
On July 11, 2025 at 6 p.m., at the Hôtel de Métropole in Montpellier, we debated Mathieu Lafourcade, of LIRMM and the Université de Montpellier, before the Journal Club IA of Les Halles de l’IA. The question was one that filled meeting rooms at the time: when we delegate to AI, do we lose abilities?
An MIT study had just put a name on that worry, “cognitive debt”. It was widely cited and far less widely read. Its conclusions rest on a specific protocol, a small sample and a single task, which rules out generalizing the results. So the debate was as much about what the study shows as about how it had been used.
Our position came down to one distinction. An employee who asks for an answer and copies it learns nothing, that much is agreed. An employee who asks for three approaches, compares them and argues for a choice exercises more judgment than with a search engine. The same tool produces both behaviors, and what separates them is the instruction the company gives, not the technology.
That is why our training sessions spend time on checking and challenging answers, not on a catalog of features. An executive who rolls out AI without saying what it is for, and where the trust it deserves ends, sets up the first behavior by default.
Thanks to Mathieu Lafourcade for a disagreement held without caricature on either side.