Artificial intelligence is now establishing itself as a structural lever for transforming intellectual work. It promises to accelerate processes, enhance human capabilities, and redistribute efforts toward tasks of higher strategic value. However, this promise rests on a rarely questioned assumption: that of cost-free cognitive transfer. Yet recent work in cognitive sciences and neurosciences suggests a more ambivalent reality. AI does not merely assist thought—it modifies the very conditions of its production. The challenge is therefore not simply technological; it is profoundly cognitive and epistemological.
The analogy often drawn with historical tools such as calculators or search engines deserves to be nuanced.
As shown by the work of Christian R. Klein and Reinhard Klein, generative AI models introduce a qualitative break. They automate not only procedural tasks but also cognitive integration processes such as reasoning, synthesis, and argumentation. This capability profoundly transforms the nature of cognitive offloading. Where previous tools delegated limited operations, AI can now substitute for central stages of thought, creating an unprecedented tension between cognitive extension and atrophy.
Yet learning rests on a fundamental constraint: effort.
The models of cognitive psychology, notably those developed by Daniel Kahneman, distinguish between a fast, intuitive system and a slow, analytical system that is costly in resources but indispensable for building lasting knowledge. Generative AI, however, is precisely designed to eliminate this friction. It provides immediate, fluent, and plausible answers, thus favoring what Klein and Klein call “cognitive bypass”—a systematic circumvention of the deliberative processes necessary for constructing cognitive schemas. In this configuration, the answer precedes reflection, creating confusion between access to information and appropriation of knowledge.
The implications of this phenomenon are beginning to be documented empirically.
Neuroscience research points to a reduction in the engagement of cognitive networks when using generative tools, including a decrease in activity associated with cognitive load. Some experimental studies report a significant decline in cortical activity accompanied by impaired memorization, suggesting the emergence of a form of cognitive debt. The brain, less solicited, mobilizes fewer verification, error detection, and memory consolidation processes. Yet neuroplasticity obeys a simple principle: circuits that are not activated weaken. By reducing cognitive effort, AI could thus undermine the very mechanisms of learning.
This impact is not uniform. It depends strongly on the user’s level of expertise.
Recent work highlights an expertise duality whereby AI acts as an equalizing factor for novices and as an amplifier for experts. In simple and structured tasks, inexperienced individuals benefit from rapid performance gains. But as complexity increases, their dependence becomes a handicap, as they lack the knowledge necessary to detect errors or biases. The expert, by contrast, uses AI as a validation and exploration tool, drawing on their knowledge to question responses. AI does not replace expertise; it reveals its strategic importance.
To these cognitive mechanisms is added a psychological and motivational dimension. The concept of the “Sovereignty Trap,” developed by Klein and Klein, describes the tendency of users to delegate their judgment when faced with systems capable of producing credible and coherent answers. This delegation follows a logic of least effort and relies on well-documented biases such as automation bias. AI produces immediate satisfaction by giving the impression of having understood, when the user has merely accessed an answer. This illusion of competence profoundly modifies the relationship to learning by short-circuiting the mechanisms that enable the construction of lasting understanding.
The long-term risk is not a disappearance of cognitive capacities, but a silent transformation of their use.
The state described as a “hollowed mind” corresponds to a form of emptied mind characterized by dependence on external tools and superficiality of knowledge. This phenomenon extends earlier observations related to the “Google effect” or the illusion of knowledge, but now extends to more complex cognitive functions. The individual does not stop thinking, but activates less spontaneously the processes necessary for autonomous thought.
Faced with these developments, the notion of cognitive sovereignty appears as a central issue.
It refers to the capacity to remain the arbiter of one’s intellectual processes, to verify information, and to understand the limits of the tools used. This sovereignty does not exclude the use of AI, but it redefines its conditions. It requires an active posture, vigilance, and the capacity to resist ease. In an environment where access to information is instantaneous, the key competence becomes the ability to evaluate, not merely to access.
In this perspective, the authors introduce the concept of the “fortified mind,” which refers to a cognitive architecture adapted to an environment saturated with artificial intelligence. This mind rests first on a foundation of deep knowledge enabling the detection of inconsistencies and limitations in generated responses. It also requires developed metacognition to reflect on one’s own thought processes. Finally, it assumes a disposition toward intellectual effort allowing the activation of analytical reasoning despite its cognitive cost.
These dimensions are interdependent and condition the capacity to use AI without becoming dependent on it.
The implications of this analysis extend beyond the individual level. They also concern system design. Current interfaces primarily favor fluidity and instantaneity of responses. This logic progressively reduces engagement in the cognitive processes necessary for analysis, verification, and deep learning. An evolution toward systems integrating mechanisms of productive friction could allow the reintroduction of conditions favorable to learning. We should design tools capable of stimulating reflection rather than replacing it, by introducing contradictions, making uncertainties explicit, or requiring active validation.
Artificial intelligence does not destroy human cognition—it redefines its equilibria.
It introduces a tension between efficiency and learning, between access and mastery, between delegation and sovereignty. The central question is therefore not whether AI is useful, but to understand under what conditions it strengthens or weakens our capacity to think. In a world where the answer is immediate, the determining competence becomes the ability to suspend this immediacy to reconstruct reasoning.
This transformation calls for heightened vigilance and opens a field of research that remains largely unexplored. How can we measure the real impact of AI on cognitive capacities in the long term? Can educational systems integrate these issues without sacrificing efficiency? Will AI reduce skill gaps or amplify them? Can we design interfaces that actively promote critical thinking rather than the consumption of answers? These questions now structure a debate that goes beyond technology to question the very foundations of knowledge.
Sources
https://www.sciencedirect.com/science/article/pii/S2666920X26000342
