EU rules now require AI model makers to publish training summaries

The process wall

From 2 August 2025, obligations under Article 53 of the EU's AI Act, Regulation (EU) 2024/1689, apply to providers of general-purpose AI models, the regulation's own text states, with that date set by the Act's entry-into-application provisions. The wall a rights holder met before this date was practical: an artist could suspect a model had trained on scraped work but had no official document to point to, and no mechanism for objecting that a provider was bound to honour. Article 53 answers both problems at once.

What the documents show

The regulation requires a general-purpose AI model's provider to keep technical documentation, to maintain a policy for complying with EU copyright law, and to publish a sufficiently detailed public summary of the content used to train the model, drafted to a template the EU's AI Office is responsible for providing. The copyright-compliance duty is defined by cross-reference: providers must identify and comply with rights reservations made under Article 4(3) of the earlier Copyright in the Digital Single Market Directive, which lets a rights holder opt a work out of the text-and-data-mining exception. For content made available online, that directive says a reservation is appropriate only if expressed by machine-readable means, such as metadata or terms and conditions a crawler can parse, not a plain-text notice alone. A third document, the General-Purpose AI Code of Practice, published 10 July 2025 and endorsed by the Commission, gives providers a voluntary route to demonstrate compliance with these Article 53 duties, including the copyright policy.

Authorship and limitation

The mechanism is disclosure and machine-readable consent, not a licence or a guarantee. A training-content summary tells a rights holder what categories of material went into a model without necessarily naming individual works, and an Article 4(3) opt-out only binds providers going forward, for training after a reservation is in place and legible to a crawler; it neither unwinds earlier training nor creates a payment right. A provider that signs the voluntary code is presumed compliant with the letter of Article 53, a procedural benefit for the company, not a substantive promise about what any specific artist's work was or was not used for.

What survives

What survives for an artist is a concrete, if narrow, lever: a machine-readable reservation attached to hosted work, and a summary document to check against it. Editorially, the practical step this regulation supports is publishing an opt-out signal in a form a crawler reads, not just a human, since the directive treats that as the threshold for an appropriate reservation.

• Read a general-purpose model provider's published training-content summary against the AI Office's own template requirements.
• Check whether a hosting platform offers a machine-readable opt-out flag consistent with Article 4(3) of the 2019 directive.
• Compare a signatory's Code of Practice copyright chapter against the bare text of Article 53.

The obligations are new and the enforcement record is not yet established as of this writing; a maker should treat the summary and the opt-out as documents to check, not as proof that disclosure or consent has actually happened in a given case.