History of Artificial Intelligence: From Turing to ChatGPT

The origins: 1943–1955

Even before the term “artificial intelligence” existed, mathematicians and neurologists were asking whether thought could be described through calculation.

1. 1943The first artificial neuron

   Neurologist Warren McCulloch and logician Walter Pitts publish a simplified mathematical model of the neuron. It is the ancestor of today’s neural networks.

2. 1950Alan Turing asks the question

   In the paper “Computing Machinery and Intelligence,” British mathematician Alan Turing proposes a test, now called the Turing test: can a machine hold a conversation without the person on the other end being able to tell it apart from a human?

3. 1955–1956The Dartmouth workshop

   John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon propose a summer workshop at Dartmouth College (United States) in 1955, held in 1956. John McCarthy popularizes the term “artificial intelligence” there. The birth of the discipline is traditionally dated to this event.

First successes and first disappointments: 1956–1980

1. 1958The perceptron

   American psychologist Frank Rosenblatt introduces the perceptron, a single-layer neural network able to learn to classify simple patterns.

2. 1966ELIZA, the first “chatbot”

   Joseph Weizenbaum, at MIT, publishes ELIZA, a program that imitates a psychotherapist by rephrasing the user’s sentences. He is surprised to see people attributing understanding and emotions to it: this has since been called the “ELIZA effect.”

3. 1969The book “Perceptrons”

   Marvin Minsky and Seymour Papert demonstrate the limits of single-layer perceptrons. The book helps steer part of the research community away from neural networks for about a decade.

4. 1973–1974The first “AI winter”

   In the United Kingdom, the Lighthill report (1973) takes a harsh view of the results of AI research; public funding drops. The early promises, judged excessive, had not been kept.

Expert systems and the return of neural networks: 1980–2000

1. 1980sThe rise of expert systems

   Programs based on rules written by specialists (in medicine, in engineering) are adopted by companies. The AI industry grows from a few million to several billion dollars. But these systems are costly to maintain and inflexible.

2. 1982Hopfield networks

   John Hopfield shows that a certain type of neural network can memorize patterns and retrieve them. This work will earn him the Nobel Prize in Physics in 2024.

3. 1986Backpropagation popularized

   David Rumelhart, Geoffrey Hinton, and Ronald Williams publish in Nature an effective method for training multi-layer neural networks: gradient backpropagation. Neural networks come back into favor.

4. Late 1980sThe second AI winter

   The expert systems market collapses, and funding dries up once again.

5. 1989–1990Convolutional networks

   Yann LeCun applies convolutional neural networks to handwritten digit recognition; the technique will be used to read bank checks.

6. 1997Deep Blue beats Kasparov

   In May 1997, IBM’s Deep Blue supercomputer beats world chess champion Garry Kasparov (3.5–2.5 over six games). Deep Blue does not “learn”: it uses brute force to explore hundreds of millions of positions per second, with rules tuned by experts. That same year, Sepp Hochreiter and Jürgen Schmidhuber publish LSTM networks, which were important for language processing before Transformers.

The deep learning revolution: 2009–2016

1. 2009ImageNet

   Fei-Fei Li and her colleagues introduce ImageNet, a huge database of millions of annotated images. It will serve as the proving ground for image recognition.

2. 2011Watson wins at Jeopardy!

   IBM’s Watson system beats two champions of the American TV quiz show Jeopardy!, a demonstration of natural language processing in front of the general public.

3. 2012AlexNet: the spark

   Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton win the ImageNet competition with AlexNet, a deep network trained on two graphics cards: a 15.3% error rate versus 25.8% for the runner-up. GPUs, large datasets, and better training techniques come together: the deep learning era begins.

4. 2013–2014Words as numbers, generated images

   In 2013, Tomas Mikolov (Google) publishes word2vec, which represents words as vectors of numbers. In 2014, Ian Goodfellow invents generative adversarial networks (GANs), the first major family of models able to generate realistic images.

5. March 2016AlphaGo beats Lee Sedol

   In Seoul, from March 9 to 15, Google DeepMind’s AlphaGo program beats champion Lee Sedol 4 to 1 at the game of Go, long considered far more complex than chess. AlphaGo had learned by studying about 30 million moves from professional players, then by playing against itself. Its 37th move, in the second game, surprised every commentator.

Transformers and the explosion of generative AI: 2017–2022

1. June 12, 2017“Attention Is All You Need”

   Eight Google researchers (including Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan Gomez, Łukasz Kaiser, and Illia Polosukhin) publish the Transformer architecture. It replaces sequential networks with an “attention” mechanism that can be processed in parallel. It is the foundation of almost all of today’s large language models (see how ChatGPT works).

2. 2018GPT and BERT

   OpenAI publishes GPT-1 (June); Google publishes BERT (October). The principle: pre-train a large model on raw text, then adapt it to specific tasks.

3. February 2019GPT-2

   OpenAI announces GPT-2 (1.5 billion parameters), capable of generating credible text. Training reportedly cost about $50,000, according to published estimates.

4. May 28, 2020GPT-3

   OpenAI introduces GPT-3: 175 billion parameters, trained on about 300 billion text “tokens” (drawn from a corpus of roughly 500 billion). The model impresses with its ability to perform tasks without specific training, from just a few examples in the prompt.

5. 2020–2021AlphaFold and science

   DeepMind’s AlphaFold 2 AI predicts the 3D shape of proteins with unprecedented accuracy. Its creators, Demis Hassabis and John Jumper, will share the 2024 Nobel Prize in Chemistry with David Baker.

6. 2021–2022Generative imagery reaches the general public

   OpenAI introduces DALL-E (2021), then DALL-E 2 (2022). Midjourney opens its public beta in the summer of 2022. Stable Diffusion, released openly in August 2022, democratizes image generation (see generative AI for images).

7. November 30, 2022ChatGPT launches

   OpenAI opens ChatGPT, a conversational assistant based on GPT-3.5, to the public. It passes one million users in five days and 100 million in about two months, making it one of the fastest-growing consumer services in history.

AI today: 2023–2026

1. 2023The race for models

   GPT-4 (March); first public launches of Claude (Anthropic, March); open models from Meta (LLaMA); creation of Mistral AI, a French startup founded on April 28, 2023, by Arthur Mensch, Guillaume Lample, and Timothée Lacroix, which releases Mistral 7B in September.

2. 2024Regulation and scientific recognition

   The European AI Act enters into force on August 1, 2024. In October, two Nobel Prizes recognize work related to AI: Physics (John Hopfield and Geoffrey Hinton) and Chemistry (David Baker, Demis Hassabis, John Jumper).

3. January 20, 2025DeepSeek-R1

   Chinese company DeepSeek releases R1, an open-weight reasoning model rivaling the best American models. The announcement triggers a drop in Nvidia’s stock and reignites the debate over the cost of training.

4. 2025–2026AI that reasons and acts

   Assistants can now search the web, analyze documents, write code, and chain actions together (what we call “agents”). According to Stanford University’s 2026 AI Index, 88% of surveyed organizations already use AI, and the performance gap between the best American and Chinese models has become very small.

To find out what lies behind these models, see how an AI is built. To learn about the limits and the rules, see limits, risks, and regulation.

What this history teaches us

• Ideas come before technology. Neural networks date from 1943 and backpropagation from 1986: they only took off in 2012, once data and computing power were sufficient.
• Enthusiasm is cyclical. Two “AI winters” are a reminder to be wary of excessive promises.
• The pace is accelerating. Less than six years passed between the Transformer (2017) and ChatGPT (2022).

Frequently asked questions