What the AI Act is
The AI Act is the world's first comprehensive AI law. It follows a risk-based approach: the higher the risk an AI system poses to health, safety and fundamental rights, the stricter the rules. It applies to providers who develop AI systems and deployers who use them in the EU, wherever they are based. CertifyAI is built around the high-risk tier, where most documentation and certification work sits.
The five risk tiers
Unacceptable risk — prohibited
Art. 5Applies since 2 February 2025
Practices banned outright: subliminal manipulation, exploitation of vulnerabilities, social scoring, predictive policing based solely on profiling, untargeted facial-image scraping and emotion recognition in workplaces and schools. Related biometric categorisation rules apply from 2 August 2025.
High risk
Art. 6, Annexes I & IIIStand-alone Annex III systems: 2 December 2027 · Annex I product safety components: 2 August 2028 · core obligations: 2 August 2026
AI used in critical areas — biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration and justice — plus AI as a safety component of regulated products (machinery, medical devices, radio equipment and others). These systems must meet the full obligations below.
General-purpose AI (GPAI)
Art. 51 et seq.Applies since 2 August 2025
Foundation models must keep technical documentation, honour the copyright directive and publish training-content summaries. Models with systemic risk face extra evaluation, adversarial testing and incident reporting duties.
Transparency risk
Art. 50Applies since 2 August 2026
People must know when they interact with AI: chatbots disclose they are AI, synthetic content is marked, and emotion-recognition or biometric categorisation is disclosed to those exposed.
Minimal risk
recitals, Art. 95No mandatory obligations
All other AI — spam filters, game AI, recommendations. Voluntary codes of conduct are encouraged under Article 95.
Key obligations for high-risk systems
| Article | Duty | What it means |
|---|
| Art. 9 | Risk management system | Continuous, documented risk assessment across the lifecycle. |
| Art. 10 | Data governance | Training, validation and testing data must be relevant, representative and checked for bias. |
| Art. 11 & Annex IV | Technical documentation | The Annex IV technical file: system description, design, testing and monitoring evidence. |
| Art. 12–13 | Logging & transparency | Automatically generated logs and clear instructions so deployers understand outputs and limits. |
| Art. 14 | Human oversight | Oversight measures so humans can monitor, intervene or stop the system. |
| Art. 15 | Accuracy & robustness | Declared accuracy metrics, resilience to errors and cybersecurity. |
| Art. 17 | Quality management | A QMS covering resources, procedures and post-market feedback. |
| Art. 27 | Fundamental rights impact assessment | Deployers of certain high-risk systems assess impacts on affected groups before first use. |
| Art. 43 & 47 | Conformity & declaration | Internal control or Notified Body assessment, then a signed EU Declaration of Conformity and CE marking. |
| Art. 49 & 71 | EU database registration | Registration of the system and declaration before placing on the market. |
| Art. 72–73 | Post-market & incidents | Post-market monitoring plan and serious-incident reporting on statutory clocks (2/10/15 days). |
Official EU documents
Always verify deadlines and duties against the official sources — CertifyAI's tools support the process but never replace the law itself.
How CertifyAI helps
CertifyAI turns these duties into a working process: classify systems against Articles 5 and 6, assemble the Annex IV technical file with zero-storage evidence, run Article 27 FRIA and Article 73 incident clocks, and dispatch sealed dossiers along the four Article 43 conformity routes. See How It Works for the full method, and the AI Act compliance statement for the app's own transparency disclosures. This page is informational and not legal advice.