Level : Beginner
A brief history of artificial intelligence, from Turing to autonomous AI agents
By The Info Agent IA editorial team ·
Artificial intelligence was officially born in the summer of 1956, but its ideas go back to the 1940s, and its history swings between big hopes, long disappointments and sudden breakthroughs. From the first artificial neurons to assistants like ChatGPT, and now AI agents that can act on their own, here are the milestones that really matter, without the jargon.
Before the name: artificial neurons and thinking machines
1943: a neuron on paper
It all starts with a mathematical idea. In 1943, Warren McCulloch and Walter Pitts publish a highly simplified model of a neuron: a small unit that receives signals and decides whether to pass a signal on. It is the distant ancestor of the neural networks that power today’s assistants.
1950: Alan Turing asks the question
In October 1950, British mathematician Alan Turing publishes “Computing Machinery and Intelligence” in the journal Mind. Rather than argue endlessly over whether machines can think, he proposes an experiment: the “imitation game.” A judge chats in writing with a human and a machine, and has to guess which is which. This is the famous Turing test.
1956: Dartmouth gives AI its name
On August 31, 1955, four researchers, John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon, sign a research proposal. In it, they use the phrase “artificial intelligence” and announce a summer workshop at Dartmouth College in New Hampshire.
The workshop takes place in the summer of 1956. Its starting assumption is bold: every aspect of learning or intelligence could, in principle, be described so precisely that a machine could simulate it. This moment is widely seen as the birth of the field.
The early optimism (1956-1973)
The next two decades are marked by huge optimism. Two broad families of ideas exist side by side.
- Symbolic AI: you write rules and manipulate symbols, the way a logician would. Programs prove theorems or solve puzzles.
- Learning: you let the machine improve from examples. In 1958, Frank Rosenblatt unveils the perceptron, a neural network that can learn to recognize simple shapes. In 1959, Arthur Samuel at IBM describes a checkers program that gets better by playing, and talks about “machine learning.”
In 1966, Joseph Weizenbaum at MIT presents ELIZA, a program that mimics a psychotherapist by rephrasing what the user types. The system is very basic, yet some users credit it with real understanding. That human reflex is still with us when we talk to today’s chatbots.
The AI winters
The promises run ahead of reality. In 1969, Marvin Minsky and Seymour Papert publish Perceptrons, a book showing the limits of single-layer neural networks. Interest in that approach fades for years.
In 1973, in the United Kingdom, the Lighthill report delivers a harsh verdict on AI research. Funding drops on both sides of the Atlantic: this is the first AI winter, in the mid-1970s.
The 1980s bring a revival with expert systems. These programs encode specialists’ knowledge as “if… then…” rules, to diagnose a fault or configure a computer. Companies invest heavily. But the systems are expensive to maintain and brittle outside their narrow field. By the late 1980s, the market collapses: this is the second AI winter.
Meanwhile, an older idea makes a comeback. In 1986, David Rumelhart, Geoffrey Hinton and Ronald Williams popularize backpropagation, a method for training neural networks with several layers. It would become the foundation of deep learning.
1997: Deep Blue and the triumph of raw computing
On May 11, 1997, IBM’s Deep Blue computer defeats world chess champion Garry Kasparov in a match. The news makes headlines worldwide. Yet Deep Blue learns almost nothing: it searches through enormous numbers of possible moves at high speed, guided by evaluation rules written by people.
In February 2011, another IBM system, Watson, wins the US quiz show Jeopardy! against former champions. These spectacular wins remain specialized feats.
Quiet AI, everywhere
Away from the spotlight, the 1990s and 2000s see AI settle quietly into everyday tools. The term “AI” has fallen out of fashion, and researchers prefer to talk about statistics or machine learning. Yet this is when spam filters learn to sort email from examples, search engines rank web pages, online stores recommend products and handwriting recognition helps read addresses and checks.
In October 2011, Apple introduces Siri with the iPhone 4S: an assistant you talk to. These uses prepare the ground. They get people used to machines that “understand” a request, even in a limited way.
2012: the deep learning breakthrough
The real turning point comes in 2012. At the ImageNet image recognition competition, Geoffrey Hinton’s team, according to the ACM, cuts the error rate almost in half using a deep neural network. The ideas were old; what changed was the combination of three ingredients: masses of digital data, suitable computer chips (GPUs) and networks with many layers.
Deep learning then takes over computer vision, followed by speech recognition and translation. In March 2016, DeepMind’s AlphaGo program beats Lee Sedol, one of the world’s best Go players, 4 games to 1. Go had long been considered too complex for a machine.
2017: the Transformer changes everything
In June 2017, Google researchers publish “Attention Is All You Need.” The paper introduces the Transformer, an architecture that processes all the words in a text in parallel and works out which ones matter most to each other. It was first aimed at machine translation.
By 2018, OpenAI with GPT and Google with BERT show that a Transformer trained on vast amounts of text can then be used for many different tasks. This marks the birth of large language models, or LLMs, which we explain in detail in our guide to LLMs.
November 30, 2022: ChatGPT brings AI to everyone
On November 30, 2022, OpenAI launches ChatGPT. For the first time, anyone can chat with a large language model, for free, in a simple chat window. AI leaves the lab and enters daily life: drafting an email, summarizing a document, planning a trip. To understand what happens when you ask it a question, read how ChatGPT works.
Competitors follow fast. In March 2023, Google launches Bard, later renamed Gemini, and OpenAI unveils GPT-4. In Europe, Regulation (EU) 2024/1689, known as the AI Act, enters into force on August 1, 2024: the first major legal framework dedicated to AI.
The data confirms the acceleration. According to Our World in Data (figures retrieved October 22, 2026), global investment in generative AI rose from 3.7 billion dollars in 2022 to 25.4 billion in 2023, then to 143.8 billion in 2025.
What the data shows: a race between the US and China
The research institute Epoch AI tracks the AI models it considers “notable.” According to its documentation, consulted on October 22, 2026, these are models that improved the state of the art on a recognized benchmark, are highly cited, have historical relevance or see significant use. Here is how many there are per year, by the country of the first organization listed.
| Year | United States | China | United Kingdom | South Korea | Canada | Other |
|---|---|---|---|---|---|---|
| 2012 | 5 | 0 | 0 | 0 | 6 | 2 |
| 2014 | 13 | 2 | 7 | 0 | 4 | 7 |
| 2016 | 32 | 4 | 5 | 0 | 1 | 7 |
| 2017 | 23 | 9 | 10 | 3 | 3 | 7 |
| 2019 | 51 | 5 | 7 | 1 | 0 | 4 |
| 2021 | 49 | 18 | 4 | 1 | 1 | 6 |
| 2022 | 61 | 10 | 13 | 3 | 1 | 3 |
| 2023 | 81 | 23 | 2 | 1 | 1 | 11 |
| 2024 | 55 | 30 | 1 | 1 | 1 | 10 |
| 2025 | 62 | 34 | 1 | 8 | 0 | 2 |
| 2026 (as of October 22) | 49 | 29 | 0 | 5 | 0 | 1 |
Source: Epoch AI, “Data on AI Models,” retrieved October 22, 2026. Selected years.
Three trends stand out. The United States has clearly led since the mid-2010s. China gains ground from 2021 and reaches or exceeds 30 notable models a year in 2024 and 2025. The United Kingdom, very present during the rise of deep learning, appears far less often since 2023.
Capabilities have moved just as fast. On GPQA, a test of PhD-level science questions designed to resist a simple web search, the best score recorded by Epoch AI rose from 35.7 out of 100 (GPT-4, first quarter of 2023) to 96.4 (Gemini 4 Argon, third quarter of 2026).
The age of AI agents (since 2024)
The latest stage is no longer just about answering, but about acting. An AI agent receives a goal, breaks the work into steps, uses tools (a browser, files, software) and checks its results.
A few milestones:
- November 25, 2024: Anthropic announces the Model Context Protocol, an open protocol for connecting assistants and agents to tools and data.
- January 23, 2025: OpenAI introduces Operator, an agent that can use a web browser to carry out tasks.
- 2025-2026: the major assistants add agent modes, and tools for building your own agents multiply.
This is a deep shift. A classic assistant produces text that you reread before using it. An agent can click, fill in a form or send a message on your behalf. Mistakes therefore have more direct consequences, and trust becomes the central question: what should you hand over, with which access rights, and how do you check what it did?
As of October 22, 2026, here is the top of the model ranking built from Epoch AI’s benchmarks. The models powering these agents are now replaced within months.
| Rank | Model | Organization | Index |
|---|---|---|---|
| 1 | Claude Sonnet 5.5 | Anthropic | 95.9 |
| 2 | Claude Opus 5.5 | Anthropic | 95.9 |
| 3 | Gemini 4 Argon | Google DeepMind | 95.3 |
| 4 | GPT-6 Astra | OpenAI | 94.7 |
| 5 | Claude Opus 5 | Anthropic | 92.7 |
Check yourself Which of these came first: the Turing test, the Dartmouth workshop or the first artificial neuron model?
The artificial neuron model by Warren McCulloch and Walter Pitts, published in 1943. Next comes Alan Turing’s paper in 1950, then the Dartmouth workshop in the summer of 1956.
Key takeaways
Remember, too, that the big breakthroughs are rarely new ideas. Neural networks date from 1943; they waited nearly seventy years for the data and computing power they needed. For a refresher on the basics, our beginner’s guide to artificial intelligence covers the essentials. And to explore every date and every notable model since 2012, open the interactive AI timeline.
Frequently asked questions
When was artificial intelligence invented?
Its birth is usually dated to the Dartmouth workshop in summer 1956, set up by a proposal dated August 31, 1955, in which the term artificial intelligence first appears. Earlier work paved the way, such as the artificial neuron model of 1943 and Alan Turing's 1950 paper.
Who is the father of artificial intelligence?
There is no single father. Alan Turing raised the question of thinking machines in 1950, and John McCarthy proposed the name of the field with Marvin Minsky, Nathaniel Rochester and Claude Shannon. More recently, Yoshua Bengio, Geoffrey Hinton and Yann LeCun are often described as the pioneers of deep learning.
What is an AI winter?
It is a period when funding and enthusiasm for AI collapse after unkept promises. Two are usually counted: one in the mid-1970s, and another in the late 1980s and early 1990s. They are a reminder that AI progress has never been a straight line.
Why has AI improved so much since 2012?
Three ingredients came together: deep neural networks, huge amounts of digital data and suitable computer chips, the GPUs. The Transformer architecture, published in 2017, then made it possible to build the large language models behind today's assistants.
Where can I see all the key dates in AI history?
Info Agent IA's interactive timeline brings together the major historical milestones and the AI models that the research institute Epoch AI considers notable since 2012. You can filter by field and organization, and zoom in on a period.