GPAI — General-purpose AI models
General-purpose AI model: the subject of Chapter V of the Regulation, whose providers are subject to the obligations of Articles 53 and 54 and, if the model is classified as having systemic risk, also to those of Article 55.
It is the model underneath the tool, and its obligations belong to whoever makes it, not whoever uses it.
Which obligations it carries
A legal obligation of the provider of the model, applicable from 2 August 2025. Article 111(3) adds a transitional period: providers of general-purpose AI models that have been placed on the market before 2 August 2025 shall take the necessary steps in order to comply with the obligations laid down in the Regulation by 2 August 2027; Regulation (EU) 2026/1744 did not touch that paragraph (its point 39 replaces paragraph 2 and adds paragraph 4). Four obligations in Article 53(1): technical documentation of the model available to the AI Office and the national authorities; information and documentation for providers of systems that integrate the model; a policy to comply with Union law on copyright, including the identification of a reservation of rights; and a sufficiently detailed public summary of the content used for training. Models with systemic risk add those of Article 55: evaluation with adversarial testing, assessment and mitigation of systemic risks at Union level, reporting of serious incidents to the AI Office, and cybersecurity of the model and its physical infrastructure. For the deployer, verifying its provider’s GPAI documentation is an internal methodological control: there is no legal mandate requiring it by this route, unless the situation escalates to Articles 25, 26, 27 or 50, or to Article 5.
What it is not
They are not obligations of whoever uses the tool: the whole of Chapter V looks at the provider of the model. A company that contracts an assistant built on a general-purpose model acquires no obligation under Article 53 by doing so. And not all apply to all models: Article 53(2) exempts from points (a) and (b) models released under a free and open-source licence that meet the conditions the paragraph itself sets. The exemption does not reach points (c) and (d).
The nuance almost nobody captures
Point (d) is the most singular in the Regulation and the one with most practical consequences outside it: it requires making publicly available a sufficiently detailed summary of the content used to train the model, according to the template provided by the AI Office. It is information that did not exist before and that, once it exists, changes what a customer can demand from its provider without invoking the Regulation. And point (c) links the AI Act to copyright law by a route that surprises many: the reservation of rights expressed pursuant to Article 4(3) of Directive (EU) 2019/790.