The idea of a machine that can think, learn and decide is far older than the computer. Philosophers, mathematicians and engineers have speculated for centuries about mechanisms that might imitate human reasoning, from Aristotle’s formal logic to the automata and algorithmic devices described by medieval Islamic scholars. The scientific field that now carries the name artificial intelligence, however, has a specific starting point.
Dartmouth, 1956
In 1955, John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon wrote a proposal for a summer research project at Dartmouth College in New Hampshire. The proposal, which coined the term “artificial intelligence,” rested on a conjecture: that every aspect of learning or any other feature of intelligence can in principle be described so precisely that a machine can be made to simulate it. The workshop took place in the summer of 1956 and is widely treated as the formal birth of the field.
The early programs that followed were modest by today’s standards. They played checkers and chess, proved simple theorems and solved puzzles, and they did so by following rules written by their programmers. Computers of the period had limited speed and memory, so progress was slow and often fell short of the optimistic predictions made in the field’s first decade. Periods of reduced funding and interest, later called “AI winters,” followed.
Expert systems and the turn to learning
In the 1980s, expert systems opened a new chapter. These programs encoded the knowledge of human specialists as large sets of if-then rules and could offer advice in narrow domains such as medical diagnosis, chemical analysis and industrial configuration. Their commercial use brought AI back into the mainstream, though they proved expensive to build and brittle when faced with situations their rules did not cover.
The more consequential shift was toward machine learning: systems that improve from data and experience rather than from hand-written rules. Instead of programming every behavior, researchers trained statistical models on examples and let them infer patterns. This approach gradually displaced rule-based methods across most of the field.
Data, computing power and deep learning
After 2000, two trends converged. The internet produced vast quantities of text, images and other data, and computing hardware, particularly graphics processors, became fast enough to train very large models. Neural networks, an idea dating back to the 1940s and 1950s, became practical at scale, and the approach known as deep learning produced sharp improvements in speech recognition, image classification and machine translation during the 2010s.
These systems began handling tasks once thought to be distinctively human: understanding language, recognizing faces and objects, interpreting speech and, more recently, generating text, images and code. Self-driving vehicle prototypes, facial recognition systems and AI-assisted medical imaging all grew out of this lineage.
An idea, not a single invention
None of this came from one breakthrough. The history of AI is better understood as the long working-out of a single premise, the one stated at Dartmouth: that intelligence can be described as a structure and a set of processes that can be reproduced outside a biological brain. Whether that premise is fully correct remains an open scientific question, and the explainer on machine consciousness elsewhere on this site covers the current debate.
What the history does show is that each advance began with a question that earlier generations could not answer with the tools available to them. The field is still in motion, and the questions it raises about human intelligence, and about the limits of imitation, remain unresolved.
