The CEM has recently embarked on a major project to create an online resource centre to bring together studies, analyses and best practices that will enable everyone to learn about the positive and measurable impacts of music when it is placed at the heart of society, with particular reference to each of the 17 Sustainable Development Goals in the United Nations' Agenda 2030.
We wanted to kick off the ‘Analyses’ section of the resource centre with a contribution from François Pachet on the subject of music and artificial intelligence, in the spirit of his speech at our event in Mafra on 30 April.

Ode to singularity

F. Pachet, artificial intelligence researcher

"AI feeds on our work, on the work of artists, it steals from them.”

This phrase—or its many variations—has become a refrain among ardent defenders of copyright. Yet it is false, in just about every possible sense. But to understand this, we must first look at what these systems actually do—and listen, too, to what these generative AIs tell us about ourselves.

Those who take the time to delve into the inner workings of the underlying algorithms, these famous deep neural networks, notice one simple thing: these machines do not copy works that are provided to them, they detect similarities between these works. They learn to effectively represent what comes back, what is repeated, what is similar.

Consider the case of sequences—whether textual, musical, or visual. If you feed an AI radically different sequences, with no commonalities, the result is clear: the algorithm learns nothing. It sees only uninterpretable chaos, noise. Of course, even in works that seem very distinct to us, it can detect regularities that escape us: syntactic turns of phrase, harmonic structures, recurring sequences. But it will only do so if these regularities are actually present.

Above all, what it retains, what it encodes, are not the works themselves. It is these regularities, these patterns. AI does not preserve the style of an author, nor the voice of a singer, nor the touch of a painter: it learns to produce “more of the same,” as Andrej Karpathy says in his famous educational videos. It systematizes. It extends. It stretches what has already been done.

Therefore, the real issue is not what AI steals, but what we produce. And it's clear that we produce a lot of self-similar content. A lot of "yet another song," "yet another article," "yet another series," as French writer Céline lamented in a famous interview from the 1950s. We live in a society saturated with similar content, and this is precisely what makes AI so effective: because it trains on material where originality is rare, and repetition is abundant.

Imagine a dataset containing a single book by Pascal, a treatise by Descartes, a play by Shakespeare, a song by Brassens, a poem by Lamartine, a speech by De Gaulle. It's likely that the AI wouldn't get much out of it. The learning curve would be meager, and the result unconvincing. That's a great thesis topic, by the way: training an AI on a deliberately heterogeneous, fundamentally singular corpus—and observing the disaster at the end.

In music, it's estimated that over 100 million tracks are available on streaming platforms, with nearly 100,000 new tracks released every day. And diversity, according to all available metrics, is declining (see [Pachet, 2024]). Standardization is a reality, and it can be seen as objective.

Faced with this, “opting out”, i.e. wanting to remove their works from training sets, as proposed by the European AI Act, is both futile and illusory. There will always be thousands of other works containing the same structures, the same motifs, the same ideas—rehashed, recycled, reproduced. The problem isn't that AI is copying us; it's that we are copying ourselves.

So, what does AI tell us? Perhaps this: if you want to be impossible to imitate, be truly original. Create what no one knows how to do. Be dissonant. Singular. Astonishing. What is new is, by definition, difficult to predict. What breaks habits resists automation. Of course, if your creation becomes popular, it will be imitated, it will in turn enter the great whole digestible by machines. This is the price of diffusion, of popularity, and we must accept it: we cannot have success, paternity, and refuse filiations. Or else we must do what Kafka advocates: stay at our table, wait for the world to come to us. But this injunction, however beautiful, does not long resist the irrepressible desire for popularity.

Our society is not made up of brilliant, isolated individuals, each producing a sacred work worthy of eternal protection. We are, above all, a network of individuals driven by desires (sometimes mimetic), shared influences, and an irrepressible tendency to imitation. This must be acknowledged—and perhaps even celebrated.

AI can be seen as a punitive force, a pagan curse, revealing the banality of our repetitive productions. But it can also be heard as a call: dare to think outside the box. Stop repeating. Embody the new. This is the only way, today, to make AI obsolete—or at least to leave it behind, unable to keep up.