In September 2026, OpenAI began rolling out GPT-6 to business testers, its latest step towards what the company calls artificial general intelligence. Seventy-six years earlier, a British mathematician had asked whether a machine could think, and then, sensibly, declined to answer the question as phrased.
Between those two moments lies one of the strangest stories in science: a field that kept promising too much, lost its funding, kept working anyway, and was finally rescued by more data, faster chips and a better way of paying attention.
Artificial intelligence is the branch of computer science that builds systems to perform tasks we associate with human thinking, such as recognising images, understanding language and generating text or pictures, usually by learning patterns from data rather than following hand-written rules.
The definition is tidy. The history is not.
1950 and 1956: a question and a name
Alan Turing’s paper “Computing Machinery and Intelligence” appeared in the philosophy journal Mind in 1950. It opens with “Can machines think?” and then argues that both “machine” and “think” are too slippery to define. So Turing swapped the question for a party game.
In his imitation game, a judge exchanges written messages with two hidden players and tries to work out which is which. Replace one of them with a computer, and the test becomes: can the judge tell the difference? Turing predicted that, by the end of the century, machines would play the game well.
The move was clever, and it has haunted the field ever since. Turing made intelligence a matter of performance, of what an observer believes, rather than of what is happening inside the box. Remember that.
Five years later, on 2 September 1955, four researchers, John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon, sent a proposal for “a 2-month, 10-man study of artificial intelligence” at Dartmouth College in New Hampshire the following summer. Its central conjecture was bold. Every aspect of learning or intelligence could, in principle, be described so precisely that a machine could simulate it.
The 1956 workshop itself was loose. People drifted in and out; only three stayed the whole time. But the proposal did something lasting. It gave the field its name.
McCarthy chose “artificial intelligence” partly to steer clear of rival labels such as automata theory and cybernetics, each with its own agenda and its own gatekeepers. It was, in effect, a branding decision: a neutral, ambitious, slightly science-fictional name that nobody else owned. It has outlived almost every technical idea discussed that summer.
Why did early AI promise so much?
Because early results looked like magic, and because the people producing them said so.
In 1958, after a press event organised by the US Navy, The New York Times described Frank Rosenblatt’s perceptron, a simple learning machine, as the embryo of a computer that might one day walk, talk, see and be conscious. The hardware version, the Mark I Perceptron, was built to recognise images and first demonstrated publicly in June 1960.
Through the 1960s, leading researchers predicted that machines would match human abilities within a generation.
Meanwhile, at MIT, Joseph Weizenbaum built ELIZA between 1964 and 1967. It imitated a therapist by turning users’ words back into questions, and it understood nothing. Yet Weizenbaum’s own secretary reportedly asked him to leave the room so she could speak to it in private.
Turing’s point about observers suddenly looked less academic. Weizenbaum was disturbed by how quickly ordinary people attributed understanding to a few hundred lines of pattern matching, and the tendency now bears his program’s name: the ELIZA effect.
For anyone who builds brands, this is the most important idea in the whole history. People do not judge intelligence. They judge the experience of being understood.
The first winter
The bills arrived in the late 1960s and early 1970s.
In the US, a 1966 committee report concluded that machine translation was slower, less accurate and more expensive than human translators, and government support was cut. In 1969 Minsky and Seymour Papert published Perceptrons, which showed that single-layer networks could not learn even simple functions such as exclusive-or.
The book is often misread as dismissing all neural networks. Its effect was real all the same: interest and money drained out of the approach for more than a decade.
Then, in 1973, a report for the UK government by the mathematician James Lighthill criticised AI for failing to meet its “grandiose objectives”. British AI research was largely dismantled. In America, military research funding now had to serve specific missions, and open-ended AI work struggled to find backers. The period from roughly 1974 to 1980 became the first AI winter.
Why did the second AI winter happen?
Because the thaw was built on rules, and rules are brittle.
A boom in expert systems
The recovery came from a more modest product: the expert system. Instead of trying to build a general mind, engineers interviewed specialists and encoded their knowledge as thousands of if-then rules.
It worked, up to a point. XCON, developed at Carnegie Mellon for Digital Equipment Corporation to configure computer orders, was estimated to have saved the company tens of millions of dollars. Corporations built AI departments. A hardware industry grew up around specialised Lisp machines.
Asia, not for the last time, made the boldest bet. In the early 1980s Japan’s Ministry of International Trade and Industry launched the Fifth Generation Computer Systems project, a national programme to build machines that could converse, translate and reason. It frightened Western governments into launching programmes of their own.
And the bust
An expert system knew exactly what it had been told and nothing else. Every new exception meant another expensive round of interviews and code. In 1984, at the annual meeting of the American Association for Artificial Intelligence, Roger Schank and Marvin Minsky warned that hype was outrunning results and gave the coming crash a name: an “AI winter”.
Within three years they were proved right. In 1987 the Lisp machine market collapsed as cheaper desktop workstations caught up. Expert systems proved too costly to maintain. Japan’s Fifth Generation project closed in 1992 without reaching its goals.
The name that had once attracted money was now a liability. Much of the work carried on under humbler labels, machine learning among them.
The quiet decades, and the year the pictures changed
While the label was out of fashion, the work went on, and some of it was public. In May 1997 IBM’s Deep Blue beat the world chess champion Garry Kasparov in New York, 3½ to 2½. It was brute-force search rather than learning, but it put machine intelligence back on front pages.
The more important shift happened in the background. The internet was producing data at a scale nobody had seen. Graphics chips built for video games turned out to be very good at the arithmetic neural networks needed. And a small group of researchers, among them Geoffrey Hinton, Yann LeCun and Yoshua Bengio, kept working on neural networks long after it was fashionable.
The turning point had a dull name: the ImageNet Large Scale Visual Recognition Challenge. In 2012 a University of Toronto team, Alex Krizhevsky, Ilya Sutskever and Hinton, entered a deep neural network trained on two consumer Nvidia graphics cards. AlexNet’s top-five error rate was 15.3 per cent, more than ten percentage points ahead of the runner-up.
Ten points in a benchmark competition is not an improvement. It is a change of era.
The three researchers formed a small company and sold it to Google. Within a few years, deep learning was running photo search, speech recognition and translation for billions of people. Hinton, LeCun and Bengio shared the 2018 Turing Award, and in 2024 Hinton shared the Nobel Prize in Physics with John Hopfield for foundational work on neural networks.
2016 and 2017: a move and a mechanism
Move 37
If 2012 convinced engineers, March 2016 convinced everyone else. It happened in Seoul.
At the Four Seasons Hotel, AlphaGo, from the Google-owned lab DeepMind, played the South Korean champion Lee Sedol in a five-game match of Go, a game long thought too intuitive for computers. AlphaGo won four games to one. In the second game it played a move, the now-famous Move 37, that professional commentators called creative and unlike anything a human would consider. Chinese-language coverage of the first game alone reached around 60 million viewers.
DeepMind’s own documentary on the Seoul match, Move 37 and Lee Sedol’s response to losing to a machine.
In November 2019, Lee retired from professional play. His reason: AI had become an entity that could not be defeated. Few products have received a more sobering testimonial.
Attention, please
The architecture behind today’s AI arrived in a 2017 paper by eight Google researchers with the cheerful title “Attention Is All You Need”. It introduced the transformer. Earlier language models processed text step by step; the transformer used a mechanism called attention, letting a model weigh every word in a passage against every other word at once.
Better still for the economics, transformers could be trained in parallel on vast clusters of chips. Feed one enough text and enough computing power, and it picks up grammar, facts, style and a surprising amount of reasoning as by-products of predicting the next word. By 2026 the paper had been cited more than 250,000 times.
DeepMind, meanwhile, showed what learning systems could do in science. Its AlphaFold 2 topped the CASP14 protein-structure competition in 2020, and in 2024 Demis Hassabis and John Jumper shared half of the Nobel Prize in Chemistry for the work.
30 November 2022 and after
ChatGPT launched on 30 November 2022 as a low-key research preview. It reached 100 million monthly users in about two months. By December, Google’s leadership had reportedly declared a “code red” over the threat to search.
Image and video generators turned text prompts into pictures. Open-weight models spread.
In January 2025 the Chinese start-up DeepSeek, funded by the hedge fund High-Flyer, released its R1 reasoning model. Within a week it had overtaken ChatGPT as the most downloaded free app on the US App Store, and Nvidia’s share price fell sharply on fears that frontier AI might need fewer expensive chips than investors had assumed.
Southeast Asia has been building its own. AI Singapore, a national programme hosted by NUS, keeps releasing SEA-LION, an open family of models built for eleven or more Southeast Asian languages; the latest version arrived in September 2026.
By February 2026, ChatGPT reported around 900 million weekly users. The European Union’s AI Act, in force since August 2024, has been phasing in obligations, starting with bans on the riskiest uses in 2025.
Is this boom another bubble?
Anyone who works with these tools ends up asking, and the arguments are well rehearsed. Sceptics point to start-ups that are thin wrappers around someone else’s model, and to capital poured into data centres. Defenders point to tasks that used to take a day and now take two hours.
Engineers who lived through the 1980s add a twist. Today’s systems are far more flexible than expert systems ever were. They are also far more willing to produce a fluent, confident answer that happens to be wrong.
The field has a habit of repeating its mistakes: assuming progress on narrow tasks is a step towards general intelligence, assuming what is hard for people is hard for machines, and describing software with human words such as “understand” and “learn” that quietly inflate what it does.
The honest answer? AI is now useful enough that a full winter looks unlikely, and hyped enough that a sharp correction would surprise nobody.
What seventy-six years of AI teach brands
Read it as a reputation story and four lessons stand out.
Names do heavy lifting. “Artificial intelligence” won funding in 1956, became a liability after the second winter and is now stamped on everything from phones to fridges. Labels attract attention, then attract scepticism. Use the phrase only where it describes something a customer can actually feel.
Overpromising has a half-life. Every winter followed a gap between what was claimed and what was delivered. Brands that market AI features should plan for the moment a customer tests the claim.
The ELIZA effect cuts both ways. People warm to systems that seem to understand them, and turn on them when the illusion breaks. If your chatbot speaks in your brand’s voice, its mistakes are your brand’s mistakes.
And AI is now one of your audiences. Answer engines such as ChatGPT and Google’s AI Overviews summarise companies for millions of people, and what they say depends on what the web consistently says about you. Clear, consistent facts are the new shelf placement.
Quick answers for the boardroom
When did AI start?
The field is usually dated to the 1956 Dartmouth summer workshop, whose 1955 proposal coined the term “artificial intelligence”. It built on Alan Turing’s 1950 paper “Computing Machinery and Intelligence”.
What were the AI winters?
AI winters were periods of collapsed funding and interest after overpromising, roughly 1974 to 1980 and from 1987 into the 1990s. They followed reports such as Lighthill’s in 1973 and the collapse of the expert-system and Lisp machine markets.
What made modern generative AI possible?
Large datasets, powerful graphics chips and the transformer architecture, introduced by Google researchers in 2017, made it possible to train very large language models. ChatGPT brought them to the public in November 2022.
The machines are now describing your brand to your customers, so it is worth deciding what they should say. We can help you shape it.
