriate to the system's purpose and is crucial for identifying situations that present a risk, facilitating post-market monitoring, and investigating incidents.
- Biometric Specifics: For remote biometric identification, logs must record the period of use, the reference database checked against, the input data that led to a match, and the identity of the personnel verifying the result.
5. Transparency & Instructions
Operation must be sufficiently transparent to enable deployers to understand how the system works and use it appropriately.
- Instructions for Use: Must be concise, correct, clear, and accessible.
- Required Disclosures: Instructions must detail the intended purpose, expected accuracy/robustness, known limitations, and circumstances where the system might pose risks to health, safety, or fundamental rights.
6. Human Oversight
Systems must be designed with appropriate human-machine interfaces allowing natural persons to effectively oversee their functioning and intervene if needed.
- Intervention: Overseers must be able to understand the system's capacities, remain aware of automation bias, and be able to disregard output or interrupt the system safely (a "stop" button).
- The "Four-Eyes" Principle: For high-risk biometric identification systems, no action or decision can be taken by the deployer based on an AI match unless that match has been separately verified and confirmed by at least two natural persons (except in certain law enforcement situations).
7. Accuracy, Robustness & Cybersecurity
Systems must achieve an appropriate level of accuracy, robustness, and cybersecurity and perform consistently throughout their lifecycle.
- Feedback Loops: Systems that continue to learn after deployment must eliminate or reduce the risk of biased outputs influencing future operations.
- Cyber Resilience: Must be resilient against attempts by unauthorised third parties to alter their use, outputs, or performance.
- AI-Specific Threats: Must include measures to prevent and control attacks trying to manipulate the training dataset (data poisoning), adversarial examples, or model evasion.
General-Purpose AI (Ch. V)
Rules for foundational models capable of competently performing a wide range of tasks.
All GPAI Models
Obligations apply universally to ensure downstream providers have necessary information. (Articles 53-54 • Recitals 97, 101, 105)
- Tech Docs & Transparency: Keep up-to-date documentation; provide info to downstream providers integrating the model.
- Copyright: Put in place a policy to respect EU copyright laws, honoring rights reservations (opt-outs).
- Training Data: Publish a sufficiently detailed summary of the content used for training.
High-Impact GPAI
Models trained with computation > 10^25 FLOPs, presumed to have high-impact capabilities causing systemic risks to health, security, or rights. (Articles 51, 55 • Recitals 110, 111)
Added Obligations:
- Perform model evaluations & adversarial testing (Red Teaming).
- Assess and mitigate systemic risks at the Union level.
- Keep track of and report serious incidents without undue delay.
- Ensure adequate cybersecurity for the model and physical infrastructure.
Transparency Rules (Article 50)
Comprehensive obligations to prevent impersonation, manipulation, and deception by AI systems. Click each section to expand details.
1. Direct Interaction with AI (Chatbots)
Providers must design and develop AI systems intended to interact directly with natural persons in such a way that the individuals are notified that they are interacting with an AI system.
- Exception: Notification is not required if it is obvious from the point of view of a reasonably well-informed, observant, and circumspect person, taking the context of use into account.
- Vulnerable Groups: The notification design must appropriately consider the characteristics of vulnerable groups (e.g., age, disability) if the system is specifically intended to interact with them.
- Law Enforcement Exemption: Does not apply to AI systems authorised by law to detect, prevent, investigate, or prosecute criminal offences (unless the system is available for the general public to report crimes).
2. Machine-Readable Marking (Generative AI)
Providers of AI systems that generate synthetic audio, image, video, or text content must ensure the outputs of the AI system are marked in a machine-readable format and are detectable as artificially generated or manipulated.
Source: Digital Omnibus on AI: Introduced a transitional period delaying enforcement of this specific obligation until December 2026 for systems already on the market.
- Technical Requirements: Solutions must be effective, interoperable, robust, and reliable (e.g., using watermarks, cryptographic provenance, or metadata), reflecting the state of the art.
- Exceptions: This marking is not required if the AI only performs an assistive function for standard editing, or if it does not substantially alter the original input data or its semantics.
3. Deepfakes Disclosure
Deployers of an AI system that generates or manipulates image, audio, or video content constituting a "deep fake" (content resembling existing persons, objects, places, or events that would falsely appear to be authentic) must clearly and distinguishably disclose that the content has been artificially generated or manipulated.
- Artistic & Satirical Exception: If the content is evidently part of a creative, satirical, or artistic work, the disclosure is still required, but it must be provided in an appropriate manner that does not hamper the display or enjoyment of the work.
- Law Enforcement Exemption: Does not apply where use is authorised by law to detect, prevent, investigate, or prosecute criminal offences.
4. AI-Generated Text of Public Interest
Deployers of an AI system that generates or manipulates text published with the purpose of informing the public on matters of public interest must disclose that the text has been artificially generated or manipulated.
- Human Review Exception: Disclosure is not required if the AI-generated text has undergone a process of human review or editorial control, and a natural or legal person holds editorial responsibility for its publication.
5. Emotion & Biometric Categorisation
Deployers of an emotion recognition system or a biometric categorisation system must inform the natural persons exposed thereto of the operation of the system.
- Data Protection: Any personal data must be processed in strict accordance with the GDPR (Regulation (EU) 2016/679) and other relevant data protection laws.
- Exceptions: This obligation does not apply to systems permitted by law to detect, prevent, or investigate criminal offences, subject to appropriate safeguards.
Governance & Fines
Enforcement structures and penalties for non-compliance.
Governance Bodies
Central Union expertise; exclusive powers to supervise and enforce rules specifically on General-Purpose AI models. (Digital Omnibus Update: AI Office's exclusive competence over GPAI and VLOPs formally centralized in Art 75).
Composed of Member State reps. Advises on consistent application, harmonisation, and guidelines.
Independent experts alerting the AI Office to systemic risks and advising on GPAI classifications.
Notifying authorities and market surveillance authorities in each Member State enforcing rules on standard AI systems.
Penalties (Article 99 • Recital 168)
- €35M / 7%For non-compliance with the Prohibited AI Practices (whichever is higher, based on global turnover).
- €15M / 3%For failing to meet High-Risk obligations, GPAI rules, or transparency duties.
- €7.5M / 1%For supplying incorrect or misleading information to authorities.
* Source: Digital Omnibus on AI: Fines proportionality and privileges for SMEs extended to Small Mid-Caps (SMCs).
Presumption of Conformity
Allows providers of high-risk AI systems or general-purpose AI (GPAI) models to legally presume that their system satisfies specific mandatory requirements without needing to demonstrate compliance from scratch.
1. Compliance with Harmonised Standards (Article 40)
- Standardization Deliverables: High-risk AI systems or GPAI models that conform to harmonised standards (or parts thereof) whose references have been published in the Official Journal of the European Union (OJEU) are automatically presumed to comply with the corresponding requirements in Chapter III (Section 2) or Chapter V.
- Role of Standardisation Bodies: The European Commission issues standardisation requests to European standardisation organisations (such as CEN, CENELEC, and ETSI) covering risk management, data quality, transparency, human oversight, accuracy, and cybersecurity.
2. Compliance with Common Specifications (Article 41)
- Fallback Technical Rules: Where harmonised standards do not exist, are delayed, or insufficiently address fundamental rights concerns, the Commission can adopt common specifications via implementing acts.
- Presumption Mechanism: High-risk AI systems or GPAI models that comply with these common specifications (or parts of them) are presumed to satisfy the corresponding statutory requirements to the extent those specifications cover them.
3. Specific Presumptions for Contextual Data Governance (Article 42(1))
- Context-Specific Data Training: High-risk AI systems trained and tested on data reflecting the specific geographical, behavioural, contextual, or functional setting in which they are intended to operate are presumed to comply with the data governance requirements of Article 10(4).
4. Cybersecurity Certification Schemes (Article 42(2))
- Cybersecurity Act Alignment: High-risk AI systems certified (or issued a statement of conformity) under an approved EU cybersecurity scheme pursuant to Regulation (EU) 2019/881 (Cybersecurity Act) are presumed to fulfill the technical cybersecurity requirements under Article 15.
5. Presumption for Notified Bodies (Article 32)
- Conformity Assessment Bodies: Third-party conformity assessment bodies that satisfy relevant harmonised standards published in the OJEU are presumed to meet the organizational, competence, and independence requirements for notified bodies under Article 31.
While relying on harmonised standards or common specifications grants a presumption of conformity, providers must still draw up the technical documentation (Article 11), establish a quality management system (Article 17), issue an EU Declaration of Conformity (Article 47), and affix the CE marking (Article 48) prior to placing the system on the market.
Timeline & Milestones
Article 113 mandates a phased rollout following publication in the Official Journal, including recent updates from the Digital Omnibus on AI.
August 1, 2024
Entry into Force20 days after publication in the Official Journal (July 12, 2024).
February 2, 2025
+ 6 MonthsProhibited AI Practices (Chapters I & II) become strictly enforced.
Source: Digital Omnibus on AI: The original AI literacy obligation for businesses was replaced with a requirement for Member States and the Commission to foster AI literacy.
August 2, 2025
+ 12 MonthsGPAI rules apply. Penalties apply. Governance structures (Board, Notifying authorities) must be operational. Codes of practice finalised.
December 2026
Digital Omnibus ProhibitionsIntroduction of new strict prohibitions surrounding Child Sexual Abuse Material (CSAM) and non-consensual deepfakes.
Source: Digital Omnibus on AI
December 2027
High-Risk (Annex III)Rules for High-Risk AI systems (Annex III stand-alone systems) now apply. Member States must have at least one AI regulatory sandbox operational.
Source: Digital Omnibus on AI: These dates were extended via the Omnibus to tie enforcement to the availability of harmonized standards, with a hard deadline set for Dec 2027.
August 2028
High-Risk (Annex I)Rules apply for High-Risk AI embedded as safety components in regulated products (Annex I, e.g., machinery, medical devices, toys).
Source: Digital Omnibus on AI: These dates were similarly extended via the Omnibus with a hard deadline set for Aug 2028.
December 31, 2030
Long-Term MilestoneAI systems acting as components of large-scale EU IT systems (Annex X, e.g., Schengen Information System) must be brought into compliance.
Abbreviations
Common terminology used within the regulation.
Explorer created with AI support.
Wrap-up
A coming article will take a combined look at the CRA, AI, and EN IEC 62443/A11, not least triggered by the recent OpenAI and Hugging Face incident.