On the Paris stage at the inaugural VivaTech Bloomberg Awards in June 2026, researcher Yann LeCun pointed out a self-evident truth that the luxury industry sometimes prefers not to voice aloud. Artificial intelligence can pass a bar exam. It still doesn’t know how to pick up a fragile object without breaking it.
The gesture that no one knows how to code
This observation has a name in computer science: Moravec’s paradox. It refers to the gap between what a machine can do effortlessly—solving an equation, drafting a contract, answering any question—and what it still fails to replicate: grasping an object, assessing resistance, or the sensory judgment that guides a movement.
For a senior management team that oversees the balance between technological investment and the preservation of know-how, the stakes are far from trivial. It shifts the focus of the question. It is no longer a matter of whether artificial intelligence should be integrated into the organization, but of determining precisely where its role ends.
What twenty thousand billion words don’t say
The technical reason for this paradox lies in the nature of data. Language is a reduced, quantified description of the world. Touch, sight, and the balance of a movement cannot be written down—they must be experienced. Yet large-scale models can still only learn from what can be put into words.
The Details — A general-purpose language model is trained on nearly all of the digitized text available worldwide, amounting to approximately 20,000 billion words. It would take 400,000 years to read this entire corpus. A four-year-old child, through vision alone, absorbs a comparable amount of sensory information in 48 months.
At Hermès, it takes 18 months to train a saddler. Not so that he can replicate a stitch. It’s so they can feel, in their hands, when the leather still offers resistance and when the needle is pushing beyond what’s reasonable. This knowledge isn’t found in any manual. It’s passed down through repetition, mistakes, and correcting one movement with another.
The boundary that cannot be delegated
The temptation for senior management, enticed by the promise of increased productivity, would be to erase this boundary in the name of speed. The opposite argument—to shield the entire workshop as a matter of principle—would be just as unwise. Artificial intelligence is already effectively handling demand forecasting, content translation, and logistics optimization. No serious fashion house would do without it anymore.
The risk, therefore, is not the use of artificial intelligence. It is the lack of a clear policy regarding its limits. A management team that has not formally defined what should never be delegated unwittingly exposes the intangible heritage it claims to protect: the craftsmanship, the tradition of the trade, and the very legitimacy of the word “craft” on a label.
This is precisely what this paradox illustrates, from an unexpected perspective—that of artificial intelligence research itself, rather than luxury marketing. The machine does not lie about its own limitations. It lays them bare with a rigor that few corporate narratives dare to adopt.
The question facing executive committees now is no longer whether artificial intelligence will make its way into the shop floor. It’s already there. What remains to be decided, company by company, is where it stops—and who within the organization has the authority to draw that line before it is drawn on its own, by default, as budgets are allocated.
Cette publication est également disponible en :
