One particularly telling example is Sadin's debate with the French philosopher and host Raphaël Enthoven, staged on Le Figaro's debate program.

Enthoven's central piece of evidence that AI cannot think is an experiment in which he was set the task of writing a philosophical dissertation. His essay received the highest possible mark, 20 out of 20, while ChatGPT scored only 11. The reasoning was that the language model had failed to frame the problem properly. For this media philosopher, the very act of posing a problem is proof enough that AI cannot think.

Yet this claim proves nothing beyond the fact that a language model produced a mediocre essay in 2023. It is no reason to conclude that AI, by its very nature, cannot think. It is rather like beating a computer when the first chess programs appeared, then concluding from that single victory that a machine could never, by definition, play chess better than a human.

In this debate, then, Enthoven mainly plays the reassuring philosopher. Nothing to worry about, he seems to say: the gulf between the human mind and AI remains vast.

Two Languages, Two Kinds of Existence

Sadin takes a different view. His analysis reaches far beyond the narrow definition of the human mind as something merely capable of framing a problem. His argument begins instead with language itself.

In AI, language production is a matter of probabilistic computation. The language model draws on statistical relationships extracted from vast quantities of text and uses them to generate the most probable continuation. Its output, therefore, does not arise from any experience of the world, but from a mathematical model of the patterns embedded in human language.

For Sadin, this stands in sharp contrast to how humans use language. Human speech, he argues, is rooted in freedom: no one knows in advance exactly how they will finish their next sentence, and that freedom is built into language itself. This view, however, might be considered something of an oversimplification. Humans are not entirely indeterminate in their speech, since the very use of words and shared meanings imposes limits of its own.

Even AI would struggle to produce identical text twice. Yet Sadin's analysis of human language goes beyond the notion of a mere random generator. Human freedom, he insists, does not stem from unpredictability alone.

The real difference lies elsewhere. A person does not speak solely on the basis of the words that came before. They speak from a particular place in the world, shaped by their own body, their history, their relationships and the situation in which they currently find themselves.

Their speech carries both stated and unstated intentions, and it draws on a long-accumulated personal store of language, culture and reading. It is, in short, a living language. For this very reason, Sadin does not hesitate to call the language used by AI thanatologos, the "language of death".

The Slow Handover of Human Judgment

Because language and thought are so closely intertwined, a dead language risks reshaping human thought itself. How? Even the earliest models, built not on generative AI but on the analysis of data sets too vast for any human to process, already required people to hand over a measure of their judgment to algorithms.

The rise of generative AI will only accelerate that shift. The danger is not that AI will produce near-perfect text, but that, over time, humans will simply lose their reason to write anything themselves.

Consider the parallel with GPS. People once had to study a route and find their own way. Now many follow GPS blindly and can no longer navigate unfamiliar places without it. Writing, Sadin suggests, may go the same way.

Language, after all, is not only a means of communication but also a means of thought. Over time, people may lose the ability to formulate, structure and develop their own ideas independently. In losing that ability, they lose themselves.

If a child discovers that a system can produce an essay, an argument or an interpretation from a simple prompt, the question follows naturally: why learn to write? Why read long texts? Why bother formulating a thought at all?

Finding meaning was already difficult in the postmodern era. Finding it with the help of AI may prove harder still. Surrounded by an almost limitless supply of information and flawlessly functioning tools, people may increasingly come to rely less on their own abilities.

Sadin therefore rejects the notion that every new capability AI offers must be embraced simply because it raises productivity. Technology, he argues, should reach its limit at the point where it begins to threaten human freedom, integrity, dignity, creativity or personal relationships.

Applying that standard would leave a far narrower role for generative AI than the one it occupies today. Sadin has no objection to its use in research, analysis or information processing. What troubles him is the moment the system stops merely assisting humans and instead takes over the act of writing, creating and forming ideas on their behalf.

The real question, then, is not whether AI will ever begin to think. It is whether we will stop thinking for ourselves.