Vote this week
GUIDE · TIMELINE

The history of AI

Artificial intelligence feels new, but people have been working on it for more than 80 years. Its history is a story of big ideas, bigger promises, two crashes and one remarkable comeback. Here it is in one timeline, with every milestone linked to its original source.

By Parth AminReviewed 14 min readHow we made this
XLinkedInRedditWhatsAppEmail
Illustration: a winding path from an old tape computer past a chess piece to a laptop with a chat bubble
On this page

Who invented AI?

TIP

Short answer: no single person. Alan Turing asked the founding question in 1950. John McCarthy coined the name “artificial intelligence” in 1955 and organized the 1956 Dartmouth workshop that started the field, so he is most often called its father. Newell, Simon and Shaw wrote the first AI program the same year.

The timeline

51 of 51 milestones

Foundations

1943–1955

Before the field had a name, a few mathematicians asked whether thinking could be described precisely enough for a machine to do it.

  1. A mathematical model of a neuronWarren McCulloch and Walter Pitts showed that networks of simple on-off 'neurons' could compute logic. It is the seed of every neural network since.Source: Bulletin of Mathematical Biophysics, 1943 ↗ (opens in a new tab)
  2. Key moment: Turing asks “Can machines think?”Alan Turing's paper in the journal Mind swapped that vague question for a game: can a machine's typed answers pass for a person's? He called it the imitation game; we call it the Turing test.Source: Mind, 1950 ↗ (opens in a new tab)
  3. The name “artificial intelligence”John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon asked for money for a summer study. Their proposal is the first written use of the phrase.Source: AI Magazine reprint of the proposal ↗ (opens in a new tab)

The birth of AI

1956–1959

A summer workshop gave the field its name, and the first programs that reasoned, learned and played games appeared within a few years.

  1. Key moment: The Dartmouth workshopThe summer meeting at Dartmouth College brought the first AI researchers together and launched the field under its new name.Source: Dartmouth College ↗ (opens in a new tab)
  2. Logic Theorist, the first AI programAllen Newell, Herbert Simon and Cliff Shaw wrote a program that proved theorems from Principia Mathematica, a landmark logic text. It is widely called the first AI program.Source: Britannica ↗ (opens in a new tab)
  3. The perceptron learnsFrank Rosenblatt demonstrated a machine that learned to sort inputs from examples rather than following hand-written rules. His later Mark I Perceptron hardware is in the Smithsonian.Source: Cornell Chronicle ↗ (opens in a new tab)
  4. LISPMcCarthy created the LISP programming language, which became the working language of AI research for decades.Source: Stanford News ↗ (opens in a new tab)
  5. “Machine learning” gets its nameArthur Samuel's checkers program at IBM got better by playing against itself, and Samuel coined the term machine learning.Source: IBM History ↗ (opens in a new tab)

Early optimism

1960–1972

Chatbots, robots and expert programs made rapid progress look certain. Researchers predicted human-level AI within a generation.

  1. Key moment: ELIZA, the first chatbotJoseph Weizenbaum's ELIZA matched patterns in what people typed. It understood nothing, yet users confided in it, which alarmed its creator.Source: Communications of the ACM, 1966 ↗ (opens in a new tab)
  2. Shakey the robotSRI built Shakey, the first mobile robot that could perceive its surroundings and plan its own moves. It is now at the Computer History Museum.Source: SRI ↗ (opens in a new tab)
  3. The book PerceptronsMinsky and Seymour Papert set out what single-layer perceptrons could not do. Interest in neural networks faded for years.Source: Lehigh University Libraries ↗ (opens in a new tab)
  4. MYCIN diagnoses infectionsWork began at Stanford on MYCIN, a system of about 500 rules for diagnosing blood infections that performed at roughly a specialist's level.Source: Britannica ↗ (opens in a new tab)

First AI winter

1973–1979AI WINTER

Promises outran results. Critical reviews and tighter budgets cooled the field for most of a decade.

  1. Key moment: The Lighthill reportA survey for the UK government judged that AI had fallen far short of its promises. It was followed by a televised debate and helped usher in the first AI winter.Source: University of Edinburgh ↗ (opens in a new tab)

Expert systems boom

1980–1986

Rule-based programs that captured a specialist's know-how became real business tools, and money poured back in.

  1. Expert systems go to workDigital Equipment Corporation began using R1 (later XCON), a rule-based system that configured computer orders, in daily operations. It became AI's first big commercial success.Source: AI Magazine ↗ (opens in a new tab)
  2. Japan's Fifth Generation projectJapan launched a ten-year national program to build computers for knowledge processing, prompting rival investment in the US and Europe.Source: IPSJ Computer Museum ↗ (opens in a new tab)
  3. Hopfield networksJohn Hopfield showed how a network could store memories as stable patterns. The work was cited in his 2024 Nobel Prize.Source: PNAS, 1982 ↗ (opens in a new tab)
  4. A warning about winterAt a major AI conference, leading researchers warned that inflated expectations could end in another backlash. Within a few years they were proved right.Source: AI Magazine ↗ (opens in a new tab)
  5. Key moment: Backpropagation catches onDavid Rumelhart, Geoffrey Hinton and Ronald Williams showed in Nature that backpropagation lets many-layered networks learn useful internal features. It is still how networks are trained.Source: Nature, 1986 ↗ (opens in a new tab)

Second AI winter

1987–1993AI WINTER

Specialized AI hardware lost to cheaper computers and expert systems proved hard to maintain. Funding fell again.

  1. The AI hardware market collapsesCheaper general-purpose computers overtook specialized LISP machines, expert systems proved costly to keep up, and funding dried up again.Source: Making Things Think (Holloway) ↗ (opens in a new tab)
  2. A network reads handwritingYann LeCun and colleagues trained a convolutional neural network to read handwritten ZIP codes, an idea that would later power computer vision.Source: Neural Computation, 1989 ↗ (opens in a new tab)

Machine learning rises

1994–2011

Instead of writing rules by hand, researchers let programs learn from data. More data and faster chips made it work.

  1. LSTM memorySepp Hochreiter and Jürgen Schmidhuber gave neural networks a way to remember over long sequences, which later powered speech recognition and translation.Source: Neural Computation, 1997 ↗ (opens in a new tab)
  2. Key moment: Deep Blue beats KasparovIBM's chess computer won a six-game match against the reigning world champion, 3½ to 2½. It won by searching millions of positions a second, not by learning.Source: Centre for Computing History ↗ (opens in a new tab)
  3. LeNet-5LeCun's team published the convolutional network design that read bank cheques, the template for modern image recognition.Source: Proceedings of the IEEE, 1998 ↗ (opens in a new tab)
  4. Deep networks become trainableHinton, Simon Osindero and Yee-Whye Teh showed a way to train deep networks one layer at a time, reviving interest in 'deep learning'.Source: Neural Computation, 2006 ↗ (opens in a new tab)
  5. ImageNetFei-Fei Li's team released a huge labelled collection of photos. The yearly contest built on it became the proving ground of the deep learning era.Source: CVPR 2009 ↗ (opens in a new tab)
  6. Watson wins Jeopardy!IBM Watson beat two of the show's greatest champions, answering questions posed in everyday language.Source: IBM History ↗ (opens in a new tab)

Deep learning

2012–2016

Large neural networks trained on graphics chips suddenly beat every other approach at seeing, listening and playing.

  1. Key moment: AlexNetA deep network trained on graphics chips won the ImageNet contest with a top-5 error rate of 15.3%, against 26.2% for the next best entry. The field changed course almost overnight.Source: ImageNet 2012 results ↗ (opens in a new tab)
  2. word2vecGoogle researchers found a fast way to turn words into lists of numbers that capture meaning, so 'king' sits near 'queen'.Source: arXiv ↗ (opens in a new tab)
  3. GANsIan Goodfellow and colleagues pitted two networks against each other, one forging images and one spotting fakes. It opened the door to realistic generated images.Source: arXiv ↗ (opens in a new tab)
  4. Key moment: AlphaGo beats Lee SedolDeepMind's AlphaGo beat one of the world's best Go players 4 to 1 in Seoul. Its 'Move 37' looked like a mistake to experts and turned out to be brilliant.Source: Google DeepMind ↗ (opens in a new tab)

Transformers and scaling

2017–2021

One new design, the transformer, turned out to get better simply by making it bigger and feeding it more text.

  1. Key moment: “Attention Is All You Need”Eight Google researchers introduced the transformer, a design built around attention. Nearly every large language model since uses it.Source: arXiv ↗ (opens in a new tab)
  2. GPT-1OpenAI showed that pretraining a transformer on lots of plain text, then fine-tuning it, beats training for each task from scratch.Source: OpenAI ↗ (opens in a new tab)
  3. BERTGoogle's BERT read text in both directions at once and set new records on language tasks. It soon improved Google Search.Source: arXiv ↗ (opens in a new tab)
  4. GPT-2OpenAI released its larger language model in stages, citing worries about misuse, an early public debate about releasing AI.Source: OpenAI ↗ (opens in a new tab)
  5. Turing Award for deep learningYoshua Bengio, Geoffrey Hinton and Yann LeCun received computing's highest honour for making deep neural networks work.Source: ACM ↗ (opens in a new tab)
  6. GPT-3A far larger model could do new tasks from just a few examples in the prompt, no retraining needed.Source: arXiv ↗ (opens in a new tab)
  7. Key moment: AlphaFold 2 cracks protein foldingDeepMind's system predicted the 3D shapes of proteins with accuracy the contest's organizers called a solution to a 50-year-old challenge in biology.Source: Google DeepMind ↗ (opens in a new tab)

The generative era

2022–2023

AI that writes, draws and chats reached millions of ordinary people within weeks.

  1. AI images go mainstreamDALL·E 2, Midjourney and the openly released Stable Diffusion arrived within months of each other, and anyone could make a picture from a sentence.Source: DALL·E 2 paper (arXiv) ↗ (opens in a new tab)
  2. Key moment: ChatGPT launchesOpenAI put a chat window on top of a large language model. It became one of the fastest-growing consumer apps ever and brought AI into everyday life.Source: OpenAI ↗ (opens in a new tab)
  3. GPT-4 and ClaudeOn the same day, OpenAI released GPT-4, which could also read images, and Anthropic released its first Claude models.Source: Anthropic ↗ (opens in a new tab)Source: OpenAI ↗ (opens in a new tab)
  4. GeminiGoogle launched Gemini, a family of models built from the start to handle text, images, audio and code.Source: Google ↗ (opens in a new tab)

Reasoning and agents

2024–today

Models learned to think before answering and to use tools and computers on their own, and lawmakers began to set rules.

  1. Key moment: The EU AI Act takes effectThe first broad law regulating AI entered into force in the European Union, with its rules phased in over the following years.Source: Council of the EU ↗ (opens in a new tab)
  2. Models that reason firstOpenAI's o1 was trained to work through a problem step by step before answering, the start of 'reasoning models'.Source: OpenAI ↗ (opens in a new tab)
  3. Key moment: Two Nobel Prizes for AIHopfield and Hinton won the Physics prize for neural networks. Half the Chemistry prize went to Demis Hassabis and John Jumper for AlphaFold, the other half to David Baker.Source: Nobel Prize ↗ (opens in a new tab)Source: Second source ↗ (opens in a new tab)
  4. AI uses a computerAnthropic let Claude operate a computer in a public beta: look at the screen, move the pointer, click and type.Source: Anthropic ↗ (opens in a new tab)
  5. Model Context ProtocolAnthropic published an open standard for connecting AI assistants to apps and data. It was later handed to a foundation under the Linux Foundation.Source: Anthropic ↗ (opens in a new tab)Source: Second source ↗ (opens in a new tab)
  6. DeepSeek-R1A Chinese lab released an openly licensed reasoning model it said matched the best closed ones.Source: DeepSeek ↗ (opens in a new tab)
  7. Agents that browseOpenAI previewed Operator, an agent that uses a web browser to complete tasks for you.Source: OpenAI ↗ (opens in a new tab)
  8. Gold-medal mathAn advanced Gemini model scored at gold-medal level at the International Mathematical Olympiad, graded by the competition's own coordinators.Source: Google DeepMind ↗ (opens in a new tab)
  9. 800 million weekly ChatGPT usersOpenAI's chief executive said ChatGPT reached 800 million people a week, under three years after launch. The figure is OpenAI's own.Source: TechCrunch ↗ (opens in a new tab)
  10. Most EU AI rules applyEnforcement began for most of the EU AI Act, while some deadlines for high-risk systems were pushed later.Source: Council of the EU ↗ (opens in a new tab)

The timeline above is grouped into eras. Tap an era to filter it, or a milestone to open its source. Key moments are marked with a star.

The eras in brief

Foundations (1943–1955)

Before computers were common, a few people asked whether thinking could be described precisely enough for a machine to do it. In 1943 Warren McCulloch and Walter Pitts described a mathematical neuron. In 1950 Alan Turing proposed his imitation game.

The birth of AI (1956–1959)

A summer workshop at Dartmouth College gave the field its name. Within three years there were programs that proved theorems, a machine that learned from examples (the perceptron) and the first use of the words “machine learning”.

Early optimism (1960–1972)

ELIZA chatted, Shakey the robot planned its moves and expert programs diagnosed diseases. Leading researchers predicted human-level AI within a generation. They were decades off.

The first AI winter (1973–1979)

Results lagged far behind the predictions. A critical report for the UK government, and tighter budgets in the US, cooled the field. This was the first AI winter.

The expert systems boom (1980–1986)

Programs that captured a specialist's rules became real business tools. Companies and governments, led by Japan's Fifth Generation project, poured money back in. Quietly, neural network research revived with backpropagation.

The second AI winter (1987–1993)

Expert systems were costly to maintain and brittle outside their narrow job. Cheaper desktop computers made specialized AI hardware obsolete. Funding fell again.

Machine learning rises (1994–2011)

The field changed strategy: instead of writing rules, let programs learn patterns from data. Chess fell to Deep Blue, handwriting to convolutional networks, and the ImageNet dataset set the stage for what came next.

Deep learning (2012–2016)

In 2012 a large neural network trained on graphics chips crushed the ImageNet contest. Deep learning soon led in vision, speech and games, capped by AlphaGo's win at Go in 2016.

Transformers and scaling (2017–2021)

Google researchers introduced the transformer. It kept getting better as it got bigger and read more text, the pattern behind today's large language models. AlphaFold showed AI could crack hard science too.

The generative era (2022–2023)

Image generators and then ChatGPT put AI that writes, draws and talks in front of millions of people within weeks.

Reasoning and agents (2024–today)

Models learned to reason before answering and to act as agents that use tools and computers. AI pioneers won Nobel Prizes, and the EU passed the first broad AI law.

AI's boom-and-bust history: eras, the two AI winters and six key moments.

Myths about AI history

MYTH

Myth: “Artificial intelligence” was coined at the 1956 workshop, or by Turing. McCarthy coined it in the August 1955 proposal for the workshop. Turing died in 1954, before the term existed. Source

MYTH

Myth: Turing invented “the Turing test”. He called it the imitation game. Others named it after him later. Source

MYTH

Myth: Deep Blue learned to play chess. It searched about 200 million positions a second, scoring them with a chess formula its team built and tuned. It did not learn the way today's AI does. Source

MYTH

Myth: Backpropagation was invented in 1986. Paul Werbos described it in his 1974 doctoral thesis, and others found it again. The 1986 Nature paper made it famous. Source

MYTH

Myth: OpenAI invented the transformer, the “T” in GPT. Researchers at Google published it in 2017. OpenAI built GPT on top of it. Source

Watch: two moments told as stories

The full, award-winning documentary about the 2016 match between AlphaGo and Lee Sedol, free from Google DeepMind.

Why the 2012 AlexNet result changed the field, explained visually.

What the history teaches

  • Hype cycles are normal. AI has been oversold twice and crashed twice. Healthy skepticism about any single claim is part of the tradition.
  • Ideas wait for hardware. Neural networks were invented in the 1940s and 50s but only worked at scale once chips and data caught up.
  • Learning beat rules. Hand-written rules hit a ceiling. Systems that learn from data kept improving.
  • The public notices late. Most of the breakthroughs behind ChatGPT were published years before anyone outside research had heard of them.

History in the making

The race between AI assistants is part of this story too. Every week, people vote at thisorthis.ai for the AI they think is best. This is how the leaders have changed:

The #1 AI each week

  1. W42 so farClaudeleads week 42 so far
See the rank-over-time chart →

Go deeper

TRY IT · OUR WORLD IN DATACharts on AI's growth(opens in a new tab) TRY IT · STANFORD HAIThe AI Index report(opens in a new tab) TRY IT · MIND (OXFORD ACADEMIC)Read Turing's 1950 paper(opens in a new tab) TRY IT · COMPUTER HISTORY MUSEUMAI and robotics timeline to 2015(opens in a new tab)

Questions people ask

Who invented artificial intelligence?

No single person. Alan Turing asked whether machines could think in 1950. John McCarthy coined the name “artificial intelligence” in a 1955 proposal written with Marvin Minsky, Nathaniel Rochester and Claude Shannon, and organized the 1956 Dartmouth workshop that launched the field. McCarthy is most often called the father of AI.

When was AI first invented?

The field was founded at the Dartmouth workshop in the summer of 1956. The first program usually called AI, Logic Theorist, was written the same year. The ideas go back further, to a 1943 mathematical model of a neuron.

What was the first AI program?

Logic Theorist, written in 1955–56 by Allen Newell, Herbert Simon and Cliff Shaw, proved theorems from Principia Mathematica. The first chatbot was ELIZA in 1966.

What were the AI winters?

Two periods when funding and interest collapsed after AI failed to meet its promises: roughly 1973 to 1979, after the critical Lighthill report in the UK, and roughly 1987 to 1993, after the market for AI hardware and expert systems fell apart.

When did deep learning take off?

In 2012, when a deep neural network called AlexNet won the ImageNet image-recognition contest by a wide margin. Within a few years deep learning was the dominant approach in AI.

When was ChatGPT released?

30 November 2022. It brought large language models to the general public and became one of the fastest-growing consumer apps in history.

How we made this

Parth Amin runs thisorthis.ai and has spent over a decade building software products, most recently enterprise generative-AI systems. @_aminparth on X
How we made this: drafted with AI assistance, then fact-checked against the linked sources and edited by hand. Reviewed . Spotted a mistake? Tell us.
Found this useful? Share it with someone learning AI.
XLinkedInRedditWhatsAppEmail
Add your vote to history: which AI is the best this week?Vote

Keep reading