EU law now makes AI model makers disclose training summaries

What happened

From 2 August 2025, obligations for providers of general-purpose AI models under the EU's Artificial Intelligence Act, Regulation (EU) 2024/1689, became applicable. Article 113 of the Regulation sets a staged timetable: Chapters I and II from 2 February 2025, and Chapter V, which contains the general-purpose AI provisions, from 2 August 2025. A voluntary Code of Practice, published by the European Commission on 10 July 2025, offers providers a way to demonstrate compliance with those obligations ahead of harmonised standards.

What the documents say

Article 53(1) of the Regulation requires a provider of a general-purpose AI model to "put in place a policy to comply with Union law on copyright and related rights, and in particular to identify and comply with... a reservation of rights expressed pursuant to Article 4(3)" of the 2019 Copyright in the Digital Single Market Directive, and separately to "draw up and make publicly available a sufficiently detailed summary about the content used for training" the model, using a template the Commission's AI Office is to provide. The referenced Article 4(3) of the earlier Directive is the EU's text-and-data-mining opt-out: the mining exception applies "on condition that the use... has not been expressly reserved by their rightholders in an appropriate manner, such as machine-readable means." The Commission's Code of Practice page, current as retrieved on 16 September 2026, describes a dedicated copyright chapter requiring providers to adopt such a policy, without itself detailing what a compliant training-content summary must contain in practice.

Why it matters for makers

Two separate legal mechanisms now interact for any model placed on the EU market: a rights holder's own opt-out under the 2019 Directive, and a model provider's obligation to publish what it trained on and to honour that opt-out. For a European composer or label, this is the first EU-wide requirement that a foreign or domestic model provider disclose a summary of its training sources at all, even in general terms, rather than leave the question to litigation.

What to check before you use it

Look for a provider's published training-content summary before assuming a model's outputs are clear of your own catalogue; the Regulation requires the summary to exist but does not require a work-by-work list. If you or your label have filed a rights reservation using machine-readable means under Article 4(3), that reservation is what a compliant provider must be identifying and honouring, so confirm it is actually in a format a provider's process would detect.

• Has the model I am using, or that trained on my catalogue, published a training-content summary under Article 53?
• Is my own catalogue's opt-out recorded in a machine-readable form a provider could actually find?
• Does the provider rely on the voluntary Code of Practice, a harmonised standard, or its own stated alternative means of compliance?

The obligations create a paper trail that did not exist before August 2025; they do not by themselves resolve whether a given model was trained lawfully, which remains a question of whether the opt-out was honoured in fact.

Continue in Rights & release, or Sony Music declared an opt-out from AI training everywhere.