Regulation (EU) 2024/1689

EU AI Act Article 53: what providers of general-purpose AI models must document

Article 53 sets four documentation duties for anyone who provides a general-purpose AI model in the EU. This page sets out what the text asks for, when it applies, and which parts an AI BOM can give you evidence for.

Who it applies to

Article 53 binds providers of general-purpose AI models. Article 3(3) defines a provider as whoever develops a model, or has one developed, and places it on the market under its own name or trademark, paid or free. Article 3(63) defines a general-purpose AI model as one that "displays significant generality" and can be integrated into many downstream systems.

If you integrate someone else's model into your own AI system, you are a downstream provider (Article 3(68)). Article 53 information flows to you from the model's provider. Whether fine-tuning or modifying a model makes you its provider is covered in the Commission's guidelines for GPAI providers, not in Article 53 itself.

Key dates

2 Aug 2025
Article 53 obligations apply (Article 113(b), which brings Chapter V into application).
2 Aug 2026
The Commission can fine providers up to 3% of worldwide annual turnover or EUR 15 million, whichever is higher (Article 101, excluded from the 2025 date by Article 113(b)).
2 Aug 2027
Deadline for models placed on the market before 2 August 2025 (Article 111(3)).

What Article 53(1) asks for

  1. (a) Technical documentation of the model, its training and testing process and evaluation results, kept up to date and containing at least the items in Annex XI. It is provided on request to the AI Office and national authorities, not published.
  2. (b) Information for integrators: documentation that lets providers of AI systems understand the model's capabilities and limitations, containing at least the items in Annex XII.
  3. (c) A copyright policy, including identifying and honouring rights reservations made under Article 4(3) of Directive (EU) 2019/790.
  4. (d) A public summary of training content, "sufficiently detailed", using the template the AI Office published on 24 July 2025.

Article 53(2) exempts models released under a free and open-source licence, with weights, architecture and usage information public, from points (a) and (b) only. Points (c) and (d) still apply, and the exemption does not cover models with systemic risk.

Where an AI BOM helps, and where it doesn't

AIsbom inspects model files and their Hugging Face metadata and writes the result as a CycloneDX 1.7 ML-BOM. The table maps what that document records to the Annex XI and XII items it relates to. The right-hand column is our reading of the overlap, not the regulation's.

Article 53 / Annex item What AIsbom records Coverage
Annex XI 1(d), XII 1(f): architecture and number of parameters modelCard.modelParameters: architecture family and model architecture, from the Hugging Face card or a GGUF header Architecture yes. Parameter count is not recorded today.
Annex XI 1(a), XII 1(a): intended tasks modelCard task from the Hugging Face card Partial. A card's task tag is a starting point, not a full description.
Annex XI 1(f), XII 1(h): the licence Declared licence per model file, plus aisbom:hf:license Yes, as declared. AIsbom flags restrictive and unknown licences.
Annex XI 2(c), XII 2(c): training data and its provenance Training datasets named on the model card Names only. Provenance, curation and data volume are for the provider to write.
Annex XII 1(e): versions of relevant software Framework dependencies from requirements.txt, SHA-256 of each model file Yes for what the scanned repository pins.
Article 53(1)(c): copyright policy Nothing Out of scope. A policy is an organisational document.
Article 53(1)(d): public training-content summary Dataset names can seed it Out of scope. The summary follows the AI Office template and is written by the provider.
Annex XI 2(b), 2(d), 2(e): design choices, compute, energy Nothing Out of scope. None of this is visible in the model file.

Generating the evidence

pip install aisbom-cli

# CycloneDX 1.7 ML-BOM with a modelCard block per model
aisbom scan hf://your-org/your-model --output sbom.json

# How complete is the document? Grades the same dimensions the dashboard uses
aisbom score sbom.json

# SPDX 3.0 with the AI profile, if your tooling expects SPDX
aisbom scan hf://your-org/your-model --format spdx --spdx-version 3.0 --output sbom.spdx.json

New to the format? What is an AI BOM? covers what goes into one and how CycloneDX and SPDX represent it.

See which Article 53 evidence your repo is missing

Connect a repository and the readiness view breaks your latest scan down against the Annex XI items above, listing each gap with the article it maps to. Free on every plan.

Open the readiness view Other regulations

Sources

Last reviewed 2026-10-31 against the Official Journal text. This page is general information, not legal advice. AIsbom reports what an SBOM contains. It does not assess or attest to anyone's regulatory status, and it cannot see your processes, policies or evaluations. Article references are to the Regulation as published in the Official Journal; check EUR-Lex for later amendments.