Carina Branco authors the opinion piece "Of Works and Prompts. The Challenges of Coexistence"

At a moment when artificial intelligence is reshaping the rules of both creative and economic activity, the regulation of training data has become one of the most pressing issues in European legal debate. In this opinion piece, Carina Branco, partner at Morais Leitão, analyses Arthur Mensch's (Mistral AI) proposal to create a universal levy on AI providers, assessing its legal credibility, its practical limitations, and the underlying strategic and political dimensions.

In the West, well-funded American companies operate under very permissive copyright regimes. In the East, there are Chinese open-source laboratories that train models at a lower cost, work faster, and do not worry about copyright.

Arthur Mensch (Mistral AI) suggests levying all companies marketing AI models in Europe. This would reframe the entire debate surrounding the threats posed by AI to the copyright ecosystem and create space for communal prosperity.

As Mensch rightly points out, the current system is not neutral. It is a jungle where nobody pays, enforcement is patchy, and legal uncertainty disproportionately affects those who pursue regulatory compliance. His proposal reverses this logic by introducing a predictable levy in exchange for legal certainty. While it is theoretically simple, it is complex to implement and has precedent in Europe. Private copying levies have long been applied to devices that can theoretically reproduce protected content, and companies such as HP, Canon and Apple have adapted to fee regimes in different countries under the Copyright Directive (Directive 2001/29/EC of the European Parliament and of the Council of 22 May 2001) and local transposing legislation without apparent disruption. Therefore, Mensch is not proposing anything out of the ordinary. Rather, he is extending a familiar regulatory mechanism to a new technology, which makes his proposal both legally credible and politically viable.

Private copying levies have also taught us that rudimentary instruments produce equivalent results. All buyers of printers, for example, pay the levy, regardless of whether the devices are used to copy protected works or not. This same design flaw will also apply here: all AI providers will pay, regardless of whether their training data focuses mainly on protected content.

The Spotify experience reinforces this concern. Distribution goes through a collective fund with no traceable allocation, and the money flows — though not necessarily to the creators of the content used to train the model. Independent creators are left with rhetoric but no financial reward.

From a legal perspective, the 'grand deals' model proposed by Petra Hansson remains an interesting prospect. Copyright litigation in Europe is still highly fragmented, unpredictable and costly. Consequently, assuming liability in exchange for a fee could gain regulatory traction and even win the approval of certain markets. However, difficult questions remain: can 'publicly available online' be defined with sufficient precision to withstand a legal challenge? What about oversight of non-European providers? Will we achieve an effective system, or will it be nothing more than aspirational?

Finally, a note on the strategic dimension of this issue. By proposing a universal levy on all economic operators in the sector, Mistral is positioning itself as a morally responsible actor. This would enable it to manage its own copyright exposure (following recent investigations into its training data) and increase the compliance costs of its larger, financially more powerful rivals.

That said, Mensch's idea is not naive; it merits debate and is ultimately feasible. Europe has the institutional memory and capacity to develop and implement the idea effectively.

However, it remains to be seen whether Europe has the strategic strength to negotiate its impacts, and whether there is sufficient geopolitical momentum to do so.