Dataset reference | EU AI Policy Observatory

Current download

Published document records: 237
Export timestamp: 2026-11-03T00:03:12Z
Publication cutoff: 2026-09-04

The download contains published records and their linked metadata, not a full-text archive. The timestamp identifies the data export; it is not the date of every source review. This URL is updated with the website and does not guarantee access to earlier snapshots.

Fixed versions

Version v0.4.0: fixed data package provides SQLite, JSON, a manifest with SHA-256 checksums, a table schema, field guidance and citation metadata. It remains separate from the changing download above. Later corrections receive new version identifiers.

Previous version v0.3.0 remains available unchanged.

Public field guide

The principal entities are documents, policy groupings, events, institutions, sources, concepts and relationships. They have different units and must not be combined as document counts. SQLite separates entities and links into tables; identifiers connect related records.

Identity and level: id identifies a record; slug identifies its website route. Titles and official references describe the represented text. record_level distinguishes principal records, supporting material, versions and attachments, not separate laws.
Dates and versions: document_date dates the text or version according to document_date_kind. publication_date dates the official publication represented by the primary cited source. Additional dates retain their individual meanings. Version status, legal status and editorial publication status are separate fields.
Attribution and evidence: Institution links specify roles such as author, publisher or commissioner; those roles are not interchangeable. Sources identify evidence URLs and retrieval or verification dates. Evidence locators identify supporting passages where recorded. An official host does not imply authorship.
Classifications: Sector and provenance tags, relevance classes, corpus tiers and policy groupings are editorial classifications. Multi-label assignments overlap. Concept associations are annotations, not authoritative definitions or completed thematic coding.
Relationships and events: Relationships identify endpoints, type, evidential basis and rationale. Official and analytical relationships are distinguished. An event date describes a policy event, not necessarily publication of its linked document.
Review and missing values: Expanded-review status and record-specific qualifications describe evidence coverage, not legal validity or a guarantee of accuracy. A null field or empty association is not evidence that the corresponding fact or relationship does not exist. Review credit alone does not establish independent double-review.
Annotation states: annotation_states distinguishes assigned values, reviewed non-assignments, inapplicability, insufficient evidence and unrecorded review states. Assigned values are not automatically independently verified. Existing empty arrays without explicit review evidence remain unrecorded. The original fields and filters are unchanged.

This is a reader's guide; the fixed package includes its complete SQLite table schema. Consult Methodology for inclusion criteria, date semantics, provenance and coverage limits.

Data reuse

Project-produced published metadata and editorial annotations are licensed by Yichen Hao under CC BY 4.0. Credit the author and dataset version, link the licence and indicate changes. This does not license third-party source texts, quotations, logos or personal data, and does not imply endorsement. See the licence scope and exclusions.

Citation and snapshot retention

Suggested attribution: Hao, Yichen. EU AI Policy Observatory. Public dataset, export timestamp 2026-11-03T00:03:12Z. Include this website's URL and your actual access date.

Retain the downloaded file used in your analysis and record a SHA-256 checksum of that file, together with the export timestamp, record identifiers, filters and exclusions. A checksum identifies bytes but is not an archive or proof of source accuracy. For a fixed version, use the versioned data package and cite its version URL. Independent archival preservation and public source-code reconstruction are not provided here.

When citing a policy text, cite its official source and represented version as well as any dataset record used. Access does not itself confer reuse rights in the original documents.