© 2026 Eugen Stoica

The European Copyright Society held its 10th annual conference at the Château de Versailles on 29 May 2026, under the theme Copyright, Artificial Intelligence and Cultural Heritage Institutions. The programme covered studies, legislative updates and doctrinal reviews, with presentation slides of the speakers attached here, but the most useful sessions, for my purposes, were the two roundtables at the crossroad of copyright (CDSM Directive) and AI (the EU AI Act).

First Roundtable: The Institutions

The first roundtable, moderated by Prof. Alain Strowel (Louvain University), sat under the session heading Reception and Utility of an EU Copyright Doctrine and brought together four voices from the EU institutional apparatus: Hubert de Verdelhan (Legal Secretary, Court of Justice of the European Union), Emmanuelle Du Chalard (Head of the Copyright Unit, DG Connect), Julie Samnadda (Legal Service of the European Commission) and Axel Voss MEP. The intended arc ran from adjudication through policy formation to the legislature.

De Verdelhan’s contribution was necessarily brief: there is not yet much CJEU case law on the AI/copyright interface to report on.

Julie Samnadda (European Commission Legal Service) described the current settlement as the “tokenisation of copyright”: a regime in which we no longer speak of works but of data and content, with too many rightsholders whose interests are rarely aligned. In Europe, market fragmentation is a structural feature rather than an incident. In the US there are fewer rightsholders and it is much easier to licence their works. There is not much comparable data available for other jurisdictions.

Emmanuelle Du Chalard, Head of the Copyright Unit at DG Connect, set out the state of the Commission’s call for evidence on the review of the CDSM Directive, which opened on 13 May 2026 and has now closed. The priority areas listed are AI, piracy, reserved rights and research. The proposals under consideration include a Secondary Publishing Right, a universal and mandatory TDM exception, and a dedicated research exception.

Axel Voss MEP took a more conservative line. His preference is for a legal framework that strikes the balance once and is then left to operate by itself, with the licensing market doing the adjusting; he floated the idea of something like a stock exchange for IP, where licences can be bought and sold by participants. This, he suggested, would address the fragmentation problem. On AI, his position was that crawling should not be the default route to collecting data, that permission should be sought in advance, and that AI companies should operate under a watermarking regime, with EUIPO potentially acting as a trusted third party for deposit and verification. He also invoked the territoriality principle: an LLM offered in the EU should require a licence as is already the case for regulated sectors such as banking, insurance etc. He spoke briefly about protections for press and publishers and the prospects for collective licensing on that side of the market. He closed by noting that the discussion is much bigger than copyright and AI – a reference, I took it, to the geopolitical pressure on EU regulatory autonomy.

Second Roundtable: The Stakeholders

The second roundtable, sitting under the same session heading and again moderated by Prof. Strowel, moved from the EU institutional apparatus to the constituencies outside it: Paul Keller (Director of Policy, Open Future), Burak Ozgen (Deputy Manager, GESAC — the European Grouping of Societies of Authors and Composers) and Blanche Savary de Beauregard (General Counsel, Mistral AI). Between them the three organisations cover the open-access/civil society, rightsholder and AI developer positions on which any future settlement will turn.

Keller’s opening point was that there is no necessary requirement that everything ingested into an LLM be attributed. Attribution is a design choice, not a doctrinal floor.

Ozgen offered a five-part diagnosis of the current position. First, unauthorised use of works remains widespread. Second, the licences that do exist are concentrated and structured as data partnerships rather than copyright licences in any meaningful sense, with authors seeing very little of the proceeds. Third, Article 14 CDSM (works of art in the public domain) is doing useful work but unevenly. Fourth, the AI Act’s architecture does not directly address copyright. Fifth, substitution effects are now visible, with AI outputs beginning to displace the original works on which they were trained.

Savary de Beauregard acknowledged that AI is not European (basically created in US and China), and that EU datasets are at real risk of being excluded from LLM training pools – both because the fragmentation of ownership makes licensing impractical and because the EU TDM opt-out regime, on a strict reading, gives rightsholders grounds to demand their removal from existing models. EU may be a big market and a regulatory power but if too much time is lost in negotiations, GenAI companies will sidestep European datasets. European creation, she said, needs to be remunerated if it is to remain sustainable. There is no functioning marketplace for licences, no collective organisation that represents most creators, and the time required to build a clean licensing infrastructure is time the sector does not have. Reopening the Directive is therefore on the table, and a universal levy financing authors is, on her view, a serious option.

Ozgen’s response was that a frontier model developer accepting that a levy is necessary is a substantive shift. Collective licensing has not delivered. Levies, by contrast, have an established track record in Europe: they are already used to compensate rightsholders for private copying of audio and video and for reprography (the levy on photocopiers, scanners and printers), administered through national collective management organisations under Article 5(2) of the InfoSoc Directive. Extending the model to AI training would not require inventing a new mechanism, only widening the scope of an existing one.

The two roundtables, taken together, suggest that the next round of EU copyright reform will be shaped at least as much by what the AI developers are now willing to concede as by what the rightsholders are still asking for.