Mistral AI: A Tax to Harmonize AI and Copyright

A Proposal for Royalties for AI Providers

Arthur Mensch, the founder of Mistral AI, highlights a major transformation that artificial intelligence imposes not only on usage but also on the legal frameworks governing the production and dissemination of knowledge. In Europe, this transformation is reflected in an increasing debate on how to reconcile copyright with AI models that rely on the massive ingestion of content. In this context, Mensch proposes, in an op-ed published by the Financial Times, an innovative idea: to establish a royalty based on the revenues of AI providers. This royalty would be intended to fund a support fund for creation while providing legal security for training models. This proposal, presented as a compromise, deserves particular attention as it seems less focused on protecting creators than on adapting their production to an already well-established data economy on a global scale.

Varied Responses to a Global Problem

The tension between artificial intelligence and copyright is not exclusive to Europe. It is already being arbitrated, de facto, by major tech powers, each adopting distinct approaches.

• In the United States, the approach is often described as pragmatic, even a fait accompli. Models are trained and products launched, leaving disputes to emerge later. This method allows AI players to benefit from a decisive advantage: speed. Agreements between AI players and press groups coexist with ongoing legal proceedings, thus building a legal framework a posteriori.

• For content publishers, the situation is ambivalent. Those who possess differentiating assets, such as archives or brands, manage to negotiate agreements, often in an asymmetrical power dynamic. They capture a portion of the value but become more dependent on platforms. In contrast, the majority of content producers find themselves integrated into training corpora without negotiation capacity or visibility on usage, creating a new hierarchy between creators and holders of negotiable assets.

• In China, data access is controlled but largely exploitable on a domestic scale. The state regulates industrialization without blocking it, allowing local players to benefit from a volume of data and strategic coherence that favors the rapid integration of models.

• Europe, for its part, has attempted to organize a balance upstream, notably through opt-out mechanisms allowing rights holders to oppose the use of their content. However, this framework remains fragmented and unclear for industrial players, leading to persistent legal uncertainty and a less favorable environment for scaling.

For publishers and creators, this architecture theoretically offers a higher level of protection, but under limited operational conditions. Exercising the right to oppose requires a technical and legal capacity that is rarely homogeneous, while the circulation of content online quickly dilutes the effectiveness of this right. The most structured players attempt to organize protection or negotiation mechanisms, without guarantees of comprehensive coverage, while others remain exposed, with no visibility on the use of their content. Europe thus produces imperfect protection, sufficiently constraining to hinder industrial players but insufficiently effective to restore control to creators over the exploitation of their works.

Transforming Copyright into an Economic Lever

Arthur Mensch proposes to recognize the use of content and integrate it into an economic mechanism, rather than attempting to control each usage individually. A royalty would be applied to all AI providers operating in the European market, including foreign players. The revenues collected would fund a dedicated creation fund, in exchange for which AI companies would benefit from legal security for training their models using content accessible online.

This change is significant, as copyright does not disappear but changes its nature. From a tool of control, it becomes an instrument of redistribution.

From Explicit Consent to Data Flow

This change is accompanied by a deeper transformation: the principle of explicit consent, a cornerstone of copyright, gives way to a presumed usage logic. This weakens the link between the work and its author, putting pressure on the inalienability of moral rights, central to the European legal tradition.

The value of content also tends to homogenize. In the training phase, models do not distinguish between a news article, a forum thread, or a codebase. They exploit massive corpora, where relevance results from quantity and statistical diversity, rather than the uniqueness of each work. This evolution comes into tension with the European conception of copyright and could provoke reactions from creators.

The consequence is direct: individual contribution becomes difficult to isolate, and remuneration detaches from actual usage, depending on a collective mechanism whose governance and distribution criteria remain to be defined. Content ceases to be a controlled asset and becomes a diffuse resource.

An Industrialization of Content Already Underway

This trend is not a European singularity but fits into a global dynamic where content is progressively integrated into industrial value chains.

• In the United States, this integration occurs through bilateral agreements. Players with premium assets manage to negotiate directly with AI companies. For example, the Axel Springer group has reached an agreement with OpenAI to allow the use of its content, while the Associated Press has signed a similar partnership for access to its archives.

• However, this negotiating capacity remains concentrated and requires a critical mass of content, legal structuring, and real bargaining power. The majority of content producers, particularly independents, small publishers, forums, blogs, or open databases, continue to feed training corpora without access to direct monetization mechanisms.

This results in a two-speed market structure: on one side, players capable of transforming their content into negotiable strategic assets; on the other, a mass of production integrated without identifiable value capture. Market logic does not hinder the industrialization of content but organizes the hierarchy in favor of dominant players.

In China, integration is more direct, with data mobilized within a framework driven by a national technological strategy.

Mistral's proposal introduces a third path with organized mutualization at the scale of the European market. It does not challenge the industrialization of content but proposes a regulated form of it.

At first glance, this mechanism may evoke existing systems like collective rights management, embodied by organizations like SACEM, or remuneration for private copying. In these models, the usage of works, difficult to trace individually, is compensated by collective redistribution based on upstream levies.

However, the comparison quickly finds its limits, as collective management relies on a stabilized framework where usages are identifiable and rights holders are recorded.