The GPAI baseline: Art 53 and the open-source carve-out

Lesson 3 of 5 in Transparency and General-Purpose AI: Art 50 and the GPAI Chapter.

Now switch legal objects. Everything so far in this domain regulated AI systems — products with an intended purpose. The GPAI chapter (Arts 51–56) regulates something upstream: the model itself. A general-purpose AI model (Art 3(63)) displays significant generality, competently performs a wide range of distinct tasks, and can be integrated into a variety of downstream systems — the definition even names the training signature, ‘trained with a large amount of data using self-supervision at scale’. Models used purely for pre-market research and development are excluded until placed on the market.

Why regulate the ingredient and not just the dishes? Because one foundation model feeds thousands of downstream systems, and its flaws — biases, security holes, hallucination patterns — propagate to all of them. The chapter has been applicable since 2 August 2025, supervised exclusively at Union level by the AI Office, with models placed on the market before that date owing compliance by 2 August 2027.

Every GPAI provider owes the Art 53 baseline, four duties:

  1. Technical documentation (Annex XI) — training and testing process, evaluation results — kept for the AI Office and national authorities.
  2. Downstream documentation (Annex XII) — the information package that lets system builders integrating your model understand its capabilities and limits and meet their own AI Act duties.
  3. A copyright policy — including honouring rights reservations under the text-and-data-mining regime, DSM Directive Art 4(3): rights holders who machine-readably opted their works out of TDM must have that respected in training.
  4. A public summary of training content, sufficiently detailed, using the AI Office template published in July 2025.

Then comes the carve-out everyone misquotes. Providers of models released under a free and open-source licence — with weights, architecture and usage information made publicly available, and the licence allowing access, use, modification and distribution — are exempt from duties 1 and 2 (the two documentation duties). They are never exempt from the copyright policy or the training-content summary, and the whole carve-out evaporates if the model is a systemic-risk model.

The logic: open release already delivers much of what the documentation duties chase — inspectable weights, community scrutiny — but copyright interests and training-data transparency do not become less urgent because the weights are downloadable. And at systemic-risk scale, openness is treated as a distribution fact, not a safety argument.

Non-EU GPAI providers must appoint an EU authorised representative (Art 54) before placing a model on the Union market — mirroring the high-risk value chain, and again with an exemption for open-source models unless they carry systemic risk.

Interactive sorting exercise: A provider releases a (non-systemic-risk) GPAI model under a genuine free and open-source licence, weights public. Sort the duties: which still bind, which are lifted?

Key terms: general-purpose AI, open-weight model, TDM opt-out, AI Office, downstream provider

Interactive checkpoint quiz (1 questions) — open this page in a browser to take it.