Article 53 requires providers of general-purpose AI models to draw up and keep up to date technical documentation, to provide information and documentation to downstream providers who intend to integrate the model, to put in place a policy to comply with Union copyright law, and to draw up and make publicly available a sufficiently detailed summary of the content used for training. A limited exception applies to certain free and open-source models for some of those duties, but the copyright policy and training summary duties remain relevant and systemic-risk models are treated differently.
What technical documentation does Article 53 require?
Article 53 requires technical documentation for the model, drawn up in accordance with Annex XI, and kept up to date. The documentation is intended to give the AI Office and, through the downstream information flow, other providers enough material to understand the model, its development, its capabilities and its limitations. In practice that means a provider should maintain a single controlled document set with version history, rather than scattered notes that cannot be shown to be current. The documentation should be proportionate to the model and should be updated when the model, its training, or its evaluation changes in a material way. A provider should record who owns the documentation, how updates are approved, and where the current version is held, so that a request for information can be answered from the controlled set rather than reconstructed after the event.
What information must flow to downstream providers?
Article 53 requires the model provider to provide information and documentation to providers of AI systems who intend to integrate the model into their systems, so that those downstream providers can understand the model's capabilities and limitations and can comply with their own obligations. The information is set out by reference to Annex XII and is distinct from the full technical documentation held for the authorities. Downstream teams need enough material to assess intended purpose, performance limits, foreseeable misuse, and any conditions on use, without receiving material that is not relevant to their integration. A provider should therefore design a downstream pack with a clear contents list, a version that matches the model version supplied, and a channel for updates and questions. The downstream provider should record what it received, what it still needs, and any gap it has raised with the model provider.
For the broader regulatory framework, see General-purpose AI models overview.
What must a copyright policy cover?
Article 53 requires a policy to comply with Union law on copyright and related rights. The policy should address, in general terms, how the provider seeks to respect rights reservations and lawful access when content is used for training, and how it handles the risk that a model produces infringing output. The policy is a governance document, not a statement that no infringing output can ever occur. It should name an owner, describe the processes relied on, explain how complaints or notifications are received and handled, and be kept up to date as training practices and guidance change. Providers should avoid copying a generic policy without checking that its processes match their actual crawling, licensing and filtering practices. Where training content information is unavailable or retrieval would impose a disproportionate burden, the position should be disclosed and justified in the relevant documents rather than left silent.
What is the public summary of training content?
Article 53 requires a sufficiently detailed summary of the content used for training the model, drawn up using a template provided by the AI Office and made publicly available. The summary is not a requirement to publish every training item, a dataset, or proprietary detail beyond what the template calls for. Its purpose is to give a meaningful overview of the kinds and sources of content used, so that rights holders and other readers can understand the training base in broad terms. A provider should complete the template carefully, check that the summary matches the model version it describes, keep it accessible at a stable location, and update it when a retraining run materially changes the training base. An outdated summary that describes an earlier model version can mislead readers and undermine the transparency purpose of the duty.
How does the open-source treatment work?
Article 53 provides a limited exception for general-purpose AI models released under a free and open-source licence that allows access, usage, modification and distribution, with parameters including weights and architecture publicly available. For such models, the technical documentation and downstream information duties in Article 53(1)(a) and (b) do not apply in the same way, subject to the conditions in the article. The exception does not remove the copyright policy and training summary duties, and it does not apply to a model with systemic risk in the same way. A provider relying on the exception should record the licence terms, what was made public, and why the conditions are met, and should check the position again if the licence, the release, or the classification changes. A model described as open in marketing material does not automatically meet the legal conditions.
Continue with Downstream information flow for general-purpose AI models.
What does this guide not decide?
This guide does not decide whether a provider has met Article 53 for a particular model, whether a documentation set is sufficiently detailed, whether a copyright policy is adequate, or whether a training summary meets the template. It does not assess a licence against the open-source conditions or determine the effect of a model update on the duties. Those questions depend on the actual documents, the model version, the template in force, and the consolidated legal text. A provider should have its Article 53 pack reviewed by a competent authority or legal adviser where the position is contested, and should keep the pack current rather than treating a single release as a permanent answer.
Frequently asked questions
Must a provider publish its full training data?
No. Article 53 requires a sufficiently detailed public summary of the content used for training, using the AI Office template. It does not require publication of every training item or of the dataset itself. The summary should be detailed enough to be meaningful and should match the model version it describes.
Does the open-source exception remove all Article 53 duties?
No. The limited exception in Article 53(2) concerns the technical documentation and downstream information duties for qualifying free and open-source models. The copyright policy and training summary duties remain relevant, and systemic-risk models are treated differently. The conditions and the release terms should be checked in each case.
How often should Article 53 material be updated?
Technical documentation and downstream information should be kept up to date as the model changes. A provider should update the pack when a new version is released, when training or evaluation changes materially, or when guidance or the template changes, and should record the version history so readers can tell which version a document describes.
Continue through this AI Act cluster
This article is an informational overview, not a legal determination. Check the current legal text and official implementation guidance for your system and role.
Consolidated Regulation (EU) 2024/1689 ↗AI Act Service Desk Article 53 ↗European Commission AI Act overview ↗