Invention of Artificial Intelligence: Turing, Dartmouth and the Birth of AI

Origins Evidence Timeline

How Artificial Intelligence Became a Field

Trace the ideas, name, meeting, and working programs that turned machine intelligence into an organized research field.

1 of 7 stages explored

All
Foundations
Definition
Field
Programs
Next milestone

Computability foundation

A General Programmable Machine

Current stage

Formal problem

Alan Turing’s 1936 work described abstract machines that follow symbol-processing rules and introduced a universal machine able to imitate other such machines.

What it enabled

A single programmable computer could, in principle, carry out many different procedures rather than being physically rebuilt for every new task.

What it was not

The paper was about computability, not a working intelligence program. Its relevance to AI came from making programmable symbolic activity mathematically precise.

Before intelligence could be engineered, researchers needed a clear account of what programmable machines could calculate.

Dates mark different kinds of evidence—papers, proposals, meetings, and programs—so no single year captures every meaning of AI’s invention.

Artificial intelligence was not invented by one person or completed on one date. Alan Turing supplied the theory of general computation and made machine intelligence an observable research question. John McCarthy coined the field’s name in a 1955 proposal. The Dartmouth summer project of 1956 gave that work a shared academic identity, while Logic Theorist and Arthur Samuel’s checkers program showed that reasoning and learning could be implemented in software.

Question	Most accurate answer
Who laid the theoretical groundwork?	Alan Turing, alongside earlier work in logic, computation, cybernetics, and mathematical models of neurons
Who coined “artificial intelligence”?	John McCarthy, in the 1955 Dartmouth project proposal written with Marvin Minsky, Nathaniel Rochester, and Claude Shannon
When did AI become a named field?	The name appeared in 1955; the Dartmouth summer project took place in 1956
What is often called the first AI program?	Logic Theorist, developed by Allen Newell, Herbert A. Simon, and Cliff Shaw in 1955–1956
What was an early learning program?	Arthur Samuel’s checkers system at IBM, developed through the 1950s
Was AI invented at Dartmouth in one session?	No. Dartmouth organized a field built from earlier ideas and ongoing experiments

Artificial Intelligence Was a Field Before It Was a Product

The invention of a telephone, engine, or transistor can be tied to a device whose components, demonstration, and manufacturing history can be compared. Artificial intelligence has no equivalent object. There was no first AI machine with a settled design, no single patent defining the field, and no immediate consumer product that marked mass adoption.

Its origin is better understood as a sequence of different achievements. Mathematicians first described general computation. Researchers then modeled neurons, search, learning, and symbolic reasoning. Turing proposed ways to discuss and test machine intelligence. McCarthy selected a new name. Dartmouth gathered researchers around a common agenda. Early programs then supplied evidence that machines could perform tasks associated with reasoning and adaptation.

Layer of origin	Historical evidence	What it established
Computational possibility	Turing’s 1936 work on computable procedures and universal machines	One programmable machine could imitate many rule-following machines
Neural computation	McCulloch and Pitts, 1943	Simplified neuron networks could be analyzed as logical systems
Machine-intelligence question	Turing’s 1948 report and 1950 paper	Intelligent behavior and machine learning could be treated as research problems
Field name	Dartmouth proposal, 1955	“Artificial intelligence” became the label for a planned research program
Research community	Dartmouth summer project, 1956	Separate lines of work gained a common identity
Working demonstrations	Logic Theorist and Samuel’s checkers work	Programs could search, reason with symbols, and improve from experience

This layered view resolves an apparent contradiction in popular accounts. Turing can be a founder without having used the term “artificial intelligence.” McCarthy can name and organize the field without having built the earliest program. Dartmouth can be the field’s birthplace without being the place where every idea first appeared.

Before Artificial Intelligence Had a Name

Turing’s Universal Machine Made General Computation Thinkable

In 1936, Alan Turing published work on computable numbers that described abstract machines reading and writing symbols according to explicit rules. The most far-reaching part was the universal machine: one machine capable of imitating any other machine in the same formal class when supplied with the proper description.

The paper did not propose an intelligent computer. Its later connection to AI rests on programmability. A computer did not need to be a different physical mechanism for arithmetic, chess, language, or theorem proving. Different symbolic procedures could run on the same general machine. That made intelligence-related tasks conceivable as programs rather than as separate mechanical inventions.

McCulloch and Pitts Connected Neural Activity with Logic

Warren McCulloch and Walter Pitts approached the problem from another direction in 1943. They modeled simplified neurons as all-or-none units that receive inputs and activate under specified conditions. Networks of these units could represent logical relations.

The model did not contain the layered training methods, large datasets, or adjustable networks associated with present-day deep learning. Nor was it intended as a faithful map of every biological process in a real neuron. Its contribution was narrower and more durable: nervous activity could be represented in a mathematical system that performed computation.

Chess Exposed the Search Problem

Claude Shannon’s 1950 paper on programming a computer to play chess explained why a rule-governed game could still be difficult. Legal moves are easy to define, but the tree of possible continuations expands too rapidly for complete examination. A useful program therefore needed to represent a position, generate candidate moves, look ahead to a limited depth, and evaluate the resulting positions.

Chess became more than a public contest between humans and machines. It offered a controlled setting for studying choice under limited time and memory. The same problem appeared in theorem proving and planning: a program may know every legal step yet still need a method for deciding which steps are worth exploring.

Turing Changed the Question from Mind to Behavior

Turing’s place in the invention of artificial intelligence rests on more than the test that later carried his name. In his 1948 report Intelligent Machinery, written at Britain’s National Physical Laboratory, he examined whether machinery could show intelligent behavior. He discussed search, learning, machine organization, and systems whose internal connections could be altered through training.

Two years later, Computing Machinery and Intelligence opened with the question “Can machines think?” Turing argued that ordinary definitions of “machine” and “think” would pull the discussion into endless verbal disagreement. He replaced the question with the imitation game, in which an interrogator communicates through text and attempts to distinguish a machine from a human participant.

This distinction matters because the “Turing Test” is often described as a universal meter of intelligence. Turing’s setup concerned performance in a text-mediated exchange. It excluded appearance, voice, skin, movement, and other physical features so the judgment would focus on the intellectual content of the interaction. Success would show persuasive conversational behavior under the test conditions, not settle every question about intelligence.

The Child Machine Was a Learning Proposal

Turing also rejected the idea that every mature behavior had to be written directly into a finished program. He proposed beginning with something closer to a child program and then applying an education process. The designer would vary the program, teach it, observe how it learned, and compare the result with other versions.

He discussed reward and punishment signals, symbolic instruction, experimentation, and the possibility that a trained machine might surprise its designer. This did not amount to a modern training algorithm, but it placed learning near the center of machine intelligence before “machine learning” became an established research label.

The 1955 Proposal That Named Artificial Intelligence

John McCarthy wanted a label broad enough to include learning, language, abstraction, neural models, games, and mathematical reasoning. Existing terms carried narrower associations. “Automata studies” pointed toward formal machines, while “cybernetics” was closely tied to feedback, control, communication, and the intellectual circle around Norbert Wiener. “Artificial intelligence” announced a separate and deliberately broad research subject.

The phrase appeared in a proposal dated August 31, 1955 for a summer study at Dartmouth College in Hanover, New Hampshire. The document was written by McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon. It proposed a two-month study by about ten researchers during the summer of 1956.

Proposal author	Position represented in the project	Relevant experience
John McCarthy	Organizer and Dartmouth mathematician	Mathematical logic, computation, and the effort to define a distinct research field
Marvin Minsky	Researcher in mathematics and neurology	Learning machines, neural networks, and models of the brain
Nathaniel Rochester	IBM computer engineer and research manager	Large-scale digital computer design, including work connected with the IBM 701
Claude Shannon	Bell Labs mathematician	Information theory, switching logic, automata, learning machines, and computer chess

The proposal’s central conjecture was unusually direct: aspects of learning and intelligence could, in principle, be described precisely enough for a machine to simulate them. This was not presented as a result already achieved. It was the assumption around which the proposed research would be organized.

The Original Research Agenda Still Looks Familiar

The proposal did not define AI as one algorithm. It listed a set of unresolved problems whose connections were not yet clear. Several remain recognizable in present research, although the methods and scale have changed.

Problem in the 1955 proposal	What researchers wanted	Present-day descendant	Question still open
Language	Machines able to use language for communication and problem solving	Language models, translation, speech systems, and question answering	How fluency relates to meaning, truth, and grounded knowledge
Abstraction	Programs able to form concepts rather than handle only isolated cases	Learned representations, classification, and concept discovery	How systems can generalize reliably outside familiar examples
Self-improvement	Machines that become better at a task through their own activity	Machine learning, reinforcement learning, and self-play	How improvement can remain controllable and aligned with the intended task
Neuron networks	Models that use connected units inspired by nervous systems	Artificial neural networks and deep learning	How computational models relate to biological intelligence
Creativity and randomness	Methods that produce useful novelty rather than repeating fixed sequences	Generative models, stochastic search, and evolutionary computation	How novelty, intention, authorship, and evaluation should be separated
Complex calculations	More efficient ways to direct computers toward difficult reasoning tasks	Heuristic search, planning, theorem proving, and tool-assisted reasoning	How to avoid explosive growth in the number of possible actions

The continuing relevance of these topics does not mean the 1955 authors predicted present-day systems in detail. They identified problem classes. Modern language models, deep neural networks, large training collections, specialized processors, and networked computing arose through later research that the proposal did not specify.

What Really Happened at Dartmouth in 1956?

The Dartmouth Summer Research Project on Artificial Intelligence is often compressed into the image of a conference where a fixed group met, agreed on a definition, and launched AI. The actual event was less orderly. The proposal described a two-month working study. Participants came for different periods, seminars mixed with informal discussion, and the people present changed during the summer.

McCarthy, Minsky, Shannon, Rochester, Ray Solomonoff, Oliver Selfridge, Allen Newell, Herbert Simon, Arthur Samuel, Trenchard More, Julian Bigelow, and several other visitors are connected with the event in surviving records. They were not all present at once, and historical reconstructions differ depending on whether they count planned attendees, sustained participants, short visits, or everyone appearing in notes and photographs.

It Did Not Produce One Finished Invention

Dartmouth did not end with a single machine, standard test, agreed definition of intelligence, or unified technical method. The researchers differed on how intelligence should be modeled. Some emphasized symbols and logical reasoning. Others examined neural models, pattern recognition, learning, games, or probabilistic induction.

The meeting’s effect was organizational. It gave a name to a collection of problems, brought researchers into direct contact, and helped establish AI as work that could support laboratories, graduate research, conferences, journals, programming languages, and dedicated funding. The field’s birth was institutional and intellectual rather than mechanical.

Logic Theorist Turned Reasoning into a Program

While the Dartmouth project was giving the field a name and research identity, Allen Newell, Herbert A. Simon, and Cliff Shaw were developing Logic Theorist. Their program worked on theorems from symbolic logic, representing statements and permissible operations in a form a computer could manipulate.

The main obstacle was not whether a legal proof step could be generated. It was the number of possible steps. Blindly exploring every branch would waste time on paths unlikely to reach the theorem. Logic Theorist used heuristics—practical selection rules—to direct attention toward more promising branches.

1. Represent the goalThe target theorem and available logical expressions were encoded as symbolic structures.
2. Generate permitted movesThe program applied valid transformation and inference rules to create possible next steps.
3. Rank the branchesHeuristics favored paths that appeared more likely to reduce the distance between the current expression and the goal.
4. Discard unproductive pathsSearch was narrowed so the computer did not spend equal effort on every legal continuation.
5. Assemble the proofWhen a branch reached the target, the sequence of transformations formed a machine-generated proof.

This method gave early AI one of its most durable ideas: intelligence can sometimes be modeled as search through a space of possible actions, guided by information that makes some paths more useful than others. The same general difficulty appears in games, planning, diagnosis, and automated reasoning.

Arthur Samuel Added Learning from Experience

Logic Theorist showed how a program could search through symbolic possibilities. Arthur Samuel’s checkers work at IBM demonstrated a related but different idea: a program’s later choices could improve by using information gathered from earlier play.

The IBM 701 did not have enough memory or speed to examine every possible checkers continuation. Samuel’s program looked ahead, evaluated board positions, and limited the branches it searched. It also recorded information about positions and whether they led toward wins or losses. Those stored values influenced future decisions.

Samuel’s 1959 paper, Some Studies in Machine Learning Using the Game of Checkers, helped establish “machine learning” as a name for programs that improve with experience. The system was still confined to a rule-defined board game. It did not learn arbitrary tasks, infer broad common sense, or redesign its own goals. Its importance lies in proving that all useful behavior did not have to be supplied as a fixed list of moves.

Who Invented Artificial Intelligence?

No single name answers every version of the question. The most accurate account assigns credit according to the contribution being discussed.

Person or team	Documented contribution	Claim to avoid
Alan Turing	General computation, intelligent machinery, the imitation game, and learning-machine proposals	That he coined “artificial intelligence” or attended Dartmouth
John McCarthy	Coined the field’s name, organized Dartmouth, and later created LISP for symbolic computing	That he alone invented every AI method or built the first intelligent machine
Marvin Minsky	Co-authored the proposal and brought early work on learning machines and neural models	That Dartmouth or AI was his work alone
Nathaniel Rochester	Connected the project with practical digital-computer engineering and IBM research	That the event was only a meeting of abstract theorists
Claude Shannon	Information theory, switching logic, automata studies, and computer-chess search	That he supplied a finished chess-playing AI in his 1950 paper
Newell, Simon, and Shaw	Built Logic Theorist, an early working symbolic-reasoning program	That they coined the term or organized the Dartmouth proposal
Arthur Samuel	Developed checkers programs that improved through stored experience and helped name machine learning	That his system could learn without the rules and evaluation structure designed by people

Calling Turing the “father of AI” can be a useful shorthand for his intellectual influence, while calling McCarthy the “father of AI” usually refers to naming and organizing the field. Neither phrase should replace the more exact history. The work depended on mathematicians, engineers, psychologists, neurophysiologists, programmers, and research institutions whose contributions answered different parts of the problem.

Symbolic and Neural AI Were Present Near the Beginning

AI history is sometimes told as a clean succession: symbolic reasoning came first, failed, and was replaced by neural networks. The origin period does not support that simple sequence. Both routes were already visible before and during Dartmouth.

Symbolic route	Neural and adaptive route
Represents objects, propositions, rules, and goals explicitly	Represents behavior through connected units, adjustable values, or experience
Uses logic, search, planning, and rule application	Uses activation, association, feedback, and learned evaluation
Early example: Logic Theorist	Early examples: McCulloch–Pitts networks and Samuel’s checkers learning
Strength: reasoning steps may be inspected	Strength: behavior can improve without listing every case
Difficulty: hand-built knowledge and search can become unmanageable	Difficulty: learned behavior may be hard to explain and may fail outside training conditions

These traditions changed over time and should not be treated as direct copies of their 1940s and 1950s forms. Modern neural networks differ sharply from McCulloch and Pitts’ logical model. Modern reasoning systems also differ from Logic Theorist. The older distinction remains useful because it captures two recurring questions: should intelligence be written as explicit symbolic operations, or learned from examples and feedback?

What the Founders Underestimated

The Dartmouth proposal assumed that a carefully chosen group could make major advances during one summer. That optimism was understandable. Early computers had begun to handle tasks—symbolic calculation, games, proof search—that were strongly associated with human thought. Yet success in a bounded task did not reveal how much knowledge ordinary human activity quietly depends on.

More computing speed addressed only part of the difficulty. A useful AI system also needed suitable representations, learning methods, data, memory, evaluation standards, and ways to cope with uncertainty. Language was especially resistant because words draw meaning from situations, social knowledge, unstated assumptions, and the speaker’s purpose.

The founders also underestimated how hard it would be to move from narrow competence to reliable general behavior. A program could prove theorems without understanding everyday objects. It could play checkers without recognizing a board in a photograph. It could follow a formal goal without knowing whether that goal made sense outside the program.

What the Dartmouth Agenda Got Right

The 1955 proposal nevertheless identified research problems that remained active for seven decades. Language, abstraction, learning, neural models, search efficiency, self-improvement, and machine creativity were not side topics. They became recurring centers of AI research.

The proposal also treated intelligence as a set of capacities rather than a mysterious all-or-nothing property. Researchers could work on theorem proving without solving vision, or on learning without first constructing a complete human-like mind. That division made experiments possible, though it also encouraged systems that performed impressively inside narrow boundaries.

Its strongest surviving assumption is methodological: a mental capacity may be studied by attempting to describe the information, operations, and feedback needed to reproduce part of it. The attempt can succeed, fail, or reveal that the original description of the human ability was incomplete. In each case, the program becomes an experiment about both computation and intelligence.

From Dartmouth to a Lasting Research Infrastructure

A named field needed more than a summer meeting. Researchers required programming languages suited to symbolic structures, access to expensive computers, laboratories where long projects could continue, and institutions willing to fund work whose practical payoff was uncertain.

McCarthy began developing LISP at MIT in 1958. Its list-based representation allowed programs and data to be handled in closely related forms, making it well suited to symbolic AI. AI groups formed at institutions including MIT, Carnegie Mellon, Stanford, IBM, and other research centers. The subject developed its own conferences, laboratories, textbooks, and professional networks.

Progress did not follow a steady line. Periods of ambitious promises were followed by disappointment when systems failed to transfer from demonstrations to messy real settings. Funding and attention contracted during the episodes later called AI winters. New methods, larger datasets, faster hardware, and better engineering repeatedly reopened problems that earlier systems could not handle.

AI at Seventy: Dartmouth’s 2026 Anniversary

Dartmouth is marking 2026 as the seventieth anniversary of the summer project. Its published anniversary schedule lists the Dartmouth AI Conference in Mountain View on October 16, 2026, followed by AI @70: What Must Remain Human? in Hanover from October 29 through October 31.

The wording of the Hanover event shows how the central question has shifted. The 1955 proposal asked which features of intelligence could be described and simulated. The 2026 program focuses on where AI should assist, where human judgment should lead, and who remains accountable when systems fail. Those questions do not replace technical research; they arise because systems can now perform tasks that were only proposals in the Dartmouth document.

Seventy years later, several original problems remain unsettled. Fluent language does not automatically establish truthful understanding. Improved task performance does not explain consciousness. A system may find an effective strategy without producing a human-readable reason. The invention of AI opened an engineering field, but it did not settle what intelligence is.

Artificial Intelligence Was Invented in Layers

The most defensible history does not force one inventor into every role. Turing made general computation and machine intelligence precise enough to investigate. McCulloch and Pitts showed how neural activity could be represented as logic. Shannon exposed the search problem in games. McCarthy supplied the name. Minsky, Rochester, and Shannon helped shape the founding proposal. Dartmouth created a research identity. Newell, Simon, and Shaw demonstrated symbolic reasoning in software. Samuel showed that stored experience could improve play.

AI was therefore not born when a machine suddenly became intelligent. It emerged when researchers began treating parts of intelligence as processes that could be represented, tested, programmed, and improved. The transition from question to field occurred across papers, meetings, machines, and programs rather than at one ceremonial instant.

Questions People Ask About the Invention of Artificial Intelligence

Who is the inventor of artificial intelligence?

There is no sole inventor. Alan Turing is associated with the theoretical and behavioral foundations, while John McCarthy coined the term and organized the Dartmouth project that established AI as a named field.

When was artificial intelligence invented?

1956 is commonly used because of the Dartmouth summer project. A more exact timeline separates the 1955 naming proposal, the 1956 meeting, and earlier work dating from the 1930s through the early 1950s.

Did Alan Turing invent AI?

Turing did not invent the entire field or its name. He provided several ideas that made it possible: universal computation, intelligent machinery, the imitation game, and the proposal to train a child-like program.

Who coined the term “artificial intelligence”?

John McCarthy coined the term for the Dartmouth project proposal written in 1955 with Marvin Minsky, Nathaniel Rochester, and Claude Shannon.

What was the first artificial intelligence program?

Logic Theorist is widely given that title because it was deliberately built to perform symbolic reasoning. Earlier game-playing programs complicate any absolute priority claim, so the result depends on how “AI program” is defined.

Was the Turing Test used at Dartmouth?

No evidence makes the imitation game the operating test of the Dartmouth project. Turing’s 1950 paper shaped the wider discussion, but the proposal covered a broader set of topics including language, neural networks, abstraction, learning, and creativity.