Simulation of Human Thought
The simulation of human thought is the research program, begun by Allen Newell and Herbert Simon in the mid-1950s, of writing computer programs whose step-by-step processing is meant to reproduce the way people actually solve problems, rather than merely to reach the same answers. It treats a program as a psychological theory: if the program's sequence of intermediate steps matches the sequence a person goes through, the program is a hypothesis about the mental process. This is a stricter goal than building a capable AI system, and it separated the psychological wing of early AI from the engineering wing.
Newell and Simon's method was to compare program traces with what subjects said while working. They collected thinking-aloud protocols, in which participants verbalized their thoughts while solving tasks such as logic derivations, chess positions and cryptarithmetic puzzles in which letters stand for digits, and then compared those transcripts with the trace of a program. Their General Problem Solver, described in a 1959 report by Newell, Shaw and Simon and in Newell and Simon's 1961 paper "GPS, a Program that Simulates Human Thought", used means-ends analysis: compare the current state with the goal, pick the largest difference, and apply an operator that reduces it. On logic problems its subgoal structure resembled that of human subjects, which they argued was evidence that the method modeled human thought.
Human Problem Solving, published in 1972 by Prentice-Hall, was the full statement. Running to more than nine hundred pages, it presented the information-processing theory of problem solving with detailed analyses of protocols from three task domains. The theory pictured a person as an information-processing system with limited short-term memory and a large long-term memory, working within a task environment and searching a problem space of their own construction. Programs were written as production systems, sets of condition-action rules, which became a lasting architecture in psychology and AI. Newell pushed the ambition further in Unified Theories of Cognition (1990), arguing for a single architecture and embodying it in Soar.
John Anderson's ACT family carried the approach into the experimental mainstream. Anderson published the original ACT theory in 1976, ACT* in The Architecture of Cognition (1983), and ACT-R in the 1990s, culminating in the book The Atomic Components of Thought with Christian Lebiere in 1998. ACT-R represents knowledge as declarative chunks and procedural rules, with activation values that govern retrieval, and its modules have been mapped onto brain regions seen in imaging studies. Unlike GPS, it makes quantitative predictions, such as response times and error rates, which are compared with human data, and it has been applied to education in the form of cognitive tutors for mathematics.
The central subtlety is that matching a human's performance is not the same as matching a human's process. Two programs can give identical answers by different routes, so fitting outputs establishes only weak equivalence; the program's internal steps must correspond to the person's for a strong claim, and the evidence for that is the trace comparison. Zenon Pylyshyn's distinction between a program's functional architecture and its content made this precise. The goal also pulls against engineering: a system designed to be human-like inherits human limits, which is why chess programs that win by brute force teach little about expert human play, and why a program that beats people at a task may say nothing about how people do it.
The critiques are substantial. Hubert Dreyfus argued in What Computers Can't Do (1972) that human expertise relies on embodied know-how that rules cannot capture. Richard Nisbett and Timothy Wilson's 1977 paper "Telling More Than We Can Know" documented that people often misreport the causes of their own behavior, which threatens verbal protocols as data, a concern Ericsson and Simon's 1980 paper and 1984 book Protocol Analysis answered by specifying when verbal reports reflect actual processing. Newell himself warned in 1973 that piling up isolated experimental results would not yield a theory of the mind. A further limitation is scope: the fits were strongest for slow, deliberate, puzzle-like tasks, and weaker for perception, language and everyday common sense.
Misconceptions are common. GPS was not a general problem solver in the sense of solving arbitrary problems, since it needed problems described in its own operator-and-difference format; it did not beat human experts; and the thinking-aloud method does not claim to read consciousness directly. The modern echo is the reasoning trace of a language model: such a chain of thought can look like a thinking-aloud protocol, but whether it reflects the underlying computation is exactly the question Nisbett and Wilson raised for people.
Read simulation of human thought as the demand that a program earn the name of a theory not by getting right answers but by taking the same steps a person takes, and by being willing to be wrong about those steps.