Regulation structure: article map

The AI Act (Regulation (EU) 2024/1689) has 113 articles and 13 annexes. You do not need to read them all. These are the key blocks for a legal team:

Title I (Art. 1-4): Scope and definitions

• Art. 2: Territorial scope. Applies to providers that place on the market or put into service AI systems in the EU, regardless of where they are established. Also applies to deployers established in the EU.
• Art. 3: 68 definitions. The critical ones: "AI system" (Art. 3.1), "provider" (Art. 3.3), "deployer" (Art. 3.4), "putting into service" (Art. 3.11), "biometric data" (Art. 3.34).
• Art. 4: AI Literacy. Cross-cutting obligation already in force since February 2025.

Title II (Art. 5): Prohibited practices

• Social scoring, subliminal manipulation, exploitation of vulnerabilities, real-time remote biometric identification (with law enforcement exceptions), emotion inference in work/education (with safety exceptions), mass facial scraping, biometric categorization by race/orientation/beliefs.
• In force since February 2025.

Title III (Art. 6-49): High-risk systems

The most extensive block. Defines what is high-risk, what obligations providers and deployers have, and how conformity is assessed.

Title IV (Art. 50): Transparency

Obligations for systems that interact with persons (chatbots), generate content (deepfakes, AI-generated text), or detect emotions.

Title V (Art. 51-56): GPAI models

Obligations for providers of general-purpose AI models (OpenAI, Anthropic, Google, Meta). Documentation, copyright, systemic risk.

Titles VIII-XII (Art. 64-99): Enforcement

National authorities, European AI Office, regulatory sandboxes, penalties.

Annex III: high-risk systems in detail

Annex III is the reference for determining whether a system is high-risk. It lists 8 areas:

1. Biometrics

Remote biometric identification (not real-time), biometric categorization by sensitive attributes, emotion recognition.

2. Critical infrastructure

Safety components in management of traffic, water, gas, heating, electricity, internet.

3. Education and vocational training

Determining access to educational institutions, assessment of learning outcomes, assessment of the appropriate level of education, proctoring of examinations.

4. Employment and worker management

The most relevant for companies: recruitment (CV screening, candidate filtering), promotion/dismissal decisions, task assignment based on behavior, performance evaluation.

5. Access to essential services

Eligibility assessment for public benefits, credit scoring, insurance risk assessment, emergency call classification.

6. Law enforcement

Individual risk assessment (recidivism), polygraphs, illegal content detection, evidence reliability assessment.

7. Migration and asylum

Immigrant risk assessment, travel documents, asylum/visa applications.

8. Administration of justice

Fact-finding, interpretation of law, alternative dispute resolution.

Provider obligations for high-risk AI

The "provider" is whoever develops or commissions the development of an AI system and places it on the market under their name. Main obligations (Art. 8-15):

1. Risk management system (Art. 9): continuous and iterative process throughout the entire lifecycle. Identify, assess, and mitigate risks. Document.
2. Data governance (Art. 10): training, validation, and test data must be relevant, representative, free from errors to the extent possible, and complete. Examine biases.
3. Technical documentation (Art. 11): before placing on the market. Minimum content in Annex IV.
4. Record-keeping (Art. 12): automatic logging during operation. Decision traceability. Minimum retention per applicable law.
5. Transparency (Art. 13): clear instructions of use for the deployer. Information on capabilities, limitations, residual risks.
6. Human oversight (Art. 14): designed to be overseen by humans. Ability to intervene, correct, deactivate.
7. Accuracy, robustness, and cybersecurity (Art. 15): appropriate levels. Resilient to errors, failures, adversarial manipulation.
8. Quality management system (Art. 17): documented QMS.
9. Conformity assessment (Art. 43): before placing on the market.
10. Registration in the EU database (Art. 49): public database managed by the Commission.

Deployer obligations (professional user)

The "deployer" is whoever uses a high-risk AI system in a professional context (not end consumer). Most companies are deployers, not providers. Obligations (Art. 26):

1. Use according to instructions: follow the provider's instructions.
2. Human oversight: assign competent natural persons for oversight. With adequate training.
3. Input data: ensure that input data is relevant and representative.
4. Monitor operation: monitor that the system functions correctly. Report incidents to the provider and authorities.
5. DPIA (if GDPR applies): conduct a data protection impact assessment before using the system.
6. Inform affected persons: Art. 26.7: inform workers and their representatives that high-risk AI is used in the workplace.
7. Registration (if public sector): register the use in the EU database.

Conformity assessment: step-by-step process

Art. 43 defines two pathways:

Pathway 1: Self-assessment (most cases)

The provider internally evaluates that it meets all requirements. No external body needed. Applies to most high-risk systems EXCEPT biometrics.

1. Verify compliance with Art. 8-15 (technical requirements)
2. Complete technical documentation (Annex IV)
3. Implement QMS (Art. 17)
4. EU declaration of conformity (Annex V)
5. CE marking
6. Registration in the database

Pathway 2: Third-party assessment (biometrics)

For biometric identification systems: a notified body must validate. Longer and more costly process.

Required technical documentation

Annex IV defines the minimum content. For legal teams, these are the documents you must prepare or request from the technical team:

Block 1: General description

• System name and version
• Intended purpose and reasonably foreseeable uses
• Interaction with hardware/software
• Relevant software versions

Block 2: Development

• Description of the development process
• Design decisions and trade-offs
• System architecture
• Training data: sources, size, characteristics, known biases

Block 3: Performance

• Performance metrics (accuracy, recall, F1, etc.)
• Testing results
• Documented residual risks
• Implemented mitigation measures

Block 4: Post-market

• Post-market monitoring plan
• Incident reporting procedure
• Information for deployers (instructions of use)

GPAI models: specific obligations

Title V regulates general-purpose AI models (GPT, Claude, Gemini, Llama). This affects model providers, not directly the companies that use them. But as a legal team, you need to know:

Obligations for all GPAI (Art. 53)

• Technical documentation of the model
• Information for downstream providers (companies integrating the model)
• Copyright compliance policy
• Public summary of training data

Additional obligations for GPAI with systemic risk (Art. 55)

Models with high-impact capabilities (threshold: >10^25 FLOP of training). Currently: GPT-4, Gemini Ultra, Claude Opus.

• Model evaluation and adversarial testing
• Systemic risk assessment and mitigation
• Serious incident reporting
• Adequate cybersecurity protection

Practical implication for your company

If you use ChatGPT, Claude, or Gemini via API: GPAI obligations fall on OpenAI/Anthropic/Google, not you. But if you build a high-risk system ON TOP of these models (e.g., CV screening with the Claude API), YOU are the provider of that high-risk system and have Title III obligations.

Enforcement: authorities, penalties, timeline

Authorities

• European AI Office: supervises GPAI and coordinates between member states
• National authorities: each member state designates at least one. In Spain: AESIA (Spanish Agency for AI Supervision)
• Sector authorities: retain competence in their sectors (DPAs for data, central banks for finance, etc.)

Penalties (Art. 99)

Application timeline

Compliance checklist for legal teams

Ejercicio practico

Puntos clave