History of Artificial Intelligence
The history of artificial intelligence is the roughly eighty-year story of attempts to make machines do things that would require intelligence in people, told as alternating waves of enthusiasm, disappointment and technical breakthrough. Its main thread is a competition between two programs: reasoning with explicit symbols and rules, and learning from data with statistical models and neural networks. Each wave rose on a method that worked surprisingly well on a bounded class of problems, and each ended when the limits of that method became visible.
The prehistory and founding fit into 1943 to 1956. Warren McCulloch and Walter Pitts published a logical calculus of nervous activity in 1943, showing that idealized neurons could compute logical functions. Alan Turing's 1950 paper in Mind asked whether machines can think and proposed the imitation game. In 1955 John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon submitted the proposal for a summer workshop at Dartmouth in 1956, which named the field and gathered its first generation. Newell, Shaw and Simon's Logic Theorist, demonstrated there, proved theorems of Principia Mathematica by heuristic search, an early sign that machines could do more than arithmetic.
The first boom, from 1957 to 1973, was built on search, toy worlds and early learning. Frank Rosenblatt's perceptron raised hopes for trainable networks, McCarthy's Lisp (1958) became the standard AI language, and Arthur Samuel's checkers program learned from play. Joseph Weizenbaum's ELIZA (1966) simulated a psychotherapist with simple pattern matching, and Terry Winograd's SHRDLU (1970 to 1971) understood commands in a blocks world. Confidence ran high, as documented in the entry on early optimism, and funding from ARPA was generous. The reversal came with the 1966 ALPAC report on machine translation, Minsky and Papert's 1969 book Perceptrons, which analyzed the limits of single-layer networks, and the 1973 Lighthill report in Britain, which criticized the field for failing to scale beyond toy problems.
The 1980s were the era of expert systems and of a neural-network revival. Programs encoding expert knowledge as rules, descending from Stanford's DENDRAL and MYCIN, entered business; Digital Equipment Corporation's XCON, built with John McDermott at Carnegie Mellon, configured computer orders and became the best-known commercial success, and Japan's Fifth Generation Computer Systems project, launched in 1982, spurred rival national programs. Meanwhile the 1986 paper by David Rumelhart, Geoffrey Hinton and Ronald Williams in Nature showed that backpropagation could train multilayer networks to learn useful internal representations, and the PDP volumes the same year made connectionism a major school. Expert systems proved expensive to maintain and brittle outside their domains, and a second downturn followed in the late 1980s.
The 1990s and 2000s quietly shifted AI toward statistics and data. Support vector machines, probabilistic graphical models and reinforcement learning matured, and Yann LeCun's convolutional networks were deployed to read handwritten digits on checks. The public landmark was IBM's Deep Blue defeating world champion Garry Kasparov in a match in May 1997, achieved with specialized hardware and massive search rather than human-like understanding. IBM's Watson won Jeopardy! in 2011, and in 2006 Hinton and colleagues introduced training methods for deep belief networks that helped reopen interest in deep networks.
The deep-learning era began with a vision benchmark. In 2012 AlexNet, by Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton, won the ImageNet challenge by a wide margin using a convolutional network trained on GPUs, and results across speech, vision and language followed. In March 2016 DeepMind's AlphaGo defeated Lee Sedol 4 to 1 at Go, combining deep networks, tree search and reinforcement learning, and in 2020 AlphaFold 2 achieved striking accuracy in the CASP14 protein-structure assessment. In 2017 Vaswani and colleagues at Google introduced the transformer in Attention Is All You Need, the architecture behind BERT, the GPT series and most large language models. OpenAI released ChatGPT on 30 November 2022, bringing conversational AI to a mass audience within weeks.
Three lessons recur across the timeline. Scale and data have repeatedly mattered more than predicted: neural networks were known in the 1950s and 1980s but became dominant when computing power and datasets grew. Benchmarks, from chess to ImageNet, have served as engines of progress and as sources of overclaiming. And the narrative of clean cycles hides continuity, since each downturn kept research alive, ideas from earlier waves such as reinforcement learning and probabilistic reasoning returned in later ones, and the symbolic and statistical programs are now frequently combined. A common misconception is to treat ChatGPT as the start of AI, when it sits at the end of at least seven decades of work.
Read the history of artificial intelligence as a long argument between rules and learning, punctuated by bursts of overconfidence and decisive technical advances, in which the same ideas keep returning once the data and hardware finally catch up.