Chapter 1 · Lesson 1.1
Definitions and history
Artificial intelligence aims to give machines capabilities that we associate with human intelligence.
L’Artificial Intelligence (AI) refers to the set of theories and techniques developed to create machines capable of simulating human intelligence. It seeks to reproduce cognitive functions: reasoning, learning, problem solving, perception, language understanding…
The boundary between computational software and software with AI capabilities is blurred. In principle, an AI implies a capacity to learn and has no need for explicit coding of the computation of the solution.
Weak AI and strong AI
Weak AI — ANI
Artificial Narrow Intelligence. A system designed for a specific task, with no self-awareness. Voice assistants, recommendation, image recognition, translation.
Strong AI — AGI
Artificial General Intelligence. Understanding, learning and applying knowledge in a general way: abstract reasoning, autonomous learning, creativity, self-awareness (theoretical), adaptability.
All existing AI is weak AI. Strong AI remains theoretical (debate on whether it is possible; estimates range from a few years, according to AI company executives, to several decades or even never, according to other researchers: none is established).
Historical milestones
| Date | Event |
|---|---|
| 19th c. | Ada Lovelace imagines that a machine could one day compose music. |
| 1950 | Alan Turing proposes the imitation game (Turing test). |
| 1956 | The Dartmouth conference: John McCarthy coins the term “artificial intelligence”. |
| 1964–68 | Eliza, one of the first dialogue programs. |
| 1965–1980 | Expert systems: DENDRAL, MYCIN, XCON/R1. |
| 1997 | Deep Blue (IBM) beats world chess champion Garry Kasparov. |
| 2012 | AlexNet wins the ImageNet competition: deep learning takes hold in computer vision. |
| 2016 | AlphaGo (DeepMind) beats Lee Sedol at the game of Go. |
| 2017 | “Attention Is All You Need”: theattention mechanism and Transformers. |
| 2020–2022 | GPT-3 (2020), DALL-E (2021), then ChatGPT (November 2022): large language models reach the general public. |
| 2024 | First reasoning models; Nobel Prizes in Physics (Hopfield, Hinton) and Chemistry (Hassabis, Jumper, Baker) linked to AI; the MCP protocol connects assistants to tools. |
| 2025–2026 | Rise of agents, which act with tools; gradual application of the AI Act in Europe. |
The Turing test
Three entities: two humans and a computer. An interrogator asks questions in writing to a human and to a machine, without knowing which is which. The machine passes the test if it fools the interrogator at least 30% of the time, after 5 minutes of interaction.
In 2025, Jones and Bergen ran this three-party test with recent models, on five-minute conversations: GPT-4.5, which was asked to adopt a persona, was taken for the human in 73% of cases, LLaMa-3.1 in 56%, GPT-4o in 21% and Eliza in 23% (arXiv preprint 2503.23674). This confirms that the test measures imitation: the same model, without a persona, does far worse.
Going further: what does Turing really say?
In his 1950 paper (Computing Machinery and Intelligence), Turing does not set an official “pass threshold”: he predicts that by the end of the 20th century, an average interrogator would have no more than a 70% chance of making the right identification after five minutes of questioning, that is, about 30% deception. The course’s “30% in 5 minutes” takes up this prediction, which has become a criterion in some competitions.
The test measures the ability to imitate a human conversation, not understanding or consciousness: the Eliza program, based on very simple rules, has already given some interlocutors the illusion of understanding.
AI brings together techniques that simulate cognitive functions; all current AI is “weak” (specialised). The Turing test measures the imitation of a conversation, not understanding. The history alternates between enthusiasm and disappointment: recent advances come from data, computing power and Transformers.