Brief History of AI

From idea to everyday tool

Artificial intelligence was officially named in 1955 and founded as a research field the following summer. Its history since then has not been a steady climb. It has moved in waves, with bursts of progress and investment followed by “winters” when funding and confidence collapsed.

Two broad approaches have competed throughout. One tried to write intelligence down as rules and logic; the other tried to let machines learn from examples. The rule-writers led for decades, but the learning approach now dominates, and most of today’s AI, from fraud detection to chatbots, descends from it.

What follows covers the key moments and the key advances, from the first ideas to the present. A quick-reference timeline of dates sits at the end.

Each winter followed a boom in which promises outran results. There has been no comparable collapse since the early 1990s.

Before the machines

The ambition predates the computer. In 1843 Ada Lovelace published her notes on Charles Babbage’s Analytical Engine, suggesting it could manipulate symbols other than numbers. She also doubted it could ever originate anything of its own, a doubt Alan Turing would later call Lady Lovelace’s objection and argue against.

In 1854 George Boole published The Laws of Thought, reducing logical reasoning to algebra. In 1937 Claude Shannon showed that Boole’s algebra could describe electrical switching circuits, the basis of every digital computer since. Eighteen years later, Shannon co-signed the proposal that gave AI its name.

1943 to 1956: the field is founded

One of the clearest intellectual starting points for the neural networks behind today’s chatbots came in 1943, when Warren McCulloch and Walter Pitts published a mathematical model of the neuron. They showed that networks of simple on-off units could compute logical functions.

In October 1950 Alan Turing published Computing Machinery and Intelligence in the journal Mind. He replaced the slippery question of whether machines can think with a practical test, the imitation game, now known as the Turing test. He also suggested that machines might one day learn, rather than be programmed for every task.

Britain built some of the first working AI programs. Christopher Strachey’s draughts program, first tried on the Pilot ACE at the National Physical Laboratory in 1951, was playing complete games on the Ferranti Mark 1 at Manchester by the summer of 1952. In the United States, Arthur Samuel’s checkers program at IBM improved by playing against itself, and in 1959 Samuel coined the term “machine learning”.

The name came last. A proposal dated 31 August 1955 by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon sought funding for a summer study of “artificial intelligence” at Dartmouth College in 1956. Its founding conjecture was that every aspect of learning could, in principle, be described precisely enough for a machine to simulate it.

At that workshop, Allen Newell and Herbert Simon, working with programmer Cliff Shaw, presented the Logic Theorist, a program that proved theorems from Principia Mathematica. The 1956 Dartmouth summer is generally taken as the founding of AI as a field.

1956 to 1980: the first boom, and the first winter

Early results were striking, and the promises grew faster than the results. In 1958 Frank Rosenblatt demonstrated the Perceptron, a machine that learned to classify simple images, and newspapers predicted it would soon walk and talk. The same year McCarthy created LISP, which remained the main language of AI research for three decades.

In January 1966 Joseph Weizenbaum of MIT described ELIZA, a program that mimicked a psychotherapist by turning users’ sentences back into questions. It understood nothing, yet people confided in it. The human tendency to credit fluent software with understanding is still called the ELIZA effect, and it applies directly to modern chatbots.

The reckoning followed. In 1966 the ALPAC report concluded that machine translation cost more than human translators and performed worse, and US funding was cut. In 1969 Minsky and Seymour Papert’s book Perceptrons set out the limits of single-layer networks, and neural network research stalled for over a decade.

In Britain, the Science Research Council asked the mathematician Sir James Lighthill to assess the field. His report, published in 1973, judged that AI had failed to meet its grand objectives, and it was debated on BBC television the same year. UK funding collapsed, and research survived in a handful of universities, Edinburgh chief among them.

With US defence funding also cut in 1973 and 1974, the first AI winter set in and lasted until about 1980. The cause was less outright failure than promises that had outrun what the technology could deliver.

1980 to 1993: expert systems, then the second winter

AI’s first commercial boom was built on rules rather than learning. Expert systems captured a specialist’s knowledge as hundreds or thousands of if-then rules. Stanford’s MYCIN recommended treatments for blood infections in the 1970s, and from 1980 Digital Equipment Corporation used XCON to configure customer orders for its VAX computers.

Governments joined in. Japan launched its Fifth Generation Computer Systems project in 1982; Britain answered with the Alvey Programme in 1983, and the United States with DARPA’s Strategic Computing Initiative. By the mid-1980s AI had become a billion-dollar industry.

Then it broke. In 1987 the market for specialised Lisp machines collapsed as cheaper desktop computers from Apple and IBM overtook them. Expert systems proved brittle and costly to maintain, and even XCON became too expensive to keep current.

Japan’s project closed in 1992 without meeting its goals, and funding stayed scarce until the mid-1990s. Many researchers stopped calling their work AI at all, preferring labels such as machine learning or knowledge-based systems. The field kept advancing, but largely under other names.

1986 to 2011: the quiet revolution

While expert systems faded, the learning approach revived. In October 1986 David Rumelhart, Geoffrey Hinton and Ronald Williams published backpropagation in Nature, a method for training networks with hidden layers. The mathematics had earlier roots, but this paper made multi-layer networks practical, and it underpins nearly every neural network trained today.

In 1989 Yann LeCun at Bell Labs applied convolutional networks to handwritten digits, and versions of the system were later used to read cheques in the United States. In 1997 Sepp Hochreiter and Jürgen Schmidhuber introduced LSTM networks, which gave neural networks a working memory for sequences such as speech.

On 11 May 1997 IBM’s Deep Blue won a six-game rematch against Garry Kasparov by 3½ to 2½. It was the first time a computer had beaten a reigning world champion in a match under standard tournament conditions. Deep Blue relied on brute-force search rather than learning, but it reset public expectations overnight.

Meanwhile machine learning went to work out of sight. By the 2000s it was filtering spam and flagging card fraud for millions of people who never heard the term. In 2005 Stanford’s driverless car Stanley won the DARPA Grand Challenge, completing a 132-mile desert course with nobody at the wheel.

The ingredients for the next leap came together. In 2006 Hinton’s group showed how to train deep networks layer by layer, helping to popularise the term deep learning. In 2009 Fei-Fei Li’s team presented ImageNet, a database that grew past 14 million labelled images and gave researchers the training data they had lacked.

AI then reached consumers directly. In February 2011 IBM’s Watson beat two of the strongest human champions on the US quiz show Jeopardy!, and in October 2011 Apple launched Siri on the iPhone 4S. Millions of people were now speaking to an AI assistant every day.

2012 to 2020: deep learning breaks through

The modern era has a clear start date. On 30 September 2012 a University of Toronto team (Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton) entered AlexNet in the ImageNet challenge. It achieved a top-5 error rate of 15.3%, against 26.2% for the runner-up, a margin that turned the whole field towards neural networks trained on graphics processors.

Industry moved quickly. In January 2014 Google acquired DeepMind, a London start-up co-founded in 2010 by Demis Hassabis. The same year Ian Goodfellow introduced generative adversarial networks, which made highly convincing synthetic imagery practical and accelerated research into generative models.

In March 2016 DeepMind’s AlphaGo beat Lee Sedol, one of the world’s strongest Go players, by four games to one in Seoul. Go has vastly more possible positions than chess, and many experts had expected that milestone to be a decade away. Later that year Google moved its Translate service to neural networks.

Arguably the most important paper for today’s chatbots appeared in June 2017. In Attention Is All You Need, eight Google researchers introduced the Transformer, an architecture that processes whole sequences in parallel rather than word by word. It is the T in GPT.

Language models then scaled rapidly. OpenAI released GPT-1 in 2018, the same year Google released BERT, and in 2019 OpenAI initially withheld the full GPT-2 over concerns about misuse. GPT-3 followed in 2020 and was opened to developers through an API.

Science took notice too. In November 2020 DeepMind’s AlphaFold 2 predicted protein structures with accuracy approaching laboratory methods, at the CASP14 assessment, cracking a problem biologists had worked on for 50 years.

2021 to 2026: the generative era

Generative tools arrived in stages. OpenAI showed DALL-E in January 2021, and GitHub previewed its Copilot coding assistant that June. Image generation reached the public in 2022, with DALL-E 2 in April followed by Midjourney and Stable Diffusion over the summer.

On 30 November 2022 OpenAI released ChatGPT, putting a fine-tuned GPT-3.5 model behind a simple chat window. A UBS analysis reported by Reuters estimated it reached 100 million monthly users within two months, at the time the fastest growth of any consumer application. GPT-4 followed on 14 March 2023, and rival assistants from Anthropic and Google launched the same month.

Governments responded. On 1 and 2 November 2023 the UK hosted the first AI Safety Summit at Bletchley Park, where Turing’s wartime codebreakers had worked, and 28 countries and the EU signed the Bletchley Declaration. The EU AI Act entered into force on 1 August 2024.

In July 2026 the EU adopted changes that delayed many of the AI Act’s high-risk obligations. Most stand-alone high-risk systems covered by Annex III moved to 2 December 2027, while many transparency obligations still took effect on 2 August 2026.

The scientific establishment recognised the long arc of the work. In October 2024 the Nobel Prize in Physics went to John Hopfield and Geoffrey Hinton for neural network research dating from the 1980s. Half the chemistry prize went to Demis Hassabis and John Jumper of Google DeepMind for AlphaFold.

Capabilities kept moving. In September 2024 OpenAI released o1, a reasoning model designed to spend more computation on a problem before answering. On 20 January 2025 China’s DeepSeek released R1, an open-weight reasoning model; a week later Nvidia’s shares fell about 17% in a day, wiping roughly $589 billion off its value.

Underneath these launches, the model itself was becoming more general-purpose. Large pretrained foundation models could be adapted to many tasks; instruction tuning and human feedback made them easier to steer; reasoning models added more deliberate inference; and agents extended the model beyond answering into tool use and multi-step action. By 2026 agentic AI had moved into everyday enterprise infrastructure.

What the history shows

Progress has come in waves rather than a straight line. Both winters followed periods when promises outran what the technology could deliver, and both ended when a different approach began producing results.

Big ideas usually arrive long before their impact. Backpropagation was published in 1986 but transformed industry only after 2012, once large datasets and graphics processors caught up. The Transformer appeared in a 2017 research paper, five years before it reached a mass-market chat product.

AI also tends to disappear from view once it works. Reading handwriting and filtering spam were once AI research problems; today few people think of them as AI at all. Much of what feels new is the latest visible layer of work that has been running for decades.

Quick-reference timeline

Key dates in order, from the first ideas to 2026.

Year	Event	Why it matters
1843	Ada Lovelace’s notes on the Analytical Engine	First published idea of a machine handling symbols, not just numbers
1854	George Boole, The Laws of Thought	Logic expressed as algebra
1943	McCulloch and Pitts model of the neuron	Mathematical ancestor of neural networks
1950	Turing, Computing Machinery and Intelligence	Proposes the imitation game, now the Turing test
1952	Strachey’s draughts program runs at Manchester	Among the first working AI programs
1955	Dartmouth proposal, dated 31 August	Coins the term “artificial intelligence”
1956	Dartmouth summer workshop; Logic Theorist	AI founded as a research field
1958	Rosenblatt’s Perceptron; McCarthy’s LISP	An early learning machine; AI’s main language for decades
1959	Samuel coins “machine learning”	Programs that improve with experience
1966	Weizenbaum’s ELIZA	First chatbot; origin of the ELIZA effect
1966	ALPAC report	US machine translation funding cut
1969	Minsky and Papert, Perceptrons	Neural network research stalls
1973	Lighthill report (UK)	Helps trigger the first AI winter
1980	XCON in use at Digital Equipment Corporation	Expert systems become a business
1982	Japan’s Fifth Generation project	Government AI race begins
1983	UK Alvey Programme	Britain’s response
1986	Backpropagation paper in Nature	Makes multi-layer networks trainable
1987	Lisp machine market collapses	Start of the second AI winter
1989	LeCun’s convolutional networks read handwriting	Neural networks in commercial use
1997	Deep Blue beats Kasparov; LSTM published	Machine beats a world champion; memory for sequences
2005	Stanley wins the DARPA Grand Challenge	Driverless car completes a desert course
2006	Hinton’s group trains deep networks layer by layer	Revival under the name deep learning
2009	ImageNet presented	Large labelled dataset for computer vision
2011	IBM Watson wins Jeopardy!; Apple launches Siri	AI reaches consumers
2012	AlexNet wins ImageNet	Start of the deep learning era
2014	Google acquires DeepMind; GANs introduced	Big tech commits; synthetic images
2016	AlphaGo beats Lee Sedol 4 to 1	Go mastered years ahead of expectations
2017	Transformer paper published	Architecture behind modern chatbots
2018	GPT-1 and BERT	Pre-trained language models
2019	GPT-2 staged release	First public debate over releasing a language model
2020	GPT-3; AlphaFold 2	Fluent text via API; protein structures predicted
2021	DALL-E; GitHub Copilot preview	Generated images and code
2022	ChatGPT launches, 30 November	Generative AI goes mainstream
2023	GPT-4; Bletchley Park AI Safety Summit	Frontier models; first global safety summit
2024	EU AI Act in force; OpenAI’s o1	Comprehensive regulation; reasoning models
2024	Nobel prizes to Hopfield, Hinton, Hassabis and Jumper	Neural network research recognised at the highest level
2025	DeepSeek R1	Open-weight reasoning model; record one-day fall in Nvidia’s value
2026	EU delays key high-risk AI Act obligations; agents go mainstream	Regulation adjusts; AI moves from answering to acting