The Illusion of Cognition: Why Behavioural Imitation Cannot Achieve Genuine Cognition

Contemporary artificial intelligence systems increasingly exhibit behaviours associated with perception, reasoning, learning, and goal-directed action. These successes have encouraged the assumption that continued scaling of existing architectures may eventually yield genuine cognition. This paper challenges that assumption by treating cognition as an organisational property that requires, but is not reducible to, behavioural success. We define cognition as an organisationally closed system of mutually implicative capacities. Based on a historical reconstruction of the cognitive vocabulary instantiated or attributed to AI systems across paradigms, we derive eight interdependent cognitive capacity groups and formalise their dependency relations. Evaluating contemporary AI systems against this organisational criterion, we show that despite extensive behavioural competence, none instantiate the closure required for cognition. We then explain why this structural absence nevertheless produces a persistent illusion of cognition. Contemporary AI systems are optimised almost exclusively for behavioural imitation: loss functions evaluate output adequacy without constraining internal organisation. Any structure sufficient to reproduce target behaviour is equally acceptable, while organisational features required for cognition—endogenous goals, internal epistemic access, and reciprocal functional integration—are systematically excluded. Scaling reduces behavioural error but leaves these absences intact, thereby intensifying the appearance of cognition without approaching it. Finally, we discuss the implications of this diagnosis. Our analysis suggests a structural bifurcation: continued scaling of behaviourally optimised systems preserves performance while precluding cognition, or pursuing AI architectures that target organisational closure beyond mere behavioural success is necessary to preserve the possibility of AGI.