Collective Licensing for Gen AI Training: Feasible or Flawed? – Part 2

On 10 March 2026, the European Parliament adopted a resolution on copyright and generative AI urging the EU Commission to clarify and potentially update the existing copyright framework for GenAI, facilitate voluntary collective licensing agreements, propose full transparency obligations for AI providers and establish a rebuttable presumption of infringement where those obligations are not met, while also calling on the EUIPO to become the trusted intermediary to manage rightsholder opt-outs. As the Commission prepares to review the CDSM Directive no sooner than June 2026, statutory licensing and collective rights management could now be on the agenda as a potential fix. But can it deliver?

This blog post series examines potential collective licensing models and finds that each carries significant structural limitations. Part 1 of this post outlined why the existing text and data mining (TDM) exceptions framework falls short, why voluntary licensing - while a useful complement - does solve its core inadequacies, and why statutory licensing either risks failing the three-step test or would need to be so narrowly drawn as to offer AI developers little meaningful legal certainty. 

Extended collective licensing: the most balanced compromise — but the devil is in the detail

Collective licensing with extended effect or extended collective licensing (hereinafter jointly referred as ECL) extends a CMO-negotiated licence to non-member rightsholders within a defined category, unless they opt out.

Two paths exist: an EU-wide ECL framework requiring legislative reform at Union level, or national ECL regimes introduced within the boundaries of Article 12 CDSM Directive. Article 12 of the CDSM Directive permits Member States to introduce ECL on a national level for uses where "obtaining authorisations from rightsholders on an individual basis is typically onerous and impractical" - a description that fits gen AI training well.

The EU Council questionnaire found ECL approach to be preferred even by CMOs. Spain became the first EU country to attempt national ECL regime tailored to generative AI, publishing a draft Royal Decree in November 2024 (on which, see Nobre). It was withdrawn two months later following strong opposition from authors' organisations - illustrating that the design details matter enormously.

Several structural challenges must be addressed in any ECL framework for AI, for example:

Territorial fragmentation. While AI training is inherently cross-border, Art. 12 ECL is effective only within the Member State which chooses to implements such regime. Unless an EU-level ECL mechanism is developed, developers would face a patchwork of national licences - negating the efficiency gains the model is supposed to deliver. Extended collective management and multi-territorial music licensing under Directive 2014/26/EU offers a partial precedent, but the 2021 Commission Report as well as the 2025 German Federal Ministry of Justice Report found that its potential has not been fully exploited.

Scope of rights and the memorisation problem. Limiting ECL to reproduction rights (mirroring the TDM exception) leaves unresolved the liability for AI outputs that reproduce training data verbatim - a risk that cannot be fully eliminated by technical means. While extending ECL to communication-to-the-public rights could resolve the regurgitation and memorisation issues, it could unproportionately interfere with rightsholders exclusive rights.

Interaction with the TDM exception. Since TDM exceptions stem from EU law, member-state ECL cannot override them. National ECL regimes can only operate where a valid opt-out under Article 4 has been declared - creating a cumbersome double-opt-out structure on top of a mechanism already acknowledged to be technically fragile.

Retroactivity and "unlearning." Once training data is embedded in model weights, removing specific contributions is technically very challenging without full retraining. The Spanish proposal required developers to "cease using" opted-out works within ten working days - without explaining how this would work for models already on the market.

Double remuneration. Where developers hold pre-existing direct licences, a new collective regime risks obliging payment twice for the same use.

Although ECL could become the most balanced compromise - provided it is designed to address the many considerations that arise – it remains territorially limited (unless adopted at EU level) and procedurally complex, as illustrated by the short-lived Spanish proposal. If implemented at member state level, ECL cannot override the TDM exception and thus remains available only for authors who have not opted out which creates an additional layer of legal uncertainty in addition to the existing uncertainty surrounding the rightsholders’ opt-out.

An EU-level ECL regime would require resolving several structural questions: the scope of rights covered would need to be clearly defined, as limiting ECL to reproduction rights leaves downstream liability for memorisation and regurgitation unresolved, while extending it to communication or making-available rights could face proportionality issues; the interaction with TDM exceptions and the Article 4 opt-out would need clarification, since without amending the CDSM Directive ECL could only operate where a valid reservation has been issued, producing an incoherent double-opt-out system. In short, introducing ECL for AI training would require the EU to confront not only doctrinal issues under international and EU copyright law, but also unresolved technical limits of machine-learning systems.

Beyond collective licensing: a case for digital infrastructure

The difficulties with all collective licensing models suggest that the more promising direction may be technical rather than primarily legislative: developing interoperable digital licensing infrastructure capable of operating at AI scale while relying on voluntary contractual licensing. Programmatic advertising infrastructure, which handles billions of micro-transactions daily, could serve as an inspiration for a design model for standardised licensing metadata and usage logging.

The EU Commission is already exploring a central registry of machine-readable TDM opt-outs, and the EUIPO's Strategic Plan 2030 envisages a Copyright Knowledge Centre offering services "facilitating licensing and ADR" and “work in the area of copyright and AI”. Such infrastructure could one day serve as a licensing hub - enabling developers to identify, verify, and obtain licences efficiently without requiring the complex legislative amendments that EU-wide ECL or statutory licensing would demand.

This approach is also supported by the Parliament's proposal which calls for voluntary licensing and granting EUIPO responsibility for facilitating such licensing.

Conclusion

None of the collective licensing models offers a complete answer. Voluntary licensing preserves autonomy but cannot scale. Statutory and mandatory models risk disproportionate interference with exclusive rights and face significant three-step test hurdles. ECL could be the most balanced compromise but faces many structural issues. National ECL regimes remain territorially constrained and entangled with the existing TDM exception and the dysfunctional opt-out they cannot replace. An EU-wide ECL framework would require complex legislative reform at Union level - and would still need to resolve fundamental questions around scope of permitted rights, liability for memorisation, double remuneration, and the technical impossibility of meaningful unlearning. The short-lived Spanish proposal illustrated that even carefully designed attempts will wake the underlying structural questions.

The CDSM Directive review - due no sooner than June 2026 - is the natural moment for the EU to confront these structural questions. Scalable digital licensing infrastructure - built on existing EUIPO frameworks and the planned opt-out registry - may be capable of delivering greater efficiency and legal certainty than new collective licensing regimes, at lower regulatory cost. A pragmatic step within the current framework could be enhanced cross-border and sectoral coordination between CMOs – for instance, a pan-European licensing portal or "one-stop shop". The Parliament's proposal to assign the EUIPO responsibility for facilitating such licensing is a welcome move in that direction.   

This blog post is an adapted and shortened version of the article titled “Collective Licensing for Gen AI Training: Feasible or Flawed?” published in European Intellectual Property Review (EIPR), vol. 2026, No 3, p. 143 - 158. ISSN 0142-0461. For the full length version, see the EIPR journal.

Image created by AI.