Today we publish Sharing the value of AI training: a collective remuneration mechanism, a discussion paper and our contribution to the European Commission’s targeted consultation on a better copyright environment for European creativity and innovation.
The paper builds on Beyond AI and Copyright, which we published in June 2025, and follows the call for a European content levy that Mistral made earlier this year. It sets out, as a basis for discussion, a more detailed mechanism tied to the training of AI models and built with the instruments of EU copyright and market-access law.
Two competing objectives
In its call for evidence, the Commission states that it wants to “improve the conditions for creators’ remuneration, while making it easier for providers of generative AI to access copyright-protected content”. These objectives pull in different directions. Measures aimed at one tend to come at the cost of the other: stricter reservation regimes make access to data for AI development harder, while broader exceptions leave rightholders without remuneration.
Our proposal starts from the observation that individual licensing cannot resolve this. A licensing market, no matter how well it functions, cannot reach most of the content on which AI models are trained. Where licensing does happen, it happens between the largest AI developers and the largest rightholders. It leaves out the long tail of individual creators and small rightholders, and it leaves out smaller AI developers. The way out is to stop trying to reconcile these objectives transaction by transaction, and to do it collectively instead.
How the mechanism works
The obligation is triggered by the commercial supply, on the EU market, of access to AI models trained on publicly available content. It applies wherever the company is established and wherever the model was trained. Companies pay a percentage of the EU revenue from the supply of model access into a single EU-level fund. In exchange, training on publicly available content is no longer subject to individual reservations under Article 4 of the CDSM Directive.
The remuneration takes the form of a statutory remuneration claim that cannot be waived and is subject to mandatory collective management. Research, and the development and release of models under open-source licences, fall outside the obligation; it arises only when such a model is commercially supplied.
Rightholders keep control, but they exercise it through access rather than through reservations. Content behind paywalls or other access controls stays outside the mechanism and remains available for licensing. Uses at the point of inference, such as retrieval-augmented generation and grounding, are unaffected by the proposal.
The fund splits the proceeds between two destinations. The larger part goes to creators and rightholders through collective management organisations, with a guaranteed share reserved for authors and performers. A significant, fixed share goes to the stewards of publicly available content: cultural heritage institutions such as libraries, archives and museums, open knowledge projects and public service media.
An invitation to discuss
The paper does not propose legislative text. It sets out the requirements an intervention has to meet, then breaks the mechanism into six parts: what triggers the obligation, how it operates, what it is calculated on, how the proceeds are distributed, how it relates to existing copyright rules, and what legal form it takes. For each part we state what we propose and identify the questions that remain to be settled, such as where the line runs for open models, the unit of use for deemed model access, and the size of the shares in the fund. These are genuinely open questions that can be answered, so they should not be read as reasons not to proceed.
We see the proposal as an invitation to discuss a European approach to AI and copyright: one that sustains the information ecosystem on which AI models are trained, supports cultural diversity, and gives European AI developers, including open-source developers, the legal certainty they need.
We developed this proposal in consultation with a range of stakeholders, including the AI developers Mistral and Black Forest Labs. The proposal is Open Future’s own, and we are solely responsible for its content.
