Who invented artificial intelligence?
No single person invented artificial intelligence. Alan Turing asked in 1950 whether machines can think. John McCarthy coined the term "artificial intelligence" in 1955 and, with Marvin Minsky, Nathaniel Rochester, and Claude Shannon, ran the 1956 Dartmouth workshop that founded the field.
"Invented" is the wrong verb for AI, and that is the actual finding here, not a dodge. Inventions have a patent, a date, a name on the plaque. AI has a paper that reframed the question (Turing, 1950), a proposal that named the field (McCarthy, Minsky, Rochester, and Shannon, 1955), and then seven decades of separate people solving separate pieces: how a neuron could be modeled as logic, how a machine could learn a weight, how attention could replace recurrence. Ask who created artificial intelligence and you get a list, not a name.
What did Alan Turing contribute in 1950?
Turing's 1950 paper "Computing Machinery and Intelligence" reframed "can machines think?" as a testable question. His imitation game, later called the Turing Test, asks whether a machine's answers can fool a human interrogator into thinking it is a person.
Six years before AI had a name, Turing had already done the hard part: he made the question answerable. His 1950 paper opens with "I propose to consider the question, 'Can machines think?'" and then sidesteps the philosophy entirely. Instead of defining thought, he proposes a game: an interrogator questions a hidden human and a hidden machine through text alone and has to guess which is which. If the machine's answers are good enough to fool the interrogator at the same rate a human would, Turing argued, arguing about what is happening "inside" the machine stops being useful. That reframe, from an unanswerable philosophical question to a testable behavioral one, is why historians treat 1950 as the intellectual starting line even though the field had no name yet.
How did artificial intelligence start at Dartmouth in 1956?
Artificial intelligence started as an organized field on August 31, 1955, when John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon wrote a funding proposal for a 1956 summer workshop at Dartmouth College, coining the term "artificial intelligence" in that document.
The proposal that started it all is a 17-page typescript asking the Rockefeller Foundation for about $13,500 to fund "a 2 month, 10 man study of artificial intelligence" at Dartmouth College the following summer. Its central conjecture, written in one sentence, still describes what every AI lab is doing today: "every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it." The four authors, McCarthy at Dartmouth, Minsky at Harvard, Rochester at IBM, and Shannon at Bell Labs, ran the workshop in summer 1956. Attendees drifted in and out over the two months rather than sitting through one continuous meeting, and the group did not solve intelligence that summer. What they did was give the problem a name and a home field, which is why Dartmouth 1956, not any single paper, is the date most textbooks use when they need one.
Who is the father of artificial intelligence?
John McCarthy is the person most commonly called "the father of artificial intelligence," a title used in his 2011 Los Angeles Times obituary. He coined the term, organized the Dartmouth workshop, and later invented Lisp, the first AI programming language.
"Father of AI" is journalism's shorthand, not a formal title, and it lands on McCarthy more often than anyone else. He coined the phrase, wrote the proposal that convened the field, and spent the following decades building the tools AI research actually ran on: he invented Lisp in 1958, one of the oldest programming languages still in use, and founded the Stanford AI Lab in 1965. Some writers give the title to Turing instead, for laying the conceptual groundwork six years earlier, and both claims have a real basis. If you need one name for a search box, McCarthy is the safer answer. If you want the honest one, it is both of them plus the Dartmouth co-authors, and that is a more useful answer than a single name because it explains why the field looks the way it does: a philosopher's question turned into an engineer's research program.
Why did AI go through two winters?
AI went through two funding winters because researchers overpromised and governments stopped paying. The 1973 Lighthill Report triggered the first, cutting UK funding after AI failed to scale past toy problems. The 1987 Lisp-machine market collapse and the failure of expensive, brittle expert systems triggered the second.
The first winter has a start date you can point to. In 1973, mathematician James Lighthill delivered a government-commissioned report concluding that "in no part of the field have the discoveries made so far produced the major impact that was then promised", and singled out the combinatorial explosion problem: methods that worked on small, contrived examples fell apart the moment problems got realistically large. The British government cut AI funding at most UK universities on the strength of that report, and US agencies pulled back around the same time.
The field recovered in the early 1980s on the back of expert systems, rule-based programs sold as captured human expertise, and specialized Lisp machines built to run them, priced at $70,000 to $150,000 each. Both collapsed within a few years of each other. General-purpose workstations from Apple and Sun matched Lisp-machine performance at a fraction of the cost, and the specialized-hardware market cratered starting in 1987; Symbolics, the industry's flagship company, filed for bankruptcy in 1993. Expert systems turned out to be expensive to build, brittle outside their narrow rule sets, and unable to learn, so the promised return never showed up. Two winters, two different causes, one shared lesson: hype that outruns delivered results eventually gets its funding pulled.
Where did neural networks come from?
Neural networks trace to 1943, when McCulloch and Pitts modeled a neuron as a logic gate, and to 1958, when Rosenblatt built the Perceptron, the first machine that could learn its own weights. Backpropagation made deep networks trainable, work that won Hinton, LeCun, and Bengio the 2018 Turing Award.
Neural networks are older than the term "artificial intelligence" itself. In 1943, neurophysiologist Warren McCulloch and logician Walter Pitts published a paper modeling a neuron as a simple logical unit: it takes weighted inputs and fires past a threshold, and networks of these units can compute any logical function. It was math, not hardware. The first machine came in 1958, when psychologist Frank Rosenblatt built the Perceptron, a system of photocells and adjustable weights that learned to classify patterns from examples instead of being programmed rule by rule. That learning rule, adjust the weights when the output is wrong, is the direct ancestor of how every neural network trains today.
The missing piece was training networks with more than one layer, and backpropagation solved it: an algorithm for pushing an error signal backward through a network to update every weight at once. Geoffrey Hinton, Yann LeCun, and Yoshua Bengio spent the 1980s through 2000s turning that idea into working systems, Hinton on backpropagation and Boltzmann machines, LeCun on convolutional networks, Bengio on sequence and attention models, and in 2018 the three shared the ACM A.M. Turing Award for their work on deep neural networks. Computer science's top honor is named after the same person who asked the founding question in 1950.
What changed between 2012 and 2026?
Three events built the current AI era: AlexNet's 2012 ImageNet win proved deep networks work at scale, the 2017 Transformer architecture made large language models practical, and ChatGPT's 2022 launch put a chat interface in front of everyone. By 2026, agents built on this stack are clearing real computer-use tasks two-thirds of the time.
For most of AI's history, the ideas existed before the compute did. That gap closed in 2012, when a team from the University of Toronto entered a convolutional neural network called AlexNet in the ImageNet competition and won by a margin nobody had seen before: a 15.3% error rate against a 26.2% runner-up, using GPUs instead of the CPUs everyone else was training on. Yann LeCun called it an unequivocal turning point for computer vision, and it was the moment deep learning stopped being an academic curiosity.
Five years later, a Google team published "Attention Is All You Need," introducing the Transformer, an architecture built entirely on attention instead of the recurrence every prior language model relied on. It trained faster, scaled further, and became the base architecture under every large language model that followed, including the one that wrote this sentence. Five years after that, on November 30, 2022, OpenAI shipped that architecture wrapped in a chat window and called it ChatGPT, and it hit 100 million users faster than any consumer product before it. By 2026, the frontier has moved from chat to action: the Stanford HAI AI Index reports that AI agents jumped from 12% to roughly 66% task success on OSWorld, a benchmark of real computer-use tasks, in about a year, though enterprise deployment of those agents still lags well behind the capability.
Why does this history matter for adopting AI today?
Every AI cycle since 1956 overpromised on its own hype, and every cycle still left behind durable, still-used tools: Lisp, the Perceptron's learning rule, backpropagation, the Transformer. The lesson for adopting AI now is to bet on what already works in production, not on what the current demo promises.
Run the timeline again and a pattern jumps out: overclaiming preceded both winters, and useful tools survived both winters anyway. Lisp is still a working language. Backpropagation still trains every network in production. The Transformer, five years old, is still the architecture under the model you are probably running this year. The hype cycle and the tool's durability are two different variables, and betting on the second one is what has actually paid off across 76 years of this field.
We build AI systems for businesses today, and that is the practical takeaway we work from: ship what is already proven in production, keep a human checking the parts that are new, and treat this year's frontier demo the way the field should have treated expert systems in 1985, as promising and unproven until it survives contact with a real workload.
