The complete history of AI, from ancient myths to ChatGPT

Short version: the idea of artificial intelligence is ancient, but the field itself was named in 1955 and launched at a summer workshop at Dartmouth College in 1956. Early researchers thought human-level AI was 20 years away. It wasn’t, and funding collapsed twice (the “AI winters” of the 1970s and late 1980s). What finally worked wasn’t hand-writing the rules of intelligence. It was letting computers learn from huge amounts of data, which only became practical once three things lined up around 2012: the internet’s mountain of data, cheap graphics chips (GPUs) and better learning methods. A 2017 design called the transformer led directly to ChatGPT in November 2022, and to the AI assistants and agents businesses use today. Almost every big idea in that story was decades old before the hardware caught up with it.

AI’s history at a glance

Before computers: myths, automata and logic

Ancient dreams

The Greeks told of Talos, a bronze giant built by the god Hephaestus to patrol the island of Crete, and Homer’s Iliad describes golden handmaidens Hephaestus made that could think and speak. China, India and Egypt have their own stories of artificial servants. None of them were real, but they show how old the question is: could something made also think?

Machines that moved by themselves

In 1206 the engineer Ismail al-Jazari, working in what is now Turkey, wrote a book describing water-powered automata, including a band of mechanical musicians whose rhythm could be changed by moving pegs on a rotating drum. That’s a program, more than 700 years before software. European clockmakers later built writing and drawing dolls of astonishing detail.

Not every “thinking machine” was honest. The Mechanical Turk (1770) toured Europe beating people at chess, including, so the story goes, Napoleon. It was a cabinet with a skilled human player hidden inside. It’s worth remembering whenever a demo looks too good to be true.

Thinking as calculation

Meanwhile, philosophers were asking whether reasoning itself could be mechanical. Around 1300 Ramon Llull built paper wheels that combined ideas to produce new statements. In the 1670s Gottfried Leibniz built a calculating machine and dreamed of a universal language of logic, so that arguments could be settled by saying “let us calculate”. In 1854 George Boole showed that logic could be written as algebra, with true and false as 1 and 0. Every computer since runs on Boolean logic.

The first program, and the first objection

In the 1830s Charles Babbage designed the Analytical Engine, a steam-age general-purpose computer that was never finished. In 1843 Ada Lovelace published notes on it that included a method for calculating Bernoulli numbers, widely regarded as the first published computer program. She also saw that the machine could work on music or symbols, not just numbers. But she drew a line: the Engine “has no pretensions whatever to originate anything”, it can only do what we tell it to. Turing later called this “Lady Lovelace’s objection”, and people are still arguing about it every time an AI writes a poem.

The word robot arrived in 1920, in Karel Čapek’s play R.U.R., from the Czech robota, forced labour. The fear of machines turning on their makers came with it.

The foundations: logic, neurons and a test (1936–1955)

Three ideas from this period underpin everything since.

What a computer is. In 1936 Alan Turing described an imaginary machine of a tape and a short table of rules, and showed that one “universal” machine could imitate any other if you gave it the other machine’s rules as data. That idea is the basis of the stored-program computer, the kind of machine every AI runs on. (We explain how it works, with one you can run in your browser, in The Turing machine.)

What a neuron does. In 1943 Warren McCulloch and Walter Pitts showed that a very simplified brain cell, one that “fires” when its inputs add up past a threshold, could compute logic. Wire enough together and you can, in principle, compute anything. In 1949 the psychologist Donald Hebb suggested how real brains learn: connections between neurons that fire together get stronger. Neural networks still learn by adjusting the strength of connections.

How to judge intelligence. In 1950 Turing published “Computing Machinery and Intelligence”. Rather than argue about the word “think”, he proposed the imitation game, now called the Turing test: if a judge chatting by text can’t reliably tell the machine from a person, what grounds do we have to say it isn’t thinking? He predicted that by about 2000 a computer would fool an average judge 30% of the time in a five-minute chat.

Electronic computers arrived at the same time. Australia built one of the first: CSIRAC, which first ran in Sydney in 1949, was the first computer to play music and is now the world’s oldest intact first-generation electronic computer, on display at Scienceworks in Melbourne.

In Britain, Christopher Strachey had a draughts (checkers) program running on the Ferranti Mark 1 by 1952. In the US, Arthur Samuel at IBM wrote a checkers program that improved by playing itself, and in 1959 coined the phrase machine learning.

1956: the summer AI got its name

In August 1955 John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon proposed a summer research project at Dartmouth College in New Hampshire. Their pitch was bold: every aspect of learning or intelligence can, in principle, be described so precisely that a machine can be made to simulate it. They called the subject artificial intelligence, a name McCarthy picked partly to set it apart from neighbouring fields.

The 1956 workshop didn’t produce a breakthrough, but it gathered nearly everyone who would lead AI for the next 20 years. The most impressive thing there came from Allen Newell, Herbert Simon and Cliff Shaw: the Logic Theorist, a program that went on to prove 38 of the first 52 theorems in chapter 2 of Whitehead and Russell’s Principia Mathematica, a famous work of mathematical logic. One of its proofs was neater than the original. Most historians call it the first AI program.

The golden years and the big promises (1956–1974)

The next two decades were heady. Government money, mostly US defence funding, flowed to labs at MIT, Stanford and Carnegie Mellon, and in Edinburgh, and the results looked like magic for the era.

• LISP (1958). McCarthy’s programming language became the language of AI research for 30 years.
• The Perceptron (1958). Frank Rosenblatt’s perceptron, first run as a program on an IBM computer and then built as the Mark I Perceptron with a grid of 400 light sensors, learned to recognise simple shapes from examples. The New York Times reported that the US Navy expected it to “walk, talk, see, write, reproduce itself and be conscious of its existence”. It couldn’t.
• ELIZA (1966). Joseph Weizenbaum’s program at MIT imitated a psychotherapist by turning your words back into questions (“Why do you say you are unhappy?”). It understood nothing, yet people confided in it, and Weizenbaum’s own secretary famously asked him to leave the room. He spent the rest of his career warning about how easily we trust machines. Sixty years later, that warning applies to every chatbot.
• Shakey the robot (1966–72). Built at the Stanford Research Institute, Shakey was the first mobile robot that could reason about its own actions: look at a room, plan a route, push a block. The A* route-finding method invented for it still guides characters in video games and directions in map apps.
• Expert knowledge (1965 onward). Stanford’s DENDRAL worked out the structure of molecules from chemistry data the way an expert chemist would, and is often called the first “expert system”.
• SHRDLU (1968–70). Terry Winograd’s program could hold a sensible conversation in English about a small world of coloured blocks: “Pick up a big red block.” “What does the box contain?”

Researchers were wildly confident. In 1960 Herbert Simon wrote that machines would be capable, “within twenty years, of doing any work that a man can do”. In 1967 Marvin Minsky wrote that within a generation “the problems of creating ‘artificial intelligence’ will be substantially solved”.

The trouble was that all these programs worked in tiny, tidy worlds. A blocks world has a dozen objects; the real world has millions, and the number of possibilities to check explodes as problems grow (researchers called it the combinatorial explosion). Computers of the day had a tiny fraction of the memory in a modern phone. And in 1969 Minsky and Seymour Papert’s book Perceptrons showed that a single-layer perceptron couldn’t learn even simple patterns like “one or the other but not both”. Interest in neural networks collapsed for more than a decade.

The first AI winter (1974–1980)

The bills came due. In the US a 1966 report (ALPAC) had already concluded that machine translation was slower, less accurate and dearer than human translators, and funding for it was cut. In 1973 the British government’s Lighthill report judged that AI research had not delivered the major impact it had promised, and UK funding was slashed. In the US, defence funders shifted to projects with direct military use and wanted results, not research.

Money and enthusiasm drained away. This was the first AI winter: a period when “artificial intelligence” became a phrase that made funders nervous. Some researchers kept working under other names, like “pattern recognition” and “informatics”.

Expert systems: AI goes to work (1980–1987)

AI came back by aiming lower. Instead of general intelligence, expert systems captured the know-how of human specialists as hundreds or thousands of “if this, then that” rules, in one narrow area.

• MYCIN, developed at Stanford in the 1970s, recommended antibiotics for blood infections and in tests did about as well as specialists. It was never used routinely on real patients, held back by practical hurdles and by the question of who would be responsible if it got one wrong. That question hasn’t gone away.
• XCON (also called R1) configured computer orders for Digital Equipment Corporation from 1980 and was reported to save the company tens of millions of dollars a year. Suddenly AI was a business tool.

By the mid-1980s most large US companies had an AI group, a market for specialised “Lisp machine” computers had sprung up, and Japan launched its Fifth Generation Computer Systems project in 1982 to build intelligent machines. The US and UK hurried out their own programs in response.

Two quieter developments mattered more in the long run. In 1982 John Hopfield showed a new kind of neural network that could store and recall memories. In 1986 David Rumelhart, Geoffrey Hinton and Ronald Williams popularised backpropagation, a way for networks with several layers to learn by passing their errors backwards and nudging every connection to do a little better next time. It’s still how almost every AI model in the world is trained.

Australia contributed too. Rodney Brooks, who was born and grew up in Adelaide, argued at MIT in the mid-1980s that robots didn’t need a detailed model of the world to act sensibly in it: simple behaviours layered on top of each other could produce smart-looking action. He co-founded iRobot, whose Roomba vacuum (2002) became the first robot in millions of homes.

The second AI winter (1987–1993)

In 1987 the market for Lisp machines collapsed almost overnight, as ordinary desktop computers from Apple and IBM became powerful enough to do the same work far more cheaply. Expert systems turned out to be expensive to keep up to date, and brittle: ask them about anything slightly outside their rules and they gave confidently wrong answers. Japan’s Fifth Generation project ended in 1992 without meeting its goals.

AI’s reputation sank again. But this time the research didn’t stop. It changed direction.

The quiet comeback: AI learns from data (1990s–2011)

Through the 1990s and 2000s AI researchers moved away from hand-writing rules and towards statistics and machine learning: give the computer lots of examples and let it find the patterns itself. Much of the success was barely called AI, because it simply worked:

• Reading handwriting. Yann LeCun’s neural network for reading digits (LeNet, 1989 onward) ended up reading more than one in ten US bank cheques by the late 1990s.
• Remembering sequences. In 1997 Sepp Hochreiter and Jürgen Schmidhuber published the LSTM, a network with a working memory that later powered speech recognition and translation on billions of phones.
• Chess. In May 1997 IBM’s Deep Blue beat world champion Garry Kasparov over a six-game match. It won by brute force, searching up to 200 million positions a second, not by learning, but it was a cultural turning point.
• Spam filters (2002 onward) learned which emails were junk from the words in them, and recommendation engines at Amazon and Netflix learned what you might buy or watch next.
• Self-driving cars. In the 2004 DARPA Grand Challenge, no robot car finished the desert course; the best managed about 12 km. In 2005 five did, and Stanford’s car Stanley won.
• Deep learning gets its name. In 2006 Geoffrey Hinton and colleagues showed new ways to train networks with many layers, and the field began calling it “deep learning”.
• ImageNet (2009). Fei-Fei Li’s team released a dataset of millions of labelled photos, and from 2010 an annual contest to recognise them. It gave the field a shared scoreboard.
• In 2011 IBM’s Watson beat the two greatest Jeopardy! champions on TV, and Apple put Siri in the iPhone 4S. AI had moved into living rooms and pockets.

Behind the scenes, the internet was producing more data than anyone had ever had, and the graphics chips (GPUs) built for video games turned out to be extremely good at the maths neural networks need. The ingredients were in place.

2012: the deep learning big bang

In 2012 a neural network called AlexNet, from Hinton’s lab in Toronto (Alex Krizhevsky, Ilya Sutskever and Hinton), entered the ImageNet contest. Trained on two gaming GPUs, it got 15.3% of its top-five guesses wrong, against 26.2% for the next-best entry. In a field used to inching forward, that was a landslide, and almost overnight the whole AI world switched to deep learning.

The same year, a Google team showed a giant network 10 million YouTube thumbnails, with no labels, on 16,000 processor cores, and it learned by itself to recognise cats. What followed was a decade of progress that kept surprising the experts:

• 2014: GANs (generative adversarial networks), two networks competing to create and spot fake images, the start of AI-generated pictures. Google bought the London lab DeepMind.
• 2015: a Microsoft network called ResNet beat the commonly quoted human error rate on ImageNet. OpenAI was founded in San Francisco.
• 2016: DeepMind’s AlphaGo beat Go champion Lee Sedol 4–1 in Seoul. Go has more possible positions than there are atoms in the universe, and experts had thought a win was a decade away. AlphaGo’s 37th move in game two looked so strange that commentators first took it for a mistake. It was brilliant. The same year, Google switched Google Translate to neural networks and its quality jumped overnight.
• 2017: AlphaZero taught itself chess, shogi and Go from nothing but the rules, in hours, and beat the best programs in the world.

Transformers and the road to ChatGPT (2017–2022)

In June 2017 eight researchers at Google published a paper with a cheeky title: “Attention Is All You Need”. It introduced the transformer, a design that looks at every word in a passage at once and works out which words matter to which (that’s the “attention”). It trained far faster on GPUs than earlier designs and kept getting better the bigger you made it. It’s the “T” in GPT, and it’s inside almost every AI tool you use today.

• 2018: Google’s BERT showed a new way to understand language (it made Google Search much better at understanding questions from 2019). OpenAI’s first GPT (Generative Pre-trained Transformer) learned language by predicting the next word across thousands of books.
• 2019: GPT-2 wrote such convincing text that OpenAI at first released it in stages, worried about fake news. That seems quaint now.
• 2020: GPT-3, with 175 billion adjustable connections, could write, summarise and even code from a few examples in plain English. The same year, DeepMind’s AlphaFold 2 predicted the 3D shapes of proteins almost as accurately as lab experiments, solving a 50-year-old problem in biology. By 2022 it had predicted the structures of more than 200 million proteins, nearly every one known to science, free for any researcher.
• 2021–22: AI started making pictures from a sentence: DALL·E, then Midjourney and Stable Diffusion. GitHub Copilot began suggesting code to programmers as they typed.
• 2022: researchers showed that training language models on human feedback about which answers were better (a method called RLHF) made them far more helpful.

Then, on 30 November 2022, OpenAI released ChatGPT as a “research preview”. It had a million users in about five days and an estimated 100 million within two months, one of the fastest-growing consumer apps ever. The technology wasn’t brand new; the chat box was. For the first time, anyone could talk to a capable AI in plain language, for free.

The generative AI era (2023–2026)

Since ChatGPT, the pace has been dizzying. The highlights:

• 2023: OpenAI’s GPT-4 and Anthropic’s first Claude model both arrived in March; Google followed with Bard, then Gemini. Meta released its Llama models with downloadable “open weights” anyone could run. Microsoft launched Copilot for Microsoft 365. Governments woke up: an open letter in March called for a pause on giant AI experiments, and in November 28 countries and the European Union, Australia included, signed the Bletchley Declaration on AI safety at Bletchley Park, where Turing had broken codes during the war.
• 2024: AI could see, hear and speak in real time. The EU AI Act, the first broad AI law, came into force in August. In October, AI took two Nobel Prizes: Physics to John Hopfield and Geoffrey Hinton for the foundations of neural networks, and Chemistry to Demis Hassabis and John Jumper of DeepMind (shared with protein designer David Baker) for AlphaFold. In September, OpenAI’s o1 introduced reasoning models that work through a problem step by step before they answer, and later in the year AI began to operate computers directly, clicking and typing like a person. Australia published a Voluntary AI Safety Standard for businesses (since updated as the Guidance for AI Adoption).
• 2025: China’s DeepSeek released a reasoning model in January that matched leading US models at a fraction of the reported cost; Nvidia’s share price fell about 17% in a day. The year’s big word was agents: AI that doesn’t just answer but carries out multi-step tasks, such as researching, writing and testing code, or working through a pile of email, using standard connectors to reach a business’s tools. In December the Australian Government’s National AI Plan dropped earlier plans for mandatory AI guardrails, choosing to rely on existing laws and regulators, backed by a new AI Safety Institute.
• 2026: AI agents are becoming normal workplace tools, working alongside staff with a person approving what matters. In July the Australian Government changed course, announcing a new Office of AI and a national framework of AI standards, some of them mandatory, with legislation expected in early 2027. Our team works this way every day: this article was researched and drafted with AI, then checked by people before it went live. (We wrote about that in My AI is writing to your AI.)

Whether today’s AI truly understands anything is a separate, lively debate. We covered it in Does AI really understand, or is it just autocomplete?

Five lessons from 70 years of AI

1. The ideas are old; the fuel is new. Artificial neurons date from 1943, backpropagation from the 1970s and 80s, neural networks that read handwriting from 1989. What changed after 2012 was data (the internet), computing power (GPUs) and the money to combine them. Many “overnight” breakthroughs took 40 years.

2. Hype and winters come in cycles. Every era overpromised: thinking machines in 20 years (1960s), expert systems in every company (1980s). Each time, the useful parts survived the crash. It’s sensible to expect bumps in today’s boom too, without assuming the tools will disappear. They won’t.

3. Once it works, we stop calling it AI. Route planning, spam filtering, speech-to-text and face unlock were all AI research once. Now they’re just features. Researchers call this the AI effect: as the computer scientist Larry Tesler put it, intelligence is whatever machines haven’t done yet.

4. Learning beat rules. Decades of hand-coding expert knowledge lost out to methods that learn from data and scale with computing power. The AI researcher Rich Sutton called this “the bitter lesson” in 2019. It’s why today’s AI is good at things nobody explicitly taught it, and why it can be confidently wrong in ways nobody predicted.

5. The hard part was never the chess. Hans Moravec noticed in the 1980s that computers find “hard” things like chess and maths easy, and “easy” things like walking, folding laundry or picking up a cup very hard. That’s still true: AI can draft a legal summary in seconds, while a robot that tidies a kitchen remains a research project.

And one lesson from ELIZA in 1966 that hasn’t aged at all: people trust fluent machines more than they should. Check the important things.

What it means for your business

You don’t need to know the history to use AI well, but it helps to keep a few things in mind:

• Expect the tools to keep changing, fast. Pick tools that fit how your team already works, and plan to review them every six to twelve months rather than betting the business on one product.
• Keep a person in the loop for anything that matters. AI is brilliant at first drafts and at grunt work. Its mistakes look exactly like its successes, so a human should check anything that goes to a client, touches money or is hard to undo.
• Set some ground rules. Decide what data staff may put into which AI tools. Our AI usage policy guide has a plain-English starting point.
• Invest in people, not just licences. The slowest part of AI adoption isn’t the technology, it’s us. (More in Keeping up with AI.) A few hours of practical training usually pays for itself quickly.

If you’d like help working out where AI fits in your business, our AI training is hands-on and in plain English, for small teams and big ones. Get in touch and we’ll talk it through.

Common questions

Sources

All checked 14/10/2026.

• Stanford Report, Ancient myths reveal early fantasies about artificial life (Adrienne Mayor on Talos, 2019). news.stanford.edu
• MacTutor, Ada Lovelace’s notes (Note G, 1843). mathshistory.st-andrews.ac.uk
• A. M. Turing, Computing Machinery and Intelligence, Mind, vol. 59 (October 1950). academic.oup.com
• CSIRO, Our first computer turns 75 today (CSIRAC, November 2024). csiro.au
• Museums Victoria, Australia’s first computer finds new home at Scienceworks. museumsvictoria.com.au
• Quote Investigator, Herbert Simon’s “any work that a man can do” (1960). quoteinvestigator.com
• Quote Investigator, Marvin Minsky’s “within a generation” (1967). quoteinvestigator.com
• Buchanan and Shortliffe, Rule-Based Expert Systems: The MYCIN Experiments (1984), chapter 31. columbia.edu
• IEEE Spectrum, SRI’s pioneering mobile robot Shakey honored as IEEE Milestone. spectrum.ieee.org
• Yann LeCun, biography (cheque-reading systems). yann.lecun.com
• Encyclopædia Britannica, Deep Blue. britannica.com
• He, Zhang, Ren and Sun, Deep Residual Learning for Image Recognition (2015). arxiv.org
• Vaswani et al., Attention Is All You Need (2017). arxiv.org
• Rich Sutton, The Bitter Lesson (2019). incompleteideas.net
• The Nobel Prize in Physics 2024 and in Chemistry 2024. nobelprize.org
• UK Government, The Bletchley Declaration (1 November 2023). gov.uk
• Department of Industry, Science and Resources, Voluntary AI Safety Standard. industry.gov.au
• Department of Industry, Science and Resources, National AI Plan (2 December 2025). industry.gov.au
• Department of the Prime Minister and Cabinet, Office of AI. pmc.gov.au
• Baker McKenzie, Australia charts a new course on AI regulation and policy (August 2026). bakermckenzie.com
• Larry Tesler, Adages and coinages (“Intelligence is whatever machines haven’t done yet”). nomodes.com

Photographs and images (all via Wikimedia Commons, used under the licence shown, checked 14/10/2026)

• TONIODELBARRIO6464, “Vase de Talos à Ruvo di Puglia”, Wikimedia Commons, CC BY-SA 4.0 (resized). file page
• Farrukh ibn Abd al-Latif (copyist), “Al-Jazari - A Musical Toy”, Wikimedia Commons, public domain (resized). file page
• after Wolfgang von Kempelen, “Tuerkischer schachspieler windisch4”, Wikimedia Commons, public domain (resized). file page
• Alfred Edward Chalon, “Ada Lovelace portrait”, Wikimedia Commons, public domain (cropped and resized). file page
• Daderot, “Analytical engine trial model, Charles Babbage, 1834-1871 - Science Museum, London - DSC05565”, Wikimedia Commons, CC0 (resized). file page
• Doug butler, “CSIRAC at Scienceworks, Melbourne”, Wikimedia Commons, CC BY-SA 4.0 (resized). file page
• null0, “John McCarthy (computer scientist) Stanford 2006 (272020300)”, Wikimedia Commons, CC BY 2.0 (resized). file page
• John C. Hay and Albert E. Murray, “Mark I Perceptron, Figure 2 of operator's manual”, Wikimedia Commons, public domain (resized). file page
• “ELIZA conversation”, Wikimedia Commons, public domain (resized). file page
• The wub, “SRI Shakey robot, 1969, Computer History Museum”, Wikimedia Commons, CC BY-SA 4.0 (resized). file page
• Larry D. Moore, “Roomba original”, Wikimedia Commons, CC BY 4.0 (resized). file page
• Anton Chiang, “IBM Deep Blue at Computer History Museum (9361685537)”, Wikimedia Commons, CC BY 2.0 (resized). file page
• DARPA, “Stanley2”, Wikimedia Commons, public domain (resized). file page
• DeFacto, “Bletchley Park Mansion”, Wikimedia Commons, CC BY-SA 4.0 (resized). file page
• Arthur Petron, “Geoffrey E. Hinton, 2024 Nobel Prize Laureate in Physics (cropped)”, Wikimedia Commons, CC BY-SA 4.0 (resized). file page

Diagrams by Digital Greenhouse. The timeline is an illustration of the story, not measured data.