Ideas before computers
People imagined thinking machines long before electronic computers existed. Myths and stories described artificial servants and mechanical beings, and philosophers and mathematicians asked whether reasoning could be reduced to rules. The development of formal logic and the mathematics of computation in the nineteenth and early twentieth centuries gave these questions a concrete form. In the 1930s, Alan Turing and others worked out what it means for something to be computable, laying groundwork for the idea that a machine could follow any well-defined procedure.
The 1950s: a field is born
In 1950, Turing published a paper asking whether machines can think and proposed what became known as the Turing test, a way of judging a machine by whether its conversation could be told apart from a person’s. It is still discussed, though many researchers do not treat it as a practical measure.
The field got its name in 1955-1956. John McCarthy and colleagues proposed a summer research project at Dartmouth College, and the proposal used the phrase “artificial intelligence.” The 1956 Dartmouth workshop is widely treated as the starting point of AI as a research field. Its organizers were optimistic that aspects of learning and intelligence could be described precisely enough for a machine to simulate them.
Early programs from this period could prove simple mathematical theorems and play checkers. Frank Rosenblatt’s perceptron, introduced in the late 1950s, was an early learning system loosely inspired by brain cells and an ancestor of modern neural networks.
The 1960s and 1970s: ambition and limits
Research in this period explored symbolic reasoning, search, and language. ELIZA, a 1960s program by Joseph Weizenbaum, imitated a conversation using simple pattern matching, and some users found it surprisingly engaging despite its lack of any real understanding. Robots and planning systems also appeared.
Progress turned out to be harder than early predictions suggested. Programs that worked in tiny, controlled “toy worlds” often failed in real-world settings, where there are far too many possibilities and too much unstated common sense. Computers of the time were also slow and had little memory. A critical book on the limitations of simple perceptrons, published in 1969, is often credited with cooling interest in that approach.
By the mid-1970s funding and enthusiasm had dropped in several countries, a period commonly called the first “AI winter.”
The 1980s: expert systems and another cycle
Interest revived with expert systems, programs that encoded the knowledge of specialists as many if-then rules for tasks like diagnosing equipment faults or configuring computer orders. Some were commercially useful, and companies invested heavily.
Expert systems proved expensive to build and maintain, and they were brittle outside their narrow domains. Meanwhile, researchers revived and refined methods for training multi-layer neural networks, with the backpropagation algorithm becoming widely known in the mid-1980s. By the late 1980s and into the 1990s, disappointment with expert systems and specialized hardware led to another downturn, sometimes called the second AI winter.
The 1990s and 2000s: quieter progress
Even during lean years, work continued, increasingly built on statistics and probability rather than hand-written rules. Machine learning methods improved at tasks such as speech recognition and spam filtering. A well-known public milestone came in 1997, when IBM’s Deep Blue chess computer defeated world champion Garry Kasparov in a match. Deep Blue relied largely on massive search and specialized chess knowledge rather than learning from data.
The growth of the internet supplied enormous amounts of data, and computing power kept rising, setting conditions for the next shift.
The 2010s: deep learning takes off
Around 2012, a deep neural network trained on graphics processors achieved a large improvement on a widely used image recognition benchmark. Combined with large labeled datasets and faster hardware, deep learning methods began to outperform older approaches in image recognition, speech and translation.
In 2016, DeepMind’s AlphaGo defeated top Go player Lee Sedol, a game long considered too complex for brute-force search. In 2017, researchers introduced the transformer architecture, which handles sequences such as text efficiently and became the foundation of modern language models.
The present
Systems that generate text, images and other media became widely available to the public in the early 2020s. They intensified both excitement and debate about accuracy, jobs, copyright, safety and ethics. The history also offers a reminder: earlier waves of enthusiasm were followed by disillusionment when claims outran results.
Lessons from the history
- Progress is uneven. Long quiet stretches have been followed by sudden advances, usually when new data, computing power and methods arrived together.
- Hype cycles repeat. Overly confident predictions have often been wrong in both directions.
- Approaches compete and merge. Symbolic reasoning, statistics and neural networks have each had their moment, and current research sometimes combines them.
- Hard problems stay hard. Common sense, reliability and understanding context remain open challenges.
The long view suggests treating today’s systems as the latest chapter in a decades-old effort, not as a sudden arrival, and treating confident forecasts about what comes next with caution.