Before the name existed
The idea of a machine that could reason is older than the electronic computer. Mechanical calculators and the notes of Ada Lovelace on Charles Babbage’s proposed engine already raised the question of whether a machine could do more than follow instructions. Lovelace’s own view was that it could only do what it was told to do.
In 1950 the British mathematician Alan Turing published a paper titled “Computing Machinery and Intelligence.” Rather than argue over what thinking means, he proposed an imitation game, now usually called the Turing test, in which a judge tries to tell a machine from a person by conversation alone. The test is still discussed, though many researchers regard it as a thought experiment rather than a practical goal.
Around the same time, early electronic computers were becoming available to universities. Researchers started writing programs that played checkers and chess, and the question of machine intelligence moved from philosophy toward experiment.
The Dartmouth workshop
The term “artificial intelligence” is generally credited to John McCarthy, who used it in a 1955 proposal for a summer research project at Dartmouth College. The proposal was co-signed by Marvin Minsky, Nathaniel Rochester and Claude Shannon, and the workshop took place in the summer of 1956. It was a small gathering rather than a single breakthrough event, and attendees did not agree on method. What it did was give the field a name and a community.
Programs from this period are often cited as the first working examples. The Logic Theorist, written by Allen Newell, Herbert Simon and Cliff Shaw, proved theorems from a standard logic textbook and was shown around the time of the workshop. The same group later built the General Problem Solver, which tried to separate a general search method from the details of any one puzzle.
Early optimism and its limits
The late 1950s and 1960s produced a steady stream of demonstrations:
- Programs that played checkers well enough to beat their authors.
- Systems that solved algebra word problems or simple geometry exercises.
- ELIZA, written by Joseph Weizenbaum in the mid-1960s, which imitated a conversation by matching patterns in typed text. Weizenbaum was reportedly uneasy that people read more understanding into it than it had.
- The perceptron, a simple learning machine developed by Frank Rosenblatt, which was an early form of the neural network.
Predictions from this period were often very confident, and they were not met on the timelines given. Problems that looked like small extensions of the demonstrations turned out to be much harder. A program that handled a toy world could fail badly when faced with the messiness of real language, vision or common sense.
The first AI winter
Funding and enthusiasm cooled in the 1970s. A 1969 book by Minsky and Seymour Papert analyzed limits of single-layer perceptrons, and many readers took it as a reason to drop neural network work for a time, although historians still debate how large its effect was. In the United Kingdom, a 1973 report by the mathematician James Lighthill was critical of the field’s progress and contributed to reduced support. Research continued, but with fewer resources and more modest claims. This period is commonly called the first AI winter.
Expert systems and a second winter
In the 1980s attention shifted to expert systems, which encoded the rules of a human specialist, such as a diagnostic procedure, as a large set of if-then statements. Some were used commercially. A configuration system at Digital Equipment Corporation, often called R1 or XCON, is a frequently cited example. These systems were useful in narrow domains but were costly to maintain, and they handled situations outside their rules poorly.
Specialized hardware built for AI languages lost out to ordinary computers, and by the end of the 1980s and into the early 1990s funding fell again. This second slump is usually called the second AI winter.
Learning from data
Meanwhile, work on neural networks continued. A 1986 paper by David Rumelhart, Geoffrey Hinton and Ronald Williams helped popularize backpropagation, a method for training networks with several layers, although related ideas had appeared earlier. Statistical methods, which estimate patterns from examples instead of hand-written rules, became the dominant approach in speech recognition, translation and many other areas through the 1990s and 2000s.
A widely reported milestone came in 1997, when IBM’s Deep Blue defeated the world chess champion Garry Kasparov in a match. Deep Blue relied mostly on fast search and hand-tuned evaluation, not learning, which is a reminder that the field has always combined several approaches.
Three things then came together. Large collections of digital examples became available, partly because of the growth of big data. Graphics cards, originally built for games, turned out to be well suited to the arithmetic used in neural networks. And researchers refined deep, many-layered networks. A 2012 image-recognition result from a deep network trained on the ImageNet dataset is commonly pointed to as the moment this approach drew wide attention. In 2017, a design called the transformer, introduced by researchers at Google, became the basis for many later language models.
AI in everyday technology
Machine learning is now built into many products without being labeled as AI. Voice assistants convert speech to text, smartphones sort photos by their content, and research on self-driving cars relies heavily on trained perception systems. In robotics, industrial robots have long followed fixed programs, while newer work tries to let machines adapt to varied tasks. The tools are also used in cybersecurity to flag unusual activity.
What remains unsettled
Today’s systems can be impressive at narrow tasks, yet they make errors that people would not, and they depend on the data they were trained on. Researchers disagree about how far current methods will go, whether they amount to understanding in any meaningful sense, and how best to manage the risks. The history above suggests caution about confident forecasts in either direction: the field has repeatedly moved faster than skeptics expected in some areas and slower than enthusiasts expected in others.