license	cc-by-4.0
pretty_name	EU AI Act (Regulation (EU) 2024/1689) as structured data — SafeLegalAI
language	en
size_categories	n<1K
tags	legal law eu-ai-act regulation ai-governance compliance safelegalai
configs	config_name data_files articles split path train data/articles.parquet config_name data_files articles_as_enacted split path train data/articles_as_enacted.parquet config_name data_files amendments split path train data/amendments.parquet config_name data_files recitals split path train data/recitals.parquet config_name data_files annexes split path train data/annexes.parquet config_name data_files annexes_as_enacted split path train data/annexes_as_enacted.parquet config_name data_files definitions split path train data/definitions.parquet config_name data_files definitions_as_enacted split path train data/definitions_as_enacted.parquet config_name default data_files obligations true split path train data/obligations.parquet config_name data_files milestones split path train data/milestones.parquet config_name data_files authorities split path train data/authorities.parquet config_name data_files penalties split path train data/penalties.parquet config_name data_files member_states split path train data/member_states.parquet

EU AI Act, structured

Regulation (EU) 2024/1689 (the Artificial Intelligence Act) as tables: every article, recital, annex and definition, 677 obligations coded by actor, risk tier, application date and penalty basis, plus milestones, national competent authorities and fine tiers.

Built 2026-09-08 by SafeLegalAI (Cognesio LLP) from the official English texts served by the Publications Office of the European Union (Cellar): the consolidated text as of 27 July 2026 (CELEX 02024R1689-20260727 — the Act as amended by Regulation (EU) 2026/1744, the Digital Omnibus on AI, in force 27 July 2026) for articles, definitions, annexes and the coding; the text as enacted (CELEX 32024R1689) in the *_as_enacted tables and for recitals; and a per-article diff in amendments. Canonical pages: safelegalai.com/topics/eu-ai-act · pipeline and issues: https://github.com/SafeLegalAI/eu-ai-act-structured.

Tables

config	rows	what a row is
articles	119	one article of the consolidated text (27 July 2026): number (string for inserted articles such as 4a, 75a), title, chapter, section, numbered paragraphs with lettered points (JSON), flattened text, EUR-Lex anchor
articles_as_enacted	113	the same for the text as enacted in 2024
amendments	72	one amended or inserted article: words before/after, a word-level diff, the amending act and its entry into force
recitals	180	one recital
annexes	14	one annex: title and text
definitions	70	one Article 3 definition: number, term, definition
obligations	677	one distinct obligation, prohibition, right or institutional duty, coded by SafeLegalAI from the article text: actor[], obligation_type, risk_tier, applies_from (+ applies_from_basis), penalty_basis, legal_practice_relevance (+ note), cross_references[], and the Regulation's operative words in quote (verbatim, ≤ 60 words; every quote is machine-checked against the parsed text)
milestones	43	one dated milestone: entry into force, staged application, transitional dates, Commission deadlines — with legislative status (past, scheduled, proposed, deferred) and official source
authorities	29	one national competent authority under Article 70, per Member State, with role and designation status
penalties	9	one fine tier from Articles 99–101: conduct, maximum fixed amount, turnover percentage, rule
member_states	147	one national implementation instrument for each of the 27 Member States plus NO/IS/LI/CH/GB: implementing act, authority designation (market surveillance, single point of contact, notifying, fundamental-rights), Article 99 penalty regime, Article 57 sandbox status (deadline 2 August 2027), deployer guidance reaching legal practice, position on the Digital Omnibus — original-language title, English title (flagged when translated by us), status, date, ≤25-word quote, gazette/authority source; a state with nothing verifiable carries one authority-page row saying what was checked. Page per state: https://safelegalai.com/regulation/eu-ai-act/<cc>

Obligations by risk tier

risk_tier	rows
high-risk	352
not-tier-specific	205
gpai	40
gpai-systemic-risk	27
prohibited-practice	24
all-ai-systems	21
transparency-risk	8

Obligations by application date (Article 113 as amended by Regulation (EU) 2026/1744; applies_from_as_enacted keeps the 2024 date)

applies_from	rows
2024-08-01	1
2025-02-02	45
2025-08-02	157
2026-07-27	21
2026-08-02	296
2026-12-02	6
2027-08-02	1
2027-12-02	143
2028-08-02	7

198 rows are coded legal_practice_relevance: high — the provisions that reach a law firm, chambers, in-house team or court deploying AI, a court as public-authority deployer, or a legal-AI vendor as provider (including Annex III point 8, AI systems intended to assist a judicial authority).

What is the Regulation's and what is ours

The text (articles, recitals, annexes, definitions, every quote) is the European Union's, reused under Commission Decision 2011/833/EU — attribution: © European Union, 1998–2026, https://eur-lex.europa.eu. Only the official text at EUR-Lex is authentic; this dataset is a convenience and may lag amendments.

The coding (actor, obligation_type, risk_tier, applies_from, penalty_basis, legal_practice_relevance, summary, legal_practice_note) is SafeLegalAI's, released CC BY 4.0 — attribute SafeLegalAI (safelegalai.com), published by Cognesio LLP. It is descriptive, not legal advice; it records what the Regulation says, not what anyone should do. Corrections: safelegalai.com/report.

Method

1. pipeline/parse_eurlex.py fetches the ELI-structured XHTML from Cellar and parses chapters, sections, articles, paragraphs, points, recitals, annexes and definitions (deterministic).
2. Three coding passes (Arts 1–28, 29–70, 71–113 + annexes) wrote one row per distinct duty using only the parsed text, quoting the operative words; a fourth pass compiled milestones, authorities and penalties from official Commission and national sources.
3. pipeline/build_release.py merges, checks every quote is a verbatim substring of the parsed article/annex text, and writes JSONL, CSV and Parquet.
4. Human review: SafeLegalAI's editor re-opens a sample of rows before each release; the legal_practice_relevance column in particular is editorial coding.

Use

from datasets import load_dataset
ob = load_dataset("safelegalaidata/eu-ai-act-structured", "obligations")["train"]
firm = ob.filter(lambda r: r["legal_practice_relevance"] == "high" and "deployer" in r["actor"])

Uses

Suited to: counting and comparing what the record shows (by court, jurisdiction, date, actor, outcome, status); building watch-lists and alerts from source_url/fetched_at; grounding retrieval or summarisation on cited primary documents; teaching and library guides that need a dated, sourced list.

Not suited to: ranking products, people or courts; inferring prevalence beyond what a court or regulator has itself stated; any use that treats a coding column as a finding of fact or law. Where a row names a person or organisation it does so as they appear in a public document; anyone named may request a correction or right of reply at https://safelegalai.com/report.

Cite

SafeLegalAI (Cognesio LLP), "EU AI Act, structured", v0.3.0, 2026-09-08. https://huggingface.co/datasets/safelegalaidata/eu-ai-act-structured — text © European Union, reused under Decision 2011/833/EU; coding CC BY 4.0.

@dataset{safelegalai_eu_ai_act_structured_0_3_0,
  title        = {EU AI Act, structured},
  author       = {{SafeLegalAI (Cognesio LLP)}},
  year         = {2026},
  version      = {0.3.0},
  url          = {https://huggingface.co/datasets/safelegalaidata/eu-ai-act-structured},
  note         = {Data CC BY 4.0. Built 2026-09-08. Canonical: https://safelegalai.com}
}

Disclaimer and notices

Provided "as is", without warranty of any kind (CC BY 4.0 §5; Apache-2.0 §7). Not legal advice; Cognesio LLP is not a law firm. Only the Official Journal text of Regulation (EU) 2024/1689 is authentic; this dataset may lag amendments and corrigenda. Every coding column is SafeLegalAI's good-faith reading of the text for comparison — not an interpretation of the law you may rely on, and not a statement of how any authority will apply it. Application dates follow Article 113 as written unless the milestones table records an amending act published in the Official Journal; check milestones before relying on a date. Names of institutions and authorities identify them only. Full terms, notice-and-takedown and governing law (England and Wales): https://safelegalai.com/disclaimer · repository DISCLAIMER.md.

Manifest

{
  "regulation": "Regulation (EU) 2024/1689 (Artificial Intelligence Act)",
  "celex": "02024R1689-20260727",
  "source": "http://publications.europa.eu/resource/celex/02024R1689-20260727",
  "source_format": "application/xhtml+xml (Cellar, Publications Office of the European Union)",
  "eli": "https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng",
  "source_sha256": "5e7719f77e8a606b257dc25958ee3222c4383300a5a34270a5b850a2ce8b8715",
  "parsed": "2026-09-06",
  "counts": {
    "articles": 119,
    "articles_as_enacted": 113,
    "amendments": 72,
    "recitals": 180,
    "annexes": 14,
    "annexes_as_enacted": 13,
    "definitions": 70,
    "definitions_as_enacted": 68,
    "obligations": 677,
    "milestones": 43,
    "authorities": 29,
    "penalties": 9,
    "member_states": 147
  },
  "reuse": "Commission Decision 2011/833/EU \u2014 attribution: \u00a9 European Union, 1998\u20132026, https://eur-lex.europa.eu",
  "version": "0.3.0",
  "built": "2026-09-08",
  "contentSha256": "ed5ec8bb7470ce9ed32bd73e33e60e9f5ab744812497f34b21e866ff47fac25d",
  "canonical": "https://safelegalai.com/topics/eu-ai-act",
  "repository": "https://github.com/SafeLegalAI/eu-ai-act-structured",
  "license_text": "Commission Decision 2011/833/EU (\u00a9 European Union)",
  "license_coding": "CC BY 4.0 (SafeLegalAI, Cognesio LLP)"
}