Computational Theory of Mind: Is Thinking a Form of Computation?

The Core Idea

The computational theory of mind (CTM) rests on a deceptively simple analogy: the mind is to the brain as software is to hardware. Just as a computer program can be described at the level of algorithms and data structures without reference to the specific electronic circuits that implement it, mental processes can be described at the level of representations and computational operations without reference to the specific neurons that carry them out. This level of description, the computational or algorithmic level, is where cognitive science operates.

The theory was developed primarily by the philosopher Hilary Putnam and elaborated extensively by Jerry Fodor. Putnam argued that mental states should be identified not by their physical composition but by their functional role, the pattern of inputs, outputs, and relationships to other states that define them. Fodor built on this foundation by proposing that the mind operates using a language of thought (sometimes called mentalese), an internal symbolic system with its own syntax and semantics. On Fodor's view, thinking is literally a process of manipulating symbols in this internal language according to syntactic rules, just as a computer manipulates symbols in its programming language.

Historical Foundations

The computational theory of mind emerged from the cognitive revolution of the 1950s and 1960s, when researchers in psychology, linguistics, and computer science converged on the idea that mental processes could be understood as information processing. Alan Turing's 1936 formalization of computation provided the mathematical foundation: a Turing machine can compute any function that is computable at all, using nothing more than a finite set of symbols, a set of rules for manipulating those symbols, and a memory tape for storing intermediate results.

If the brain is a kind of computer, then understanding cognition means understanding what the brain computes and how it computes it. Noam Chomsky's work on transformational grammar showed that language could be described as a formal system with precise rules for generating and transforming sentences. Allen Newell and Herbert Simon's Logic Theorist and General Problem Solver demonstrated that computers could prove mathematical theorems and solve puzzles using strategies similar to those reported by human problem solvers. David Marr later articulated three levels at which any information-processing system could be analyzed: the computational level (what function does it compute and why), the algorithmic level (what procedure does it use), and the implementational level (how is the procedure physically realized).

The Language of Thought

Jerry Fodor's language of thought hypothesis (LOTH) is the most developed version of the computational theory of mind. Fodor argued that thinking requires a representational medium with compositional structure, meaning that complex thoughts are built from simpler components combined according to systematic rules. The thought that the cat is on the mat is composed of representations of the cat, the mat, and the spatial relationship on, combined in a way that captures the specific proposition being entertained.

Two key properties of the language of thought are productivity and systematicity. Productivity means that a finite set of symbols and rules can generate an infinite number of distinct thoughts, just as a finite vocabulary and grammar can generate an infinite number of sentences. Systematicity means that the ability to think certain thoughts is intrinsically connected to the ability to think related thoughts: anyone who can think the thought that John loves Mary can also think the thought that Mary loves John. Fodor argued that these properties are best explained by positing a language-like representational system in the mind, because languages are exactly the kind of system that exhibits both productivity and systematicity.

Classical vs Connectionist Computation

The classical version of CTM, associated with Fodor and Pylyshyn, treats cognition as the manipulation of discrete, structured symbols according to explicit rules. This approach was dominant in artificial intelligence from the 1950s through the 1980s and produced systems like expert systems, theorem provers, and planning algorithms. Classical systems excel at logical reasoning, rule following, and systematic inference, but they struggle with pattern recognition, learning from examples, and handling noisy, ambiguous input.

Connectionist models, also called neural networks or parallel distributed processing (PDP) models, offer an alternative computational framework. Developed by David Rumelhart, James McClelland, and others in the 1980s, connectionist models represent information as patterns of activation across networks of simple processing units, inspired by the structure of biological neural networks. Rather than following explicit rules, these networks learn by adjusting the strength of connections between units based on experience, using algorithms like backpropagation.

The debate between classical and connectionist approaches has been one of the most productive in cognitive science. Fodor and Pylyshyn argued that connectionist models cannot explain the systematic and compositional nature of thought without implementing classical symbolic computation at some level. Connectionists countered that their models capture important aspects of cognition that classical models miss, including graceful degradation, content-addressable memory, generalization from examples, and sensitivity to statistical regularities in the environment. Most contemporary cognitive scientists view these approaches as complementary rather than competing, with different computational frameworks being appropriate for different cognitive functions.

Challenges to the Computational Theory

The Chinese Room Argument

John Searle's Chinese Room argument is the most famous objection to the computational theory of mind. Searle imagines a person who does not understand Chinese sitting in a room, receiving Chinese characters through a slot, consulting a rule book to determine which Chinese characters to send back, and passing the responses out through another slot. From outside, the room appears to understand Chinese, passing any conversational test. But Searle argues that clearly no understanding is occurring, the person in the room is just following formal rules without knowing what the symbols mean. If the person does not understand Chinese by virtue of running the right program, then no computer running the right program truly understands anything. Computation, Searle concludes, is not sufficient for understanding or consciousness.

Defenders of CTM have offered several responses. The systems reply argues that while the person does not understand Chinese, the system as a whole (person plus rule book plus room) does. The robot reply argues that connecting the computer to sensors and actuators that interact with the physical world would provide the grounding needed for genuine understanding. The brain simulator reply asks what would happen if the program simulated the activity of every neuron in a Chinese speaker's brain. These responses remain contested, and the Chinese Room continues to generate debate about the relationship between computation and understanding.

Embodied and Situated Cognition

The embodied cognition movement challenges CTM's assumption that cognition is fundamentally abstract symbol manipulation that happens to be implemented in a biological body. Proponents like George Lakoff, Lawrence Barsalou, and Andy Clark argue that cognition is deeply shaped by the body and its interactions with the physical and social environment. Language comprehension activates motor and sensory systems, suggesting that understanding a sentence about kicking involves the same neural circuits used for actually kicking. Spatial reasoning relies on mental imagery that preserves the metric properties of physical space. Emotion and bodily states influence reasoning and decision making in ways that abstract computational models typically do not capture.

The Frame Problem

The frame problem, originally identified in AI research, poses a challenge for classical computational approaches to cognition. The problem is specifying which aspects of a situation change and which remain the same when an action is performed. When you move a cup from one table to another, the cup changes location but its color, weight, and contents remain the same, and so do essentially all other facts about the world. For a classical computer program, explicitly representing what does not change is computationally intractable because the number of unchanged facts is essentially infinite. Humans handle this effortlessly through common sense knowledge and default reasoning, but formalizing this capacity has proven extraordinarily difficult.

CTM and Modern AI

The rise of modern deep learning and large language models has reignited fundamental questions about the computational theory of mind. These systems achieve remarkable performance on language tasks, problem solving, and even creative work, but they operate through mechanisms quite different from classical symbolic computation. They do not manipulate structured symbols according to explicit rules but learn statistical patterns from massive datasets using gradient-based optimization.

This raises the question of whether the computational theory of mind was right in its broad claim that cognition is computation, but wrong in its specific assumption that the relevant computation is symbolic rule-following. If AI systems can achieve human-like cognitive performance through very different computational mechanisms, this suggests that the space of possible computational architectures for cognition may be much larger than classical CTM assumed. The ongoing comparison between human and artificial intelligence continues to refine our understanding of what computation is, what thinking is, and how the two relate.