EU AI Act obligations dataset
The EU AI Act text is public. A structured, citable version of the parts a software vendor is actually asked about is not. This is that version: every operative provision as one row, with the trigger that switches it on, who owes it, the exception, and the date it applies from — including the dates as amended by the Digital Omnibus, Regulation (EU) 2026/1744.
Download
| File | Format | Contents |
|---|---|---|
| croissant.json | JSON-LD (Croissant 1.0) | Machine-readable metadata: record sets, fields, types, checksums |
| provisions.jsonl | JSON Lines | 65 provisions, 11 fields each |
| key-dates.jsonl | JSON Lines | 7 applicability milestones with their legal basis |
| GitHub repository | JSON | Source of the dataset, issue tracker for corrections |
Load it in three lines, no account and no key:
pip install mlcroissant
import mlcroissant as mlc
ds = mlc.Dataset("https://dealrescue.grantmastermind.com/dataset/croissant.json")
for record in ds.records("provisions"):
print(record["provisions/ref"], record["provisions/rule"])Article 50 — the transparency duties, in full
Article 50 applies from 2 August 2026. The Digital Omnibus moved the high-risk chapter to 2 December 2027 and 2 August 2028; it did not move Article 50.
| Ref | Role | Trigger | Duty | Applies from |
|---|---|---|---|---|
| Art. 50(1) | provider | AI system intended to interact directly with natural persons (chatbots, assistants, voice agents) | Inform the persons concerned that they are interacting with an AI system | 2026-08-02 |
| Art. 50(2) | provider | AI systems generating synthetic audio, image, video or text content | Mark outputs in a machine-readable format, detectable as artificially generated or manipulated | 2026-08-02 |
| Art. 50(3) | deployer | Emotion recognition system or biometric categorisation system | Inform the natural persons exposed to the operation of the system | 2026-08-02 |
| Art. 50(4) | deployer | Deepfakes (image, audio or video content that is artificially generated or manipulated); AI-generated or manipulated text published with the purpose of informing the public on matters of public interest | Disclose that the content has been artificially generated or manipulated | 2026-08-02 |
| Art. 50(5) | both | — | Information must be provided in a clear and distinguishable manner at the latest at the time of the first interaction or exposure, conforming to accessibility requirements | 2026-08-02 |
Applicability dates
| Milestone | Date | Basis |
|---|---|---|
| regulation entry into force | 2024-08-01 | Art. 113 as adopted |
| article 4 ai literacy applies from | 2025-02-02 | Art. 113 as adopted |
| article 5 prohibitions apply from | 2025-02-02 | Art. 113 as adopted |
| article 50 applies from | 2026-08-02 | Art. 113 as adopted |
| article 50 2 marking transition for systems already on market until | 2026-12-02 | Art. 113 as adopted |
| high risk annex iii applies from | 2027-12-02 | Art. 113(c)(i) as amended by Regulation (EU) 2026/1744, Art. 1(40) — covers systems classified as high-risk pursuant to Art. 6(2) and Annex III. |
| high risk embedded in regulated products applies from | 2028-08-02 | Art. 113(c)(ii) as amended by Regulation (EU) 2026/1744, Art. 1(40) — covers systems classified as high-risk pursuant to Art. 6(1) and Annex I. |
Coverage
Beyond Article 50, the dataset covers the articles that show up in enterprise AI questionnaires: Article 2, Article 25, Article 3, Article 4, Article 5, Article 50, Article 53, Article 6, Article 99. That is 65 rows in total.
How to cite
ONE PERCENT NETWORK LLC (2026). eu-ai-act-article-50: EU AI Act Article 50 transparency duties as machine-readable JSON. CC BY 4.0. https://github.com/onepercentnetworkllc/eu-ai-act-article-50
Licence and limits
Creative Commons Attribution 4.0 International. Commercial use permitted with attribution. This is orientation data, not legal advice: the authoritative text is the consolidated version in the Official Journal. Found an error? Open an issue — corrections are logged in the dataset itself.
Who maintains it
Deal Rescue writes the AI-compliance sections of vendor security questionnaires — 48 hours, $490, paid after delivery. We built this dataset because we needed it to answer them without guessing. Got one on your desk?