What Is the Origin of Artificial Intelligence?

By Victor Hoang · Updated Oct 21, 2026

In 1955 the people who would found the field were not working in one place. John McCarthy was at Dartmouth and Marvin Minsky was at Harvard. Nathaniel Rochester was at International Business Machines (IBM) and Claude Shannon was at Bell Telephone Laboratories.

Artificial intelligence as a named field began on August 31, 1955, when four researchers sent a proposal for a summer study at Dartmouth College. The study ran in the summer of 1956. The idea underneath it is far older. Poets imagined mechanical helpers more than two thousand years ago. Philosophers spent the 1600s asking whether reasoning could be reduced to calculation. The throughline: people imagined thinking machines for centuries, wrote the theory in the 1930s and 1940s, and only then built the first programs.

Allen Newell, Herbert Simon and Cliff Shaw were at the RAND Corporation in Santa Monica, building a program that would prove theorems. Each group worked on a piece of the same puzzle. One typed proposal gave the puzzle a name.

Timeline at a Glance

The Mythic Machines (antiquity-1500s)

Poets imagine servants, Llull builds a reasoning wheel

Picture the forge of Hephaestus in the Iliad. When the goddess Thetis visits, the smith god is attended by "golden handmaids" who the poem says were like real young women, with sense, reason, voice and strength. In Samuel Butler's translation of the Iliad, they work for him at the forge. It is one of the oldest literary pictures of artificial helpers that think.

A few centuries later the Greek poet Apollonius Rhodius, writing in the 3rd century BCE, gave the Argonauts a harder opponent. His Argonautica describes Talos, "the man of bronze," given to Europa as the warder of Crete, who strides around the island thrice a day and keeps the Argonauts from landing. These were myths, not machines. But they set up a pattern that repeats through this whole history: people imagine the artificial worker first and build it later.

Engineers moved from myth to mechanism around the first century CE. Heron of Alexandria, whom MacTutor's biography at the University of St Andrews dates through a lunar eclipse he mentions that occurred on 13 March 62, wrote on devices worked by air, steam or water pressure. He also described an automaton theatre, a puppet show driven by strings, drums and weights. Heron's machines repeated a fixed sequence. Nobody would call them intelligent, and nobody at the time used the word.

The first real step toward mechanical reasoning came later and from an unexpected direction. Around 1305 to 1308, the Majorcan thinker Ramon Llull wrote the Ars generalis ultima. The Stanford Encyclopedia of Philosophy entry on Llull describes figures made of concentric circles that "must be rotated to form the different combinations" of letters standing for concepts. Llull wanted a tool that could help missionaries argue across religions. He got a method for generating combinations of ideas by turning a wheel.

Keep precursor and concept apart here. Llull combined concepts mechanically, but nobody called it artificial intelligence, and his wheel generated pairings rather than conclusions. Later philosophers noticed the family resemblance: Paolo Rossi's Clavis universalis, listed in the Stanford entry's bibliography, carries the subtitle "Arti mnemoniche e logica combinatoria da Lullo a Leibniz," a line from Llull to Leibniz.

For teams today: every automation wave starts as a story about a helper that never tires. The story is a signal of demand, not proof the technology is near.

The Reckoning Years (1600s-1840s)

Thinking starts to look like arithmetic

Imagine Gottfried Leibniz in Paris in the 1670s, sketching a machine that could add, subtract, multiply and divide, and carrying the designs to London. The Stanford Encyclopedia of Philosophy entry on Leibniz says he designed a calculating machine able to perform those four operations during his Paris years (1672 to 1676), and that his 1673 trip to London was meant in part to present the designs to the Royal Society.

The philosophical groundwork arrived first. In 1651, Thomas Hobbes wrote in Leviathan that "Reason, in this sense, is nothing but Reckoning," meaning adding and subtracting the consequences of agreed names. In 1666 the young Leibniz published his Dissertation on the Art of Combinations, which, per the same Stanford entry, sketched a plan for a "universal characteristic" and a logical calculus that would occupy him for much of his life. Among the names Leibniz used for this dream of a reasoning calculus was calculus ratiocinator, as the Stanford entry on his influence on logic notes.

Two hundred years later the dream got hardware. Charles Babbage designed the Analytical Engine, and Ada Lovelace translated a French account of it by Luigi Menabrea. The St Andrews MacTutor page on Lovelace records that the translation appeared in Taylor's Scientific Memoirs in September 1843, with her own notes signed AAL. Note G worked through how the engine could compute Bernoulli numbers.

And then she drew a boundary that still gets quoted in AI debates. In her Notes, Lovelace wrote that "The Analytical Engine has no pretensions whatever to originate anything." It can do whatever people know how to order it to perform, she said, and its province is to assist us in making available what we already know. Whether that holds for modern AI is argued to this day. It was the first clear statement of the question.

For teams today: the Reckoning Years left two lasting habits, to write a process down precisely and to ask what a machine can do beyond what we told it. Both are still the first steps of any automation project.

The Paper Years (1854-1950)

Logic and neurons arrive before the computers

In 1854 George Boole published An Investigation of the Laws of Thought, a treatise that, as Project Gutenberg's description puts it, sets out to give Aristotle's logic mathematical foundations through equations. Logic could now be written as algebra, which meant a mechanism could in principle run it.

The 1930s supplied the theory of what mechanisms can do. In 1936 and 1937, a young Cambridge mathematician named Alan Turing published "On Computable Numbers, with an Application to the Entscheidungsproblem," which the Turing Scrapbook describes as working out the theory of Turing machines and the definition of computability. It answered a question David Hilbert had posed about whether mathematical statements could be decided mechanically. It also gave the world the abstract blueprint for the programmable computer.

In 1943, the neurophysiologist Warren McCulloch and the logician Walter Pitts published "A logical calculus of the ideas immanent in nervous activity" in the Bulletin of Mathematical Biophysics (volume 5, pages 115 to 133). The University of Texas at San Antonio's page on the paper notes that McCulloch worked with Norbert Wiener to pioneer cybernetics and that Pitts later enrolled at MIT as Wiener's graduate student. Their paper connected brain cells to logic, and every neural network since descends from that link.

In 1948 Wiener published Cybernetics. The Computer History Museum timeline says the book had a major influence on research into artificial intelligence and control systems, and that Wiener drew on wartime work on anti-aircraft systems that anticipated the course of enemy planes. The word comes from the Greek for "steersman."

Then in October 1950, Turing published "Computing Machinery and Intelligence" in the journal Mind. The paper opens, "I propose to consider the question, 'Can machines think?'" and replaces it with an imitation game in which an interrogator, working only from typed answers, tries to tell a human from a machine. It gave the field its first measurable goal.

This era is the "theory first, practice later" stretch. For more than a century, the ideas ran ahead of anything that could execute them. The programs would have to wait for the hardware, and for a name.

For teams today: when a concept is mostly paper, treat it as a bet with a long runway. The people who understood neural networks in 1943 were not wrong. They were early.

The Dartmouth Summer (1955-1956)

Four researchers name a new field

On August 31, 1955, John McCarthy of Dartmouth College, Marvin Minsky of Harvard, Nathaniel Rochester of IBM and Claude Shannon of Bell Telephone Laboratories sent their proposal. The text, hosted by Stanford, begins by proposing "a 2 month, 10 man study of artificial intelligence" at Dartmouth in Hanover, New Hampshire, in the summer of 1956. It rests on a conjecture: that every aspect of learning or any other feature of intelligence can in principle be described precisely enough for a machine to simulate it. The proposal lists language use, abstraction, problem solving and self-improvement as targets.

Look at what that sentence assumes. It is Hobbes and Leibniz in a grant application. The proposal also did not invent the ambition, as this article has shown. What it did was name it and bring the people together. The Computer History Museum's timeline says the coining of "Artificial Intelligence" is credited to this proposal, which it calls the founding event of AI, and that five attendees returned for a 50th anniversary on the Dartmouth campus in 2006: McCarthy, Minsky, Trenchard More, Oliver Selfridge and Ray Solomonoff.

The first working program was already in motion. According to the Computer History Museum, Newell, Simon and Shaw began work on the Logic Theorist in 1955, and it eventually proved 38 theorems from Whitehead and Russell's Principia Mathematica. The museum credits it with introducing heuristics, list processing and "reasoning as search." Historian Jeremy Norman's entry on the program adds that the first version was run by hand with index cards, that Shaw later ran it on the RAND computer in Santa Monica, that it proved 38 of the first 52 theorems, and that Newell and Simon described it in RAND report P-868 on June 15, 1956.

The Logic Theorist did what Leibniz had imagined: it manipulated symbols according to rules and arrived at proofs. It also surprised its authors by finding a shorter proof of one theorem than the one in the book.

For teams today: the founding of AI was a naming and convening act, not a product launch. A shared vocabulary and a room full of people can start an industry faster than a prototype.

The Long Climb (1957-2016)

Narrow wins, with the big ambition out of reach

The founders expected fast progress. What followed was a longer climb with several detours. In 1966, Joseph Weizenbaum finished ELIZA, which the Computer History Museum describes as a natural language processing environment whose DOCTOR mode answered users like a psychotherapist, and which tricked some users into thinking they were talking to a person until it reached its limits. It showed how little understanding a convincing conversation needs.

The same museum timeline records the 1997 match where IBM's Deep Blue, evaluating 200 million positions per second, defeated world chess champion Garry Kasparov on May 11. In 2011, IBM's Watson beat Ken Jennings and Brad Rutter on Jeopardy! by analyzing natural-language questions faster and more accurately than its human rivals. Each was a narrow win in a bounded domain, which is the pattern the pillar article covers in detail, including the funding droughts known as AI winters. For the full chronology, read AI Timeline: From Turing to Large Language Models, and for what this era meant for business adoption, see How AI Moved from Research Labs to Everyday Work.

Meanwhile the older line, McCulloch and Pitts' neurons, quietly matured into machine learning and deep learning, which learn from examples instead of hand-written rules.

For teams today: narrow wins are real and worth using. Just don't extrapolate from a chess or quiz result to general capability.

The Transformer Age (2017-present)

One architecture delivers conversation at scale

In June 2017, eight researchers posted "Attention Is All You Need" on arXiv, proposing the Transformer, a network architecture based on attention. It became the design behind large language models. In 2022, OpenAI introduced ChatGPT as a free research preview, announcing it on its own page, and ordinary workers could suddenly talk to a model in plain language. That is the old Dartmouth goal of machines using language, finally within reach of a browser tab. Natural language processing moved from lab demos to daily tools.

The newest chapter extends the question Lovelace asked in 1843. Systems now propose, draft and act, as in an AI agent that works inside business software. Whether they "originate" anything is still argued. For the whole story of how we got here, start at The History of AI, and for how it reached the office, see what is business AI and the evolution of business AI. The machines themselves have their own lineage in the history of computers.

For teams today: the tools are new, but the questions are old: what can it do beyond what we told it, and who checks the output?

Key People

Key Facts

• The Dartmouth proposal is dated August 31, 1955 and called for a "2 month, 10 man study of artificial intelligence."
• Ramon Llull's Ars generalis ultima was written between 1305 and 1308 and used rotating figures to combine concepts.
• Hobbes wrote in 1651 that reason is nothing but reckoning, adding and subtracting the consequences of agreed names.
• Lovelace's Notes appeared in September 1843 and signed each note with the initials AAL.
• The Logic Theorist proved 38 of the first 52 theorems in Principia Mathematica.
• Turing's 1950 paper appeared in Mind in October 1950.

The Pattern Behind the History

Our synthesis, not published research: the origin of AI follows four beats that repeat.

• Imagine first. Myth, poetry and philosophy picked the target (a thinking helper) long before any tool could hit it.
• Formalize second. Hobbes, Leibniz, Boole and Turing turned vague ideas into rules a machine could follow.
• Name and convene third. Dartmouth did not invent the idea. It gave scattered researchers one banner.
• Overclaim, then calibrate. Every era's first demo (Llull's wheel, the Logic Theorist, ELIZA) promised more than it delivered, and Lovelace's 1843 caution was the template for the correction.

What It Teaches Teams Today

• Separate the story from the capability. Centuries of imagined machines did not bring them closer. Test the real tool on a real task.
• Write the process down before you automate it. Every step in this history, from Llull to Turing, started by making a procedure explicit.
• Name your project and bring people together. A shared vocabulary does more for early adoption than another pilot.
• Expect long runways. Neural networks waited from 1943 until the 2010s for hardware and data. Judge ideas on a decade, not a quarter.
• Keep Lovelace's question on the table. Ask what a system does beyond what you told it, and who is accountable for it.