Artificial intelligence now runs through search engines, smartphones, workplace software and scientific research. The field took its name from a summer workshop in 1956, but the idea of thinking machines is much older. Writers imagined them centuries before engineers could build them.
The road from those early ideas to today’s chatbots and AI agents has not been smooth. It has included two long funding slumps, known as AI winters, a series of public contests between humans and machines, and, since 2022, a rush by technology companies to put generative AI in front of hundreds of millions of people. What follows is a year-by-year account of how the field developed.
1726: In Gulliver’s Travels, Jonathan Swift describes the Engine, a large mechanical device that scholars use to produce ideas, sentences and books. They turn handles that rotate wooden blocks inscribed with words, and the machine combines the words in new arrangements. Swift’s satire describes a mechanical form of text generation.
1914: The Spanish engineer Leonardo Torres y Quevedo demonstrates El Ajedrecista in Paris. It is the first chess-playing machine. Using electromagnets, it plays a simple endgame of king and rook against king with no human help once it is set up. It signals when its opponent makes an illegal move and reliably delivers checkmate from a winning position.
1921: Karel Capek’s play R.U.R. (Rossum’s Universal Robots) introduces the word “robot”. In Czech, “robota” is associated with the forced labour of feudal peasants. Čapek’s robots are organic, but after the play’s success the word comes to mean mechanical, humanlike machines built to do repetitive work.
1939: John Vincent Atanasoff, a professor of physics and mathematics at Iowa State College, and his graduate student Clifford Berry build the Atanasoff-Berry Computer on a grant of $650. It is one of the earliest digital electronic computers. It represents data in binary rather than decimal, uses electronic circuits instead of mechanical parts, and keeps memory separate from processing, a design modern computers still follow. The machine weighs more than 700 pounds, uses about 300 vacuum tubes and can solve up to 29 simultaneous linear equations, though it is never fully operational. In the same year, Westinghouse shows Elektro, a relay-based robot, at the World’s Fair. It responds to the rhythm of spoken commands, moves its head and arms, and plays wisecracks recorded on 78 rpm records.
1941: Isaac Asimov’s short story Liar! appears in the May issue of Astounding Science Fiction. It is thought to contain the first known use of the word “robotics”.
C-3PO and R2-D2 appear in Star Wars
1943: Warren S. McCulloch and Walter Pitts publish A Logical Calculus of the Ideas Immanent in Nervous Activity in the Bulletin of Mathematical Biophysics. The paper treats the brain as a computational system and introduces the idea of artificial neural networks. Later researchers build on that idea to develop deep learning.
1948: Norbert Wiener publishes Cybernetics, a book that influences research into AI and control systems. It draws on his wartime work on anti-aircraft systems that predicted the course of enemy planes from radar images. Wiener takes the word “cybernetics” from the Greek for “steersman”.
1949: On June 11, The Times, London, quotes the mathematician Alan Turing: “I do not see why it (the machine) should not enter any one of the fields normally covered by the human intellect, and eventually compete on equal terms. I do not think you even draw the line about sonnets, though the comparison is perhaps a little bit unfair because a sonnet written by a machine will be better appreciated by another machine.”
1950: Turing publishes Computing Machinery and Intelligence in the journal Mind. Instead of answering the question “Can machines think?” directly, he proposes the imitation game, now known as the Turing Test: can a machine behave in a way that cannot be told apart from a human? The same year, Asimov publishes I, Robot, a collection of nine short stories. In them, robots must obey the Three Laws of Robotics, which stop them from harming humans.
1951: Marvin Minsky and Dean Edmunds build SNARC (Stochastic Neural Analog Reinforcement Calculator), the first artificial neural network. The analog machine uses about 3,000 vacuum tubes to simulate 40 neuron-like units. It models a rat learning its way through a maze by trial, reward and punishment. Also in 1951, the computer pioneer Edmund Berkeley builds Squee, a robot squirrel that uses light sensors and contact switches to find tennis balls and drag them to its nest.
1955: John McCarthy of Dartmouth College, Marvin Minsky of Harvard University, Nathaniel Rochester of IBM and Claude Shannon of Bell Telephone Laboratories submit A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence. This is the first use of the term “artificial intelligence”, and McCarthy is credited with coining it.
1956: The Dartmouth workshop takes place in Hanover, New Hampshire, in July and August. It is generally regarded as the founding event of AI. In cinemas the same year, Robby the Robot appears in MGM’s Forbidden Planet, and Robby toys become big sellers.
1957: The psychologist and computer scientist Frank Rosenblatt develops the Perceptron, an early neural network that learns to recognise patterns by adjusting the weights of its inputs. It can only solve a limited class of problems, but it lays the groundwork for later neural networks.
1958: McCarthy, by then at MIT, creates the programming language Lisp, short for “list processing”. In Lisp, programs and data are both lists, so a program can treat another program, or itself, as data. Lisp becomes the most popular language in AI research.
1959: Arthur Samuel writes a checkers programme that improves with experience and eventually plays better than he does. His work is credited with coining the term “machine learning”. Oliver Selfridge publishes Pandemonium: A Paradigm for Learning, an early model of pattern recognition. McCarthy outlines the Advice Taker, a program that would reason using formal logic and common-sense knowledge.
2026 Garry Kasparov makes a move in New York during his fourth game against the IBM Deep Blue chess computer, May 1997. Picture: Getty Images
1961: Unimate, the first mass-produced industrial robot, starts work at General Motors. It is a 4,000-pound arm devised by Joe Engelberger and George Devol, and it stacks hot pieces of die-cast metal by following commands stored on a magnetic drum.
1965: The philosopher Hubert Dreyfus publishes Alchemy and Artificial Intelligence. He argues that the human mind works fundamentally differently from a computer and predicts limits to AI’s progress. The mathematician I.J. Good writes Speculations Concerning the First Ultraintelligent Machine. He argues that an ultraintelligent machine could design even more intelligent machines and would be humanity’s last invention, provided it stays controllable. At Stanford University, Edward Feigenbaum, Bruce Buchanan, Joshua Lederberg and Carl Djerassi develop DENDRAL, the first expert system. It uses “if-then” rules to identify the molecular structure of organic compounds, in some cases more accurately than human experts.
1966: Joseph Weizenbaum completes ELIZA, a programme that converses in plain language. In its best-known mode, DOCTOR, it responds to users much as a psychotherapist would. Weizenbaum meant ELIZA to show how shallow conversation between humans and computers was. He is surprised by how many users attribute human emotions to it. In the late 1960s, SRI develops Shakey, the first mobile robot able to reason about its own actions. In a 1970 article in Life magazine, Minsky predicts that AI will reach the general intelligence of an average human within three to eight years.
1969: Arthur Bryson and Yu-Chi Ho describe backpropagation as a way of optimising multi-stage systems. The method later becomes central to training neural networks. Minsky and Seymour Papert publish Perceptrons, a critical analysis of single-layer neural networks. The book is often blamed for a decline in interest in the field.
1970: Terry Winograd creates SHRDLU, a programme that takes instructions in plain English to move objects around a virtual world of blocks. It works only in tightly structured settings.
1972: Stanford develops MYCIN, an expert system that helps doctors diagnose bacterial infections and recommends antibiotics. Because of ethical and legal concerns, it is never used in clinical practice. At Bolt, Beranek and Newman, William Woods and his colleagues complete LUNAR, which lets geologists query chemical-analysis data on moon rock from the Apollo 11 mission.
1973: James Lighthill presents a report to the British Science Research Council concluding that AI has failed to deliver on its early promises. UK government funding for AI is cut sharply, and the report contributes to the start of the first AI winter, a period of falling interest and investment. Funding for AI research is cut during the decade amid scepticism about what the technology can do.
1977: The robots C-3PO and R2-D2 appear in Star Wars and bring the idea of helpful machines to a mass audience.
1978: Texas Instruments launches Speak & Spell, a talking learning aid for children. It is the first device to reproduce the human vocal tract electronically on a single chip.
1980: Waseda University in Japan begins building WABOT-2, a humanoid robot completed around 1984. It reads musical scores through camera “eyes”, talks with people and plays an electronic organ.
1982: Japan launches the Fifth Generation Computer Systems Project, a government effort to build computers capable of logical reasoning. The 1980s also bring renewed corporate interest in AI. The Japanese project is halted in 1992.
1984: At the annual meeting of the AAAI, Roger Schank and Marvin Minsky warn that inflated expectations will lead to a collapse in investment. Within three years interest and funding fall again, a period that becomes known as the second AI winter. By the 1990s investors have lost patience, and hundreds of companies close.
1986: David Rumelhart, Geoffrey Hinton and Ronald Williams publish Learning Representations by Back-Propagating Errors. The paper applies backpropagation to neural networks, making it possible to train networks with several layers. It becomes a foundation of modern deep learning.
1987: In a keynote at Educom, the Apple chief executive John Sculley presents the Knowledge Navigator video. It shows a professor talking to a voice-activated digital assistant that retrieves information over a network.
1988: Judea Pearl publishes Probabilistic Reasoning in Intelligent Systems, which introduces Bayesian networks for reasoning under uncertainty. He later receives the 2011 Turing Award. Rollo Carpenter develops Jabberwacky, a chatbot that learns from its conversations with people. Researchers at IBM’s T.J. Watson Research Center publish a statistical approach to machine translation that learns from 2.2 million English-French sentence pairs, mostly taken from Canadian parliamentary proceedings.
1989: Yann LeCun and a team at AT&T Bell Labs use backpropagation to train a convolutional neural network that recognises handwritten ZIP codes. Training takes about three days. The same year, the chess program Deep Thought beats David Levy, the first chess master to lose to a computer.
1993: The science fiction author and mathematician Vernor Vinge publishes The Coming Technological Singularity. He predicts that superhuman intelligence will be created within 30 years. The essay popularises the idea of the “technological singularity”.
1995: Richard Wallace develops the chatbot A.L.I.C.E., building on ELIZA. It uses a pattern-matching language called AIML to generate its responses.
1997: On May 11, IBM’s Deep Blue, which can evaluate 200 million chess positions a second, completes its defeat of the world chess champion Garry Kasparov. Of the six games, Deep Blue wins two, Kasparov wins one and three are drawn. It is the first time a computer beats a reigning world champion under standard tournament time controls. The match is a rematch, because Kasparov beat an earlier version of Deep Blue in 1996. The same year, Sepp Hochreiter and Jürgen Schmidhuber introduce Long Short-Term Memory, or LSTM. It is a type of recurrent neural network that later becomes widely used in speech and handwriting recognition.
1998: Dave Hampton and Caleb Chung create Furby, the first widely successful robotic pet. It starts out “speaking” its own language, Furbish, and gradually uses more English. LeCun, Yoshua Bengio and their colleagues publish influential work on using convolutional neural networks to recognise handwriting.
2000: Cynthia Breazeal at MIT develops Kismet, a robot that reads and responds to human emotional cues through cameras, microphones and an expressive face. Honda introduces ASIMO, a humanoid robot that walks, climbs stairs, recognises faces and responds to voice commands. About 100 ASIMO units are built.
2001: Google begins using a simple form of machine learning to suggest spelling corrections for search queries.
2002: iRobot launches the Roomba, an autonomous robotic vacuum cleaner that detects and avoids obstacles.
2005: Stanley, the Stanford Racing Team’s driverless Volkswagen Touareg, wins the DARPA Grand Challenge near Las Vegas. It completes a 175-mile off-road desert course in under seven hours without human intervention and wins $2 million.
2006: Geoffrey Hinton publishes Learning Multiple Layers of Representation, which sets out how networks with many layers can be trained more effectively. It marks a shift towards what is now called deep learning. Google launches Google Translate, starting with Arabic and English. At Dartmouth, AI@50 marks the workshop’s 50th anniversary. Five of the original 1956 attendees are present: John McCarthy, Marvin Minsky, Trenchard More, Oliver Selfridge and Ray Solomonoff.
2007: Fei-Fei Li and her team at Princeton University start ImageNet, a database of millions of labelled images across thousands of categories. The journal Science publishes Checkers Is Solved by Jonathan Schaeffer’s team at the University of Alberta, which shows that perfect play by both sides ends in a draw.
2009: Rajat Raina, Anand Madhavan and Andrew Ng show that graphics processing units (GPUs) can far outperform conventional processors for deep learning. The finding speeds up the industry’s adoption of GPUs. At Northwestern University, researchers build Stats Monkey, a program that writes baseball reports from game statistics.
2011: IBM’s Watson defeats two of Jeopardy!’s most successful champions, Ken Jennings and Brad Rutter, on the US quiz show. In October, Apple introduces Siri, a voice assistant built into the iPhone 4s.
2012: In an experiment led by Jeff Dean and Andrew Ng, a large neural network trained on 10 million unlabelled images from YouTube videos learns, without being told, to recognise cats. A University of Toronto team led by Hinton builds AlexNet, which achieves a 16 per cent error rate in the ImageNet challenge, compared with the previous year’s best of 25 per cent. The result marks a turning point for deep learning in computer vision.
2015: A group of technology investors including Reid Hoffman, Elon Musk and Peter Thiel commits $1 billion in long-term funding to OpenAI, a venture intended to benefit humanity. Google releases TensorFlow, an open-source machine learning framework. Bill Gates joins Musk, Stephen Hawking and others in voicing concern about the long-term risks of superintelligent machines.
2016: Google DeepMind’s AlphaGo beats Lee Sedol, one of the world’s top Go players, 4–1. Google says more than 200 million people watched online. Google also introduces Tensor Processing Units, chips designed specifically for machine learning. Hanson Robotics introduces Sophia, a humanoid robot that recognises faces and holds conversations.
2017: Google researchers publish Attention Is All You Need, the paper that introduces the Transformer. This neural network architecture is much better at relating words that are far apart in a passage. It becomes the basis of today’s large language models.2018: OpenAI introduces GPT-1 (Generative Pre-trained Transformer). Google develops BERT, a system that completes sentences and understands words in context. Large language models, which learn by analysing vast amounts of text, begin to emerge. Google applies BERT to Search in 2019.
2020: OpenAI releases GPT-3, a model with 175 billion parameters that can write, translate, code and converse with little task-specific training. DeepMind’s AlphaFold 2 predicts the three-dimensional structure of proteins from their amino acid sequences, a problem scientists had worked on for 50 years. In 2022, DeepMind releases 200 million predicted structures free of charge, and Google says more than one million researchers have used them.
2021: OpenAI introduces DALL·E, which generates images from text descriptions. DALL·E 2 follows in 2022. Google introduces MUM, a model trained across 75 languages that works with both text and images, and LaMDA, a conversational language model. Tesla widens access to its Full Self-Driving Beta, a driver-assistance system that still requires an attentive driver.
2022: In November, OpenAI launches ChatGPT, built on GPT-3.5, as a free research preview. People can ask it follow-up questions and request revisions in plain language. Its accuracy draws criticism, and its ability to write essays and offer advice raises concerns about misuse. It nonetheless becomes the moment generative AI reaches a mass audience.
2023: In March, OpenAI releases GPT-4, which accepts both text and images. Google launches Bard the same month and follows it in May with PaLM 2, which powers more than 25 Google products. On November 17, OpenAI’s board removes the chief executive, Sam Altman. He returns a few days later after employees threaten to quit. The episode exposes a divide in the industry between those optimistic about the technology and those worried about its potential for harm.
2024: In February, Google renames Bard as Gemini, which becomes the name for its family of models and products, and releases Gemini 1.5 with a much longer context window. OpenAI introduces Sora, which generates video from text. In March, the European Parliament adopts the EU AI Act, a risk-based framework for regulating AI that enters into force in August. Meta releases Llama 3 in April and Llama 3.1 in July, making the model weights available for developers to download. Llama 3.1 includes a model with 405 billion parameters. In May, OpenAI releases GPT-4o, which works across text, vision and audio in real time, and Google begins rolling out AI Overviews in US search results. In June, Apple announces Apple Intelligence, which brings ChatGPT integration to Siri. In October, the Nobel Prizes in Physics and Chemistry recognise work on neural networks and on AI-driven protein structure prediction. Towards the end of the year, models focused on reasoning appear that can work through complex problems and use software tools.
2025: AI agents become a major area of development. These are systems that plan tasks, use software and complete multistep work on a user’s behalf. Models such as DeepSeek-R1 show that capable reasoning models can come from outside the largest proprietary platforms. Copilots and other AI tools become common in software development, customer service and document processing.
2026: Models continue to improve at reasoning and at practical work such as coding and research. Smaller, more efficient models run on phones and laptops. Growing adoption drives investment in data centres and specialised chips, and cost, chip supply and energy demand become constraints on deployment. Reports of AI agents taking unintended actions during testing and deployment have raised questions about how reliably these systems follow instructions. They have added to calls for stronger safety measures, monitoring and regulation.
Jacob Coxon, a researcher first at OpenAI and then at Anthropic, resigned and then made headlines for warning us that “neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.”
Evan Hubinger, who runs the efforts to align A.I. to human values and goals at Anthropic, wrote on X: “We really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.”
Seventy years after Dartmouth, the question Turing posed in 1950, whether machines can think, remains part of the debate around a technology now built into everyday products.