The meteoric rise of generative artificial intelligence has put intellectual property at the heart of public debate.
The meteoric rise of generative artificial intelligence has put intellectual property at the heart of public debate. Authors, publishers, artists, researchers, industrialists: all legitimately question how to protect and remunerate creation in the era of major language models. But by focusing the debate on the model training phase, we risk missing the essential point – and constructing responses that are unsuitable for the technical, cultural and economic reality of AI.
When the debate on intellectual property misses the target
Much of the current discussion focuses on the following question: do AI models have the right to train on content protected by intellectual property when it is publicly accessible? Asked like this, the question seems legitimate. However, it is poorly worded.
From an operational point of view, requiring exhaustive traceability of training data on an Internet scale is an illusion. Precisely identifying what content is protected, under what jurisdiction, with what conditions of use, is practically impossible on a large scale. Such a requirement would create a de facto insurmountable technological and legal barrier for the majority of players – particularly smaller ones, open source projects and emerging European initiatives.
This approach also carries a major cultural risk. Faced with high legal uncertainty, model designers would be encouraged to massively exclude certain corpora, certain languages or certain domains. The result would be impoverished, biased models, sometimes blind to entire sections of culture and knowledge. However, major language models are becoming one of the main gateways to information, comparable to a 21st century encyclopedia. Accepting their cultural impoverishment would be a historical error.
Finally, this debate is based on conceptual confusion. Training a language model is neither about storing nor reproducing works, but about learning linguistic structures, styles, regularities – in the same way that a human being learns by reading. Assimilating this learning to a systematic violation of intellectual property amounts to confusing inspiration and reproduction.
The real question is not what AI learns, but what it produces
Recognizing the limits of the current debate does not mean denying the rights of authors. The central question is not what a model learns about, but what it is capable of producing.
Intellectual property must be protected at the time of inference, that is, when the model generates content. If a system reproduces textually, or in a recognizable manner, a protected work, then there is an infringement of intellectual property – and it is legitimate for the rights holders to be remunerated.
This shift in focus makes it possible to address the problem where it really manifests itself: in concrete uses. It also offers a much more operational framework, compatible with the technical realities of generative AI.
Mechanisms already exist to regulate the generation of sensitive content, whether toxic, illegal or regulated. It is entirely possible to apply similar principles to intellectual property, via detection and filtering tools integrated at the time of generation, relying on databases of protected works.
Pay for creation without blocking generative AI
From this principle, several remuneration models can coexist. Model publishers can, for example, choose to subscribe to a license with a dedicated entity – a form of “SACEM for generative AI” – guaranteeing that models do not emit protected content, while ensuring shared remuneration for rights holders.
Failing this, inference detection mechanisms could make it possible to identify the content concerned and trigger remuneration a posteriori, proportionate to actual uses. In both cases, the logic is the same: protect creation where it is actually exploited.
This approach has several decisive advantages. It truly protects intellectual property, because it targets concrete uses rather than learning processes. It preserves the cultural and linguistic diversity of the models. It avoids penalizing European innovation, SMEs and open source, which have neither the legal means nor the financial capacities of the global giants. And above all, it helps establish a healthier relationship between creators and innovators.
The debate on intellectual property should not pit the world of creation against that of technology. It must make it possible to build a balanced framework, in which creation is protected and remunerated, without slowing down innovation or impoverishing our collective access to knowledge.
At a time when Europe is seeking to define its own path in terms of artificial intelligence, it is this type of pragmatic, operational compromise that respects cultural diversity that must guide our choices.
