⭐ Star on GitHub

Summary: The EU’s horizontal AI regulation: a product-safety-style law that sorts AI systems into risk tiers and attaches obligations to each, plus a separate regime for general-purpose AI models.

Sources: raw/AI-regulations/Regulation - EU - 2024_1689.md (full official text, CELEX:32024R1689)

Last updated: 2026-07-28


Signed at Brussels 13 June 2024, published OJ L 2024/1689 on 12 July 2024, in force (source: Regulation - EU - 2024_1689.md). 113 articles across 13 chapters, plus annexes.

Structure

ChapterSubjectWiki page
IScope, definitions, AI literacy (Arts 1–4)this page
IIProhibited AI practices (Art 5)prohibited-ai-practices
IIIHigh-risk AI systems (Arts 6–49)high-risk-ai-systems, high-risk-requirements
IVTransparency obligations (Art 50)article-50-transparency
VGeneral-purpose AI models (Arts 51–56)gpai-and-systemic-risk
VIInnovation measures / sandboxes (Arts 57–63)[[../laws-wiki/ai-regulatory-sandboxes
VIIGovernance (Arts 64–70)[[../laws-wiki/eu-ai-act-governance-bodies
IXPost-market monitoring, incident reporting (Arts 72–94)serious-incident-reporting
XIIPenalties (Arts 99–101)enforcement-and-penalties
XIIIFinal provisions, application dates (Arts 102–113)ai-act-timeline

The risk-tier architecture

The Act is often summarised as four tiers. That framing is a reading aid, not a term the text uses:

  1. Unacceptable — prohibited outright under Article 5.
  2. High-risk — permitted subject to the Chapter III Section 2 requirements. Two routes in: safety components of products already under Union harmonisation legislation (Art 6(1)), or the Annex III use-case list (Art 6(2)).
  3. Transparency-only — Article 50 obligations attach to four situations regardless of tier, and are explicitly “not limited to systems classified as ‘high-risk’” (source: raw/Laws/The EU AI Act's Transparency Rules...md).
  4. Minimal — everything else, unregulated by this Act.

GPAI models sit orthogonally to this. Chapter V regulates the model, not the system, with its own two-tier split (all GPAI vs GPAI with systemic risk). A single organisation can be a GPAI model provider under Chapter V and a high-risk system provider under Chapter III for different products.

The Act’s own security requirement

Article 15 is the security clause and the single most important article in this corpus for AI security work — it is where the Act names data poisoning, model poisoning, adversarial examples and confidentiality attacks. See article-15-cybersecurity.

Article 55(1)(d) is the parallel for GPAI with systemic risk: “ensure an adequate level of cybersecurity protection for the general-purpose AI model with systemic risk and the physical infrastructure of the model” (source: Regulation - EU - 2024_1689.md).

Roles

Obligations attach to roles, not organisations: provider, deployer, importer, distributor, authorised representative, product manufacturer (Arts 16, 22–26). The role is not fixed — modifying a system or model can make you its provider. See provider-vs-deployer.

Extraterritorial reach

The Act reaches non-EU actors. A staffing-sector analysis notes the Act “may apply to deployers even if they are not based in the European Union” (source: raw/Laws/What the EU AI Act Means for Staffing Businesses...md). Providers of GPAI models established in third countries must appoint an EU authorised representative by written mandate before placing a model on the Union market (Art 54(1)) (source: Regulation - EU - 2024_1689.md).