Open Access

Hermann Müller, Rouwen Cañal-Bruland, Stefan Künzell, Torsten Schubert

• The final stage of a learning process is often described as automatic, meaning that skills are executed in a seemingly effortless manner, fast, and without attentional control. Yet, over the past decades, various theories offered different concepts of what "automatic" means and how automatic and non-automatic processes relate to each other. Likewise, there is ongoing debate about the mechanisms underlying the transition from non-automatic to automatic processes. In this article, we propose a model that specifies how automatization emerges through practice. The central idea of our Multilayer Model of Automatization (MMA) is that cognitive control operates across hierarchically organized processing levels that each display characteristics of both controlled and automatic processing. The superordinate level controls the behavior of the subordinate levels by specifying tasks, determining control policies, and monitoring output, thereby incurring additional costs in terms of time andThe final stage of a learning process is often described as automatic, meaning that skills are executed in a seemingly effortless manner, fast, and without attentional control. Yet, over the past decades, various theories offered different concepts of what "automatic" means and how automatic and non-automatic processes relate to each other. Likewise, there is ongoing debate about the mechanisms underlying the transition from non-automatic to automatic processes. In this article, we propose a model that specifies how automatization emerges through practice. The central idea of our Multilayer Model of Automatization (MMA) is that cognitive control operates across hierarchically organized processing levels that each display characteristics of both controlled and automatic processing. The superordinate level controls the behavior of the subordinate levels by specifying tasks, determining control policies, and monitoring output, thereby incurring additional costs in terms of time and effort. The control units in the MMA are adaptive: they can learn to recognize regularities, anticipate tasks, and generate goal-directed outputs. During this learning process, lower level units gradually become more independent from higher-level control signals, thereby freeing higher-level processing capacities. We specify the necessary and sufficient prerequisites and mechanisms for automatization to be successfully produced by our MMA. Furthermore, we conceptualize automaticity not as a binary, all-or-none phenomenon, but as a gradual or partial property. We can observe the degree of automaticity by the amount of control signals exchanged across control levels. Finally, we explain and discuss how the properties of the MMA correspond to human automatization processes, allowing for specific predictions of automated behavior. The article concludes with an outlook on phenomena and open questions that can be addressed through the MMA framework.…