Article 53 gives providers of general-purpose AI models four duties: keep technical documentation, give downstream providers the information they need, maintain an EU copyright compliance policy and publish a summary of training content. Free and open-source models are excused from the first two, but not for models with systemic risk.
The four duties
According to Article 53(1), providers must:
- Keep up to date technical documentation of training, testing and evaluation results, including the Annex XI information, and supply it to the AI Office on request.
- Provide information to downstream providers integrating the model, enough to understand its capabilities and limits and meet their own duties, including the Annex XII elements. Intellectual property and trade secrets stay protected.
- Maintain a policy to comply with EU copyright law, including identifying and respecting rights reservations under Article 4(3) of Directive 2019/790.
- Publish a sufficiently detailed summary of training content, using the AI Office template.
When it applies
The Commission page says general-purpose AI model obligations have applied since August 2025, and that from August 2026 the AI Office can request documentation, evaluate models, require corrective measures and issue fines. Models placed on the market before 2 August 2025 have until 2 August 2027, per the timeline. The Commission also lists supporting tools from July 2025: scope guidelines, a voluntary Code of Practice and a template for the public training summary.
The open-source exception
Article 53(2) waives only the documentation duties in points one and two for models under a free and open-source licence that allows access, use, modification and distribution, with parameters, architecture information and usage information publicly available. The copyright policy and the training summary still apply. The exception does not apply to models with systemic risk.
What downstream builders should ask for
If you build on someone else's model, request:
- Capabilities, limits and known failure modes.
- Intended and unsupported uses.
- Evaluation results relevant to your use.
- Training data summary and copyright policy.
- A contact for incident and change notices.
This is the information that lets you meet your own provider or deployer duties, and it is the reason Article 53(1)(b) exists.
What providers should prepare
A documentation pack aligned with the annexes, a public training summary built on the template, a written copyright policy and a downstream information sheet. Version them with the model release.
Model versus system
Remember the split: these duties attach to the model. A product built on the model is an AI system with separate classification. See the post on what counts as an AI system.
Version control matters here. A training summary or copyright policy that does not match the model actually shipped is worse than none. Tie each document to a release identifier and review them whenever the data mix changes.
Quick check
Does an open-source model avoid all Article 53 duties?
No. It skips the documentation duties in points one and two only, and only if there is no systemic risk. Copyright policy and training summary remain.
Who gets the documentation?
The AI Office on request, and downstream providers receive the information needed to integrate the model.
Sources
- AI Act Article 53, obligations for providers of general-purpose AI models
- European Commission, AI Act regulatory framework
- AI Act implementation timeline
Last reviewed: 19 October 2026.
