History of Artificial Intelligence Timeline
From ancient automata myths and formal logic through the Dartmouth workshop, AI winters, deep learning and large language models, to October 2026.
Open interactive timeline
Key events
Babbage designs the Analytical Engine
Charles Babbage describes a general-purpose mechanical computer with a store, a mill and punched-card programs. It is never completed.
Ada Lovelace's notes on the Analytical Engine
Lovelace publishes an algorithm for computing Bernoulli numbers and suggests the engine could manipulate symbols beyond numbers, while arguing it cannot originate anything on its own.
George Boole's algebra of logic
Boole's The Mathematical Analysis of Logic (and his 1854 Laws of Thought) treats reasoning as algebra, the basis of later digital circuits.
Čapek's R.U.R. coins the word "robot"
Karel Čapek's play about manufactured workers who revolt introduces "robot," from the Czech robota (forced labor). It premieres in Prague in January 1921.
Gödel's incompleteness theorems
Kurt Gödel shows that any consistent formal system rich enough for arithmetic contains true statements it cannot prove, setting limits on formal reasoning.
Turing's "On Computable Numbers"
Alan Turing defines the Turing machine, shows a universal machine can simulate any other, and proves the Entscheidungsproblem unsolvable. Alonzo Church reaches the same result with lambda calculus.
McCulloch and Pitts model the neuron
Warren McCulloch and Walter Pitts show that networks of simple threshold units can compute logical functions, the first mathematical model of a neural network.
Norbert Wiener publishes Cybernetics
Wiener's book on control and communication in animals and machines popularizes feedback as a unifying idea for minds and machines.
Hebb's learning rule
In The Organization of Behavior, Donald Hebb proposes that connections between neurons strengthen when they fire together, a basis for later learning algorithms.
Turing proposes the imitation game
In "Computing Machinery and Intelligence" in Mind, Turing replaces "Can machines think?" with a conversational test, later called the Turing test.
Georgetown-IBM machine translation demonstration
An IBM 701 translates more than 60 Russian sentences into English, prompting predictions that translation would be solved within a few years.
Dartmouth proposal coins "artificial intelligence"
John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon propose a summer study on the conjecture that every aspect of intelligence can be described precisely enough for a machine to simulate it.
The Logic Theorist
Allen Newell, Herbert Simon and Cliff Shaw's program proves 38 of the first 52 theorems in Principia Mathematica, often called the first AI program.
Dartmouth Summer Research Project on AI
–The workshop at Dartmouth College gathers the field's founders and is widely regarded as the founding event of AI as a discipline.
General Problem Solver
Newell, Simon and Shaw build GPS, which uses means-ends analysis in an attempt at a universal problem-solving method; it is described in publications in 1959.
McCarthy creates Lisp
John McCarthy designs Lisp, which becomes the dominant programming language of AI research for decades.
Rosenblatt's Mark I Perceptron
Frank Rosenblatt demonstrates the perceptron, a trainable neural network, on an IBM 704, and the press predicts machines that will walk, talk and be conscious. His paper appears in Psychological Review in November.
Samuel coins "machine learning"
Arthur Samuel's paper on checkers describes programs that learn from experience; by 1962 his program beats a Connecticut checkers champion.
Unimate goes to work at General Motors
George Devol and Joseph Engelberger's Unimate, the first industrial robot, starts handling hot die-cast parts at a GM plant in New Jersey.
Dendral expert system begins at Stanford
Edward Feigenbaum, Joshua Lederberg and Bruce Buchanan start Dendral, which infers molecular structures from mass spectra and pioneers the expert-system approach.
Shakey the Robot
–SRI's Shakey is the first mobile robot to reason about its own actions, producing the STRIPS planner and the A* search algorithm.
Weizenbaum's ELIZA
Joseph Weizenbaum's MIT program imitates a psychotherapist with simple pattern matching; users' emotional attachment to it alarms its creator.
Winograd's SHRDLU
–Terry Winograd's MIT program understands English commands about a world of blocks, a high point of early natural-language understanding.
Minsky and Papert publish Perceptrons
Their book proves that single-layer perceptrons cannot compute functions such as XOR, and it helps drain support from neural network research for over a decade.
MYCIN diagnoses blood infections
–Edward Shortliffe's Stanford expert system recommends antibiotics using about 600 rules and performs on par with specialists in tests, though it is never used clinically.
The Lighthill Report
James Lighthill's report to the UK Science Research Council finds AI has failed its grand promises, citing combinatorial explosion, and leads to deep cuts in British AI research.
First AI winter
–After unmet predictions, the Lighthill Report and DARPA's cuts to speech and general research, funding and interest in AI fall sharply in the US and UK.
Expert systems boom begins with XCON
Digital Equipment Corporation deploys XCON (R1), a rule-based system that configures VAX computer orders and is credited with saving tens of millions of dollars a year, sparking a commercial expert-systems boom.
Searle's Chinese Room argument
In "Minds, Brains, and Programs," John Searle argues that a program manipulating symbols can appear to understand Chinese without any understanding at all.
Hopfield networks
John Hopfield shows that a recurrent network can store and recall memories as stable energy states, reviving interest in neural networks.
Japan launches the Fifth Generation Computer project
–Japan's MITI funds a ten-year effort to build logic-programming supercomputers for AI, prompting rival programs in the US and Europe. It falls short of its goals.
Backpropagation popularized
David Rumelhart, Geoffrey Hinton and Ronald Williams publish in Nature how multi-layer networks can learn internal representations by back-propagating errors, reviving connectionism.
Lisp machine market collapses
Cheaper desktop workstations from Sun and others undercut specialized Lisp machines, and the AI hardware industry collapses within a year.
Second AI winter
–Expert systems prove brittle and costly to maintain, specialist hardware collapses and funding dries up; many firms fold and "AI" becomes an unfashionable label.
Pearl's Probabilistic Reasoning in Intelligent Systems
Judea Pearl's book introduces Bayesian networks, bringing rigorous probability to reasoning under uncertainty.
LeCun's convolutional network reads handwritten digits
Yann LeCun at Bell Labs trains a convolutional network with backpropagation to read ZIP codes; later versions read a large share of US checks.
Support vector machines
Corinna Cortes and Vladimir Vapnik publish soft-margin support vector machines, which dominate machine learning practice for the next decade.
Long short-term memory (LSTM)
Sepp Hochreiter and Jürgen Schmidhuber's LSTM lets recurrent networks keep information over long sequences; it later powers speech recognition and translation.
Deep Blue defeats Kasparov
IBM's Deep Blue wins the rematch in New York 3½–2½, the first defeat of a reigning world chess champion by a computer in a match.
Stanford's Stanley wins the DARPA Grand Challenge
Sebastian Thrun's team drives 132 miles of desert autonomously; five vehicles finish, a turning point for self-driving cars.
Hinton's deep belief networks revive "deep learning"
Geoffrey Hinton, Simon Osindero and Yee-Whye Teh show deep networks can be trained layer by layer, rebranding neural network research as deep learning.
ImageNet dataset introduced
Fei-Fei Li's team presents ImageNet, millions of labeled images in thousands of categories; its annual challenge from 2010 becomes deep learning's proving ground.
DeepMind founded in London
Demis Hassabis, Shane Legg and Mustafa Suleyman found DeepMind Technologies with the aim of building general-purpose learning systems.
IBM Watson wins Jeopardy!
IBM's question-answering system beats champions Ken Jennings and Brad Rutter over three televised episodes (February 14–16).
Apple launches Siri
Apple introduces Siri on the iPhone 4S, bringing a voice assistant to millions of consumers.
AlexNet wins ImageNet
Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton's GPU-trained convolutional network cuts the top-5 error to 15.3%, far ahead of rivals, starting the deep learning boom.
DeepMind's DQN learns Atari games
A deep Q-network learns to play Atari games from raw pixels and scores; the Nature version in 2015 reaches human level on many of 49 games.
Google acquires DeepMind
Google buys DeepMind for a reported $500–650 million, and DeepMind sets up an ethics board as a condition of the deal.
Sequence-to-sequence learning and attention
Ilya Sutskever's team shows neural networks can translate whole sentences, and Dzmitry Bahdanau, Kyunghyun Cho and Yoshua Bengio add an attention mechanism, the seed of the Transformer.
Generative adversarial networks
Ian Goodfellow and colleagues pit a generator against a discriminator, enabling realistic synthetic images and, later, deepfakes.
Bostrom publishes Superintelligence
Nick Bostrom's book argues that a superintelligent AI could be catastrophic if its goals are misaligned, bringing AI risk into mainstream debate.
OpenAI founded
Sam Altman, Elon Musk, Ilya Sutskever, Greg Brockman and others launch OpenAI as a nonprofit research lab, with backers pledging $1 billion.
AlphaGo defeats Lee Sedol
–DeepMind's AlphaGo beats one of the world's best Go players 4–1 in Seoul; its move 37 in game two shows unexpected creativity.
"Attention Is All You Need" introduces the Transformer
Ashish Vaswani and seven Google co-authors propose the Transformer, an architecture built on attention alone that becomes the basis of nearly all large language models.
AlphaGo Zero and AlphaZero learn from self-play alone
AlphaGo Zero masters Go with no human games; in December AlphaZero generalizes the method to beat top programs at chess, shogi and Go.
OpenAI releases GPT-1
OpenAI's first generative pre-trained transformer (117 million parameters) shows that pre-training on raw text then fine-tuning works well across tasks.
Google releases BERT
Jacob Devlin's team's bidirectional Transformer sets records across language-understanding benchmarks and is soon used in Google Search.
OpenAI announces GPT-2 and withholds the full model
Citing misuse risks, OpenAI initially releases only smaller versions of its 1.5-billion-parameter model, publishing the full model in November.
Deep learning pioneers win the Turing Award
Geoffrey Hinton, Yann LeCun and Yoshua Bengio receive the 2018 Turing Award for breakthroughs that made deep neural networks a critical part of computing.
Scaling laws for language models
Jared Kaplan and colleagues at OpenAI show model performance improves predictably with more parameters, data and compute, guiding the race to scale.
OpenAI unveils GPT-3
The 175-billion-parameter model can perform many tasks from a few examples in its prompt; OpenAI offers it through an API from June.
AlphaFold 2 solves protein structure prediction
DeepMind's AlphaFold 2 predicts protein structures with near-experimental accuracy at the CASP14 assessment, cracking a 50-year-old biology problem.
Anthropic founded
Dario Amodei, Daniela Amodei and other former OpenAI staff found Anthropic as a public benefit corporation focused on AI safety research.
OpenAI unveils DALL-E and CLIP
DALL-E generates images from text descriptions, and CLIP learns to link images and text from internet captions.
GitHub Copilot previews AI pair programming
GitHub and OpenAI launch Copilot, powered by the Codex model, which suggests code as developers type.
InstructGPT and RLHF
OpenAI fine-tunes GPT-3 with reinforcement learning from human feedback to follow instructions, the technique behind ChatGPT.
Stable Diffusion released with open weights
Stability AI and partners release an image generator anyone can run on a consumer GPU, spreading generative AI and sparking copyright fights.
ChatGPT launches
OpenAI releases a free chatbot built on GPT-3.5; it reaches a million users in five days and about 100 million within two months, the fastest-growing consumer app to that point.
Meta releases LLaMA to researchers
Meta's 7B–65B-parameter models are released for research; the weights leak online within a week, kicking off the open-weight model movement.
Anthropic releases Claude
Anthropic's first Claude model becomes available through a chat interface and API.
GPT-4 released
OpenAI's multimodal GPT-4 accepts images and text and scores in the top 10% on a simulated bar exam.
Open letter calls for a six-month pause on giant AI experiments
The Future of Life Institute's letter, signed by Elon Musk, Steve Wozniak, Yoshua Bengio and thousands of others, asks labs to pause training systems more powerful than GPT-4.
Hinton leaves Google to warn about AI risks
Geoffrey Hinton says he quit so he could speak freely about the dangers of the technology he helped create.
Biden signs Executive Order 14110 on AI
The order requires developers of the most powerful models to share safety test results with the US government.
First AI Safety Summit at Bletchley Park
–28 countries, including the US and China, sign the Bletchley Declaration on risks from frontier AI; the UK and US set up AI safety institutes.
OpenAI board fires and rehires Sam Altman
–The board removes Altman, saying he was not consistently candid; after most staff threaten to quit and Microsoft offers them jobs, he returns with a new board.
Google DeepMind launches Gemini
Google's natively multimodal Gemini family debuts in Ultra, Pro and Nano sizes.
OpenAI previews Sora; Google announces Gemini 1.5
Sora generates minute-long video from text, and Gemini 1.5 offers a context window of up to a million tokens.
GPT-4o brings real-time voice and vision
OpenAI's "omni" model reasons across audio, images and text in real time and becomes ChatGPT's default.
EU AI Act enters into force
The world's first comprehensive AI law sorts systems by risk level, with obligations phased in over the following years.
OpenAI o1 introduces reasoning models
OpenAI's o1 is trained to think at length before answering, sharply improving math, science and coding results.
Hopfield and Hinton win the Nobel Prize in Physics
John Hopfield and Geoffrey Hinton are honored for foundational discoveries that enable machine learning with artificial neural networks.
Hassabis and Jumper share the Nobel Prize in Chemistry
The AlphaFold creators share the prize with David Baker, who is honored for computational protein design.
DeepSeek-R1 shakes the industry
Chinese lab DeepSeek releases an open-weight reasoning model rivaling o1 at a fraction of the reported cost; on January 27 its app tops the US App Store and Nvidia loses about $590 billion in market value in a day.
Stargate Project announced
OpenAI, SoftBank, Oracle and MGX announce up to $500 billion for US AI data centers, unveiled at the White House.
Claude 3.7 Sonnet and Claude Code
Anthropic releases a hybrid reasoning model and Claude Code, a command-line agent that writes and edits code; Claude Code becomes generally available in May.
AI models win gold at the International Mathematical Olympiad
Experimental reasoning models from OpenAI and Google DeepMind each solve five of six problems in natural language, a gold-medal score.
GPT-5 released
OpenAI unifies its models in GPT-5, which routes between fast and reasoning modes; user backlash leads it to restore GPT-4o for paying users within days.
OpenAI restructures as a public benefit corporation
OpenAI Group PBC is placed under the nonprofit OpenAI Foundation, and Microsoft takes about a 27% stake, after a share sale valued the company at $500 billion.
Gemini 3 released
Google's Gemini 3 Pro leads many benchmarks and ships with Nano Banana Pro image generation days later.
SpaceX completes its merger with xAI
The all-stock deal values the combined company at about $1.25 trillion, folding xAI and X into SpaceX.
Pentagon blacklists Anthropic
After Anthropic refuses to drop limits on military uses of Claude, the President orders agencies to stop using it and the Defense Secretary labels it a supply chain risk; a federal judge issues an injunction on March 26.
Project Glasswing and Claude Mythos Preview
Anthropic gives a highly capable model, Mythos Preview, only to a small group of organizations to find and fix software vulnerabilities, judging it too risky for general release.
Claude Fable 5 and Mythos 5 released
Anthropic releases Mythos-class models; days later a US Commerce Department directive briefly blocks foreign nationals from them, a restriction lifted entirely on June 30.
GPT-6 Astra launches to limited users
OpenAI releases GPT-6 Astra first to selected enterprises (reported as September 3 or 4), its first model rated "Critical" for cybersecurity under its Preparedness Framework.
142 more in the interactive timeline.
Open interactive timeline