Cognitive scientific explanations can take either a mechanistic or design perspective. Some recent philosophical works propose to apply the mechanistic perspective to the influential mental file framework. The design perspective, however, remains underexplored. This paper takes on this task, arguing that mental files have been designed by natural selection to efficiently represent property cluster structures in the environment—such as objects and kinds—to facilitate our learning about them. It also shows that the design perspective can help resolve a previous debate between two specific accounts of mental files.
Some prominent theorists argue that within a teleosemantic framework, according to which representation is grounded in biological function, organisms are unlikely to evolve the capacity to represent substances per se. Their line of reasoning is that organisms would only evolve to represent surface-level stimuli or affordances, as the capacity for developing substance representations is evolutionarily more “expensive” but practically inconsequential. I argue that this reasoning is flawed. Drawing on research in animal theory of mind and associative learning, I show that many animals, precisely for resource efficiency, tend to approximate causal modeling by positing “intervening” or “latent variables” that mediate sensory stimuli and practical outcomes. I argue that, on a teleosemantic account, these latent variables function as substance representations.