> ## Content Index
> Fetch the complete content index at: https://ai-frontier.ghost.io/llms.txt
> Use this file to discover other available public pages before exploring further.

# When Machines Began to Think
- URL: https://ai-frontier.ghost.io/road-to-ai-05/
- Published: 2026-09-29T20:40:20.000Z
- Updated: 2026-10-03T21:57:48.000Z
- Description: As programmable computers emerged, researchers began asking a new question: could machines do more than calculate? Article 5 follows the ideas that helped turn machine intelligence into a serious field of research.
- Author: AI Frontier

[The Road to Artificial Intelligence — Part 5 of 15](https://ai-frontier.ghost.io/road-to-ai/)

THE QUESTION CHANGES

The stored-program computer made flexible electronic computation practical. On its own, however, it did not explain how a machine might reason, learn, or adapt.

By the middle of the twentieth century, several intellectual traditions were converging. Mathematicians were formalizing logic. Engineers were studying feedback and communication. Researchers were modeling neural activity, designing chess-playing procedures, and building machines that could manipulate symbols.

The field later called artificial intelligence emerged from this overlap, not along a straight path but at the meeting point of different ideas about minds, machines, information, and problem-solving.

#### Logic inside the nervous system

In 1943, Warren McCulloch and Walter Pitts published "A Logical Calculus of the Ideas Immanent in Nervous Activity." Their paper represented neural activity with idealized units that could switch between states, and it connected those units to propositional logic.

The model was neither a realistic account of the brain nor a modern trainable neural network. Its importance was more abstract: it showed how networks of simple elements might perform logical operations, so that neural activity could be discussed as a form of computation as well as in biological terms. That possibility became important to neural-network research and later theories of machine intelligence. 

#### Control, communication, and feedback

Another tradition developed around Norbert Wiener. In *Cybernetics*, published in 1948, Wiener examined control and communication in animals and machines. Feedback was central: a system could compare its behavior with a goal and adjust its actions accordingly.

A thermostat is a simple example. Instead of acting once, it continually responds to changing conditions. Similar ideas could be applied to mechanical systems, organisms, and organizations.

Cybernetics was broader than what later became symbolic AI, and it did not reduce intelligence to logic alone. Even so, its emphasis on regulation, adaptation, and communication helped establish a vocabulary for thinking about machines that respond to their environments.

#### Information becomes measurable

During the same period, Claude Shannon developed information theory. His 1948 paper provided mathematical tools for describing messages, channels, signals, and noise.

Information theory was not itself a theory of AI. It did establish a powerful way of thinking about communication: information could be represented, transmitted, measured, and processed independently of the physical medium carrying it. A message could be treated as a structured pattern rather than solely as ink on paper, sound in the air, or electrical pulses in a wire. That idea would become important to computing and AI.

Shannon also examined chess as a problem for computers. His 1950 paper, "Programming a Computer for Playing Chess," discussed board representation, move search, and evaluation. It was a design study for a chess-playing procedure, not a report on a completed program. Even so, it illustrated a growing ambition: perhaps machines could perform activities traditionally associated with intelligence by representing problems formally and searching through possible actions.

#### Turing asks what machines might learn

Alan Turing approached machine intelligence from several directions.

His 1948 work "Intelligent Machinery" discussed machine intelligence, unorganized machines, and learning. Two years later, in "Computing Machinery and Intelligence," he introduced the imitation game, later known by the simplified phrase "Turing Test."

The 1950 paper did more than ask whether machines could possess minds. It proposed a practical way to discuss machine behavior, considered objections to the idea that machines could think, and gave substantial attention to learning machines.

This widened the question. A machine need not be programmed with every answer in advance. Perhaps it could be designed to change through experience.

#### A name for a research program

By 1955, several researchers believed these ideas could be brought together in a distinct research program.

John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon prepared "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence." The document proposed work on language, abstraction, problem-solving, learning, and self-improvement.

The proposal contains one of the earliest well-documented uses of “artificial intelligence” as the name of a research program. It should not be described as the first use ever, since McCarthy later recalled choosing the phrase but could not say whether he had heard it before. He is generally credited with introducing it as a field name, although the proposal had four authors. The 1955 document applied the label to a planned program, and the 1956 gathering helped establish it as a shared identity for the developing field.

The choice of language mattered. It gave researchers from different backgrounds a common label, even though they shared neither a single method nor a unified theory of intelligence.

#### Dartmouth was a beginning, but not the beginning

The Dartmouth project took place in the summer of 1956, roughly from June to August. Archival evidence suggests that its duration and attendance were more fluid than the familiar account of a neatly organized six-week conference implies. The gathering was informal, and not every proposed participant attended throughout. Those who took part came from mathematics, computing, psychology, engineering, and information theory.

Dartmouth gave a developing field a shared identity, but it did not create machine intelligence from nothing. Many relevant ideas and experiments predated the meeting, and major AI laboratories and programs later developed at institutions including MIT, RAND, Carnegie, Stanford, and Cornell.

Historian Ronald Kline has emphasized the complicated relationship among cybernetics, automata studies, and AI. Dartmouth was an important moment of naming and organization, not the sole origin of AI.

#### Machines that manipulated symbols

Among the most important demonstrations associated with this early community was the Logic Theory Machine, developed by Allen Newell, Herbert Simon, and J. C. Shaw.

It used heuristic methods to prove mathematical theorems. Instead of exhaustively trying every possibility, it searched for promising paths through a problem, treating reasoning as the manipulation of symbols according to procedures. A 1956 RAND report described its design and research purpose, and a later paper documented its programming and implementation for the JOHNNIAC computer.

Logic Theorist is best understood as one of the earliest successful symbolic AI programs rather than, without qualification, “the first AI program.” Its development came before or around the Dartmouth project, while its demonstration and discussion became part of the emerging AI community.

Its importance extends beyond the particular theorems it addressed. Logic Theorist offered a model of intelligence as search: a machine could represent a problem, generate possible steps, and use heuristics to avoid some unproductive paths.

#### General problem solving

Newell and Simon soon pursued a broader ambition. Their General Problem Solver, described in a 1958 report revised in 1959, investigated intelligent, adaptive, and creative behavior through means–ends analysis and planning. Means–ends analysis compares a current state with a desired goal, identifies the differences, and selects actions meant to reduce them.

It suggested that different tasks might share a common structure even when their subject matter differed. A machine might solve problems not because it understood them as a human does, but because many problems could be expressed in terms of states, goals, operators, and strategies.

This vision became central to symbolic AI: intelligence understood as the manipulation of explicit representations.

#### A different route: neural and adaptive machines

Not every researcher focused on symbols and formal reasoning.

In the summer of 1951, Marvin Minsky built SNARC with the assistance of Dean S. Edmonds Jr. at the Harvard Psychological Laboratory. The neural-analog machine was based on a probability model of reinforcement. It was not a digital computer in the modern sense; its design explored how networks of components might produce adaptive behavior through reinforcement.

Frank Rosenblatt later developed the perceptron, a trainable model inspired by the organization of the brain. He described its principles in 1957 and 1958, and by 1958 the Mark I Perceptron had been built at Cornell Aeronautical Laboratory.

The perceptron demonstrated an influential approach: connections between processing units could be adjusted so that a system learned to recognize patterns. It was not a general reasoning machine, but it embodied a conception of intelligence different from the symbolic programs of Newell and Simon. Symbolic systems emphasized explicit rules and representations. Perceptrons emphasized patterns, connections, and adaptation.

Both traditions would later influence AI, though their fortunes rose and fell at different times.

#### Lisp and the language of AI research

Beginning in 1958, John McCarthy led the development of Lisp (a list processor) at MIT, with contributions from other researchers involved in its design and implementation. Its central data structure, the list, made it well suited to manipulating symbolic expressions, and researchers could use the same language to represent both programs and the structures those programs operated on.

Lisp became closely associated with symbolic AI research. It let investigators write programs that handled rules, expressions, and changing problem structures with flexibility.

Programming languages shape what researchers can readily express, and Lisp's support for symbolic manipulation made it easier to explore theories of reasoning as symbol processing.

### The connection to AI

The early history of AI brought together several ideas that would continue to influence the field:

- McCulloch and Pitts showed how networks of simple units could be analyzed computationally.
- Wiener emphasized feedback, control, and adaptation.
- Shannon provided mathematical tools for representing information and communication.
- Turing framed machine intelligence as a question involving behavior, learning, and computation.
- Logic Theorist and GPS explored reasoning through symbolic representation and search.
- SNARC and the perceptron investigated adaptive behavior through neural-inspired mechanisms.
- Lisp helped researchers build programs that manipulated symbols and structures.

These were not stages in a predetermined march toward today's AI. They were different answers to different questions. Could intelligence be expressed as logic? Could it emerge from feedback? Could machines learn? Could reasoning be represented as search? Could networks of simple elements recognize patterns?

Modern AI draws on ideas from several of these traditions, but many early expectations did not unfold as researchers anticipated.

### Looking ahead

With a field name, research programs, symbolic methods, neural-inspired models, and specialized programming tools in place, researchers began building more ambitious working systems—and expectations rose with them.

The first AI boom would carry these ideas into laboratories, miniature worlds, language demonstrations, robots, and scientific applications. It would reveal both what early AI could accomplish and how easily success in bounded settings could appear more general than it was.

## SOURCES

1. Warren S. McCulloch and Walter Pitts, "A Logical Calculus of the Ideas Immanent in Nervous Activity," *Bulletin of Mathematical Biophysics*, 1943\. [https://doi.org/10.1007/BF02478259](https://doi.org/10.1007/BF02478259?ref=ai-frontier.ghost.io)
2. Norbert Wiener, *Cybernetics: Or Control and Communication in the Animal and the Machine*, 1948\. MIT Press.
3. Claude E. Shannon, "A Mathematical Theory of Communication," *Bell System Technical Journal*, 1948\. [https://doi.org/10.1002/j.1538-7305.1948.tb01338.x](https://doi.org/10.1002/j.1538-7305.1948.tb01338.x?ref=ai-frontier.ghost.io)
4. Alan M. Turing, "Computing Machinery and Intelligence," *Mind*, 1950\. [https://doi.org/10.1093/mind/LIX.236.433](https://doi.org/10.1093/mind/LIX.236.433?ref=ai-frontier.ghost.io)
5. Alan M. Turing, "Intelligent Machinery," Turing Digital Archive, 1948\. [https://turingarchive.kings.cam.ac.uk/intelligent-machinery](https://turingarchive.kings.cam.ac.uk/intelligent-machinery?ref=ai-frontier.ghost.io)
6. Claude E. Shannon, "Programming a Computer for Playing Chess," 1950\. [https://doi.org/10.1080/14786445008521796](https://doi.org/10.1080/14786445008521796?ref=ai-frontier.ghost.io)
7. John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon, "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence," 1955\. Computer History Museum archive.
8. Allen Newell and Herbert A. Simon, "The Logic Theory Machine: A Complex Information Processing System," RAND P-868, 1956.
9. Allen Newell and J. C. Shaw, "Programming the Logic Theory Machine," 1957\. [https://doi.org/10.1145/1455567.1455606](https://doi.org/10.1145/1455567.1455606?ref=ai-frontier.ghost.io)
10. Allen Newell, J. C. Shaw, and Herbert A. Simon, "Report on a General Problem-Solving Program," RAND P-1584, 1958/1959.
11. Frank Rosenblatt, "The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain," *Psychological Review*, 1958\. [https://doi.org/10.1037/h0042519](https://doi.org/10.1037/h0042519?ref=ai-frontier.ghost.io)
12. Marvin Minsky, "A Neural-Analogue Calculator Based upon a Probability Model of Reinforcement," Harvard University Psychological Laboratories, January 8, 1952\. [https://web.mit.edu/dxh/www/marvin/web.media.mit.edu/\~minsky/Bibliography.html](https://web.mit.edu/dxh/www/marvin/web.media.mit.edu/~minsky/Bibliography.html?ref=ai-frontier.ghost.io)
13. Marvin Minsky, *Neural Nets and the Brain Model Problem*, Ph.D. dissertation, Princeton University, 1954, section 4/5.2\. The dissertation identifies SNARC as constructed by Minsky "with the very generous assistance of Mr. D. S. Edmonds, Jr." at the Harvard Psychological Laboratory in the summer of 1951\. [https://people.csail.mit.edu/brooks/idocs/Marvin\_Minsky\_PhD\_Princeton\_1953\_1954.pdf](https://people.csail.mit.edu/brooks/idocs/Marvin%5FMinsky%5FPhD%5FPrinceton%5F1953%5F1954.pdf?ref=ai-frontier.ghost.io)
14. Ronald R. Kline, "Cybernetics, Automata Studies, and the Dartmouth Conference on Artificial Intelligence," *IEEE Annals of the History of Computing*, 2011\. [https://doi.org/10.1109/MAHC.2010.44](https://doi.org/10.1109/MAHC.2010.44?ref=ai-frontier.ghost.io)
15. Ray Solomonoff, "The Summer of 1956" and Dartmouth notebook materials, 1956\. [https://raysolomonoff.com/dartmouth/notebook/notebook.html](https://raysolomonoff.com/dartmouth/notebook/notebook.html?ref=ai-frontier.ghost.io)

← Previous in the series

[Part 4 — From Symbols to Stored Programs](https://ai-frontier.ghost.io/road-to-ai-04/)

Continue the series →

Part 6 — The First AI Boom  
Coming October 1