Centaur Literature
Am I the first to join these two words, “literature” and “centaur,” in the context of AI at least? The mere fact of asking says something about our time. We are at the frontier. Everything remains to be invented, including the vocabulary to name what is emerging.
Half human, half machine, centaur literature will no doubt be transitional. As the camps take shape—anti-AI versus pro-AI—hybrid uses will grow rarer, or will simply be judged irrelevant, unworthy of visibility. There will be the writers of the sanctuary on one side and, on the other, the disposable writers, fronts for the machine.
Opening up in the meantime is the possibility of a centaur literature. That is Sébastien Bailly’s next novel, if I’ve understood correctly.1 It is also, to a certain extent, Thélyson Orélien’s bestselling novel: even though Pangram judged the book to be almost entirely written by AI,2 it would have been impossible to write C’était ça ou mourir in full just by tossing in a few prompts. At the very least, the novel was surely written scene by scene, and reworked dozens or hundreds of times with prompt after prompt. That isn’t the kind of centaur literature I find very exciting, but it’s part of it.
The more AIs improve, the more the human share needed to write a quality book will shrink, like the wild ass’s skin in Balzac, until a simple prompt of a sentence or two is enough. Then the field of literature as product—the novel and fiction, at least—will be drowned, and centaur literature relegated to the realm of the experimental: unmonetizable explorations, experiments, weavings at the suture of the human and the machine.
Like mules, centaur literature will struggle to reproduce.
I am writing in the future tense, and yet several phenomena are already surfacing. All the same, it is early, it is early and at the same time it is getting late. Late, because superintelligence is at our gates—and with it, the risk of catastrophe or extinction. Early, very early, because we are only at the very beginning of the singularity. If we make it through, if we survive, we will look back on 2026 and see a dawn, a stammering.
Because it is early, so early, and because it is late, very late, we must recover prophetic speech.
Nakhon Ratchasima, Thailand
Forthcoming from Robert Laffont in early 2027. An AI generated some twenty chapters, reworked over six months. According to Bailly, between 2 and 20 percent of the initial text remains. ↩︎
Contrary to what has been said, as things stand, Pangram is very reliable. I was disappointed, in September, to see people in the literary world sharing any old nonsense on their social networks to defend Orélien at all costs: screenshots “proving” that Pangram labels obviously human texts like the Bible, Victor Hugo, etc., as “written by AI.” It was almost always the free version of Pangram or else other detection software, never endorsed by independent researchers, unlike Pangram. See Les Décrypteurs, “Controverse Thélyson Orélien : ce qu’il faut savoir sur Pangram,” Radio-Canada, September 24, 2026; Brian Jabarian and Alex Imas, “Artificial Writing and Automated Detection,” NBER, 2025. ↩︎