Old ideas about artificial minds
People imagined made-to-order thinking things long before computers. Greek myths described mechanical servants, medieval legends told of speaking statues, and writers in the nineteenth and twentieth centuries imagined artificial people and robots. These stories are not science, but they show that the idea of a constructed mind is much older than the technology.
A more technical thread began with logic and calculation. Philosophers and mathematicians asked whether reasoning could be reduced to rules. Charles Babbage designed a programmable engine in the 1830s, and his collaborator Ada Lovelace remarked that it could not originate anything by itself, a point that is still quoted in discussions of machine creativity. Formal logic developed through the nineteenth century, and in the 1930s mathematicians defined precisely what it means to compute something, which laid the groundwork for the question of machine intelligence.
The 1940s and early 1950s: groundwork
Several strands came together after the Second World War. Early electronic computers existed, and some of their builders openly speculated about whether the machines might one day think. In 1943, Warren McCulloch and Walter Pitts published a mathematical model of simplified neurons, an early link between brain-inspired ideas and logic. Norbert Wiener’s work on feedback and control, which he called cybernetics, encouraged people to think of living and mechanical systems in similar terms.
In 1950, Alan Turing published a paper asking whether machines can think. He proposed replacing the vague question with an imitation game, now usually called the Turing test, in which a judge tries to tell a machine’s typed answers from a human’s. The test is often misunderstood and has never been a formal standard in the field, but the paper helped make machine intelligence a respectable subject.
1956 and the first wave
The term “artificial intelligence” is generally traced to a proposal for a summer workshop at Dartmouth College in 1956, organized by John McCarthy together with Marvin Minsky, Nathaniel Rochester and Claude Shannon. The proposal expressed the hope that aspects of learning or intelligence could be described so precisely that a machine could simulate them. Attendees and others working at the time included Allen Newell and Herbert Simon, whose Logic Theorist program proved some theorems from a mathematics text and is often counted among the first AI programs.
The period that followed produced programs that played checkers, solved algebra word problems, manipulated symbols and held simple text conversations. The Lisp language, created in the late 1950s, became a favorite tool for this kind of symbolic programming. Researchers made confident predictions about how soon machines would match human abilities. Those predictions turned out to be far too optimistic.
The first disappointment
By the late 1960s and 1970s, it was becoming clear that methods that worked on small, tidy examples often did not scale to messy real-world problems. Computers were slow and had little memory by later standards, and programs that depended on hand-written rules struggled when situations were ambiguous. A critical book on early perceptron models by Minsky and Papert in 1969 is often credited with reducing interest in one learning approach, though historians debate how large its effect was. Reports commissioned by governments questioned the progress made, and funding in several countries was cut back.
This downturn is commonly called the first “AI winter.” The label is informal, and the severity varied by country and research area, but the general pattern of inflated expectations followed by cutbacks has repeated in the field.
Expert systems and a second boom
In the 1980s, interest returned through expert systems, programs built from large collections of rules written down by specialists in some narrow area such as diagnosing particular problems or configuring equipment. Several companies sold such systems, and governments launched large research programs. Some applications were genuinely useful. But the systems were costly to maintain, brittle outside their narrow areas, and they did not learn from experience. By the late 1980s and early 1990s, enthusiasm and funding again declined, a period often described as a second winter.
Statistics, data and learning
Even during the lean years, work continued. Researchers increasingly emphasized probability, statistics and learning from examples instead of hand-written rules. Interest in neural networks revived in the 1980s when methods for training multi-layer versions were popularized. Progress was gradual, and the results were often modest until more data and more computing power became available.
Some milestones are well known. In 1997, a chess machine built by IBM defeated the reigning world champion in a match, a landmark for game-playing, although the techniques were specialized and did not amount to general intelligence. In the 2000s, large collections of data and faster hardware made learning-based methods more practical. In the 2010s, so-called deep learning systems showed striking improvements in image recognition, speech recognition and translation, which led to the present wave of attention.
Where things stand
Today’s systems can do things that would have surprised many earlier researchers, such as producing fluent text and recognizing objects in photographs. They also have well-documented limits, including making errors, reflecting biases in their training data, and being hard to interpret. Opinions differ about how close current methods are to human-like general intelligence, and honest answers are uncertain.
The history offers a few general lessons. Progress has come in bursts, often when new ideas coincide with new hardware or data. Predictions about timing have been unreliable in both directions. And the definition of “intelligence” keeps moving: tasks once considered signs of intelligence, such as playing chess, are often reclassified as ordinary computing once machines can do them. That makes the story of artificial intelligence research as much a story about changing expectations as about technology.