In 1950, British mathematician Alan Turing published “Computing Machinery and Intelligence” in the journal Mind. Rather than debating whether machines could truly “think,” he proposed a practical test: if a human judge, communicating via text, could not reliably distinguish between a machine and a human, the machine should be considered intelligent. This reframed the question from philosophy to engineering.

Turing didn’t build an AI. He did something more important: he defined the goal. By proposing a measurable benchmark for machine intelligence, he gave researchers a target to aim at. The Turing Test remains a reference point in AI discourse 75 years later — even though modern AI has largely moved beyond it.

Turing was not working in a vacuum. During World War II, he had led the effort to crack the Enigma code at Bletchley Park, demonstrating that machines could solve problems previously thought to require human intelligence. The first programmable digital computers were being built in the 1940s, and scientists were beginning to ask whether these machines could do more than arithmetic.

In the summer of 1956, John McCarthy organized a workshop at Dartmouth College with Marvin Minsky, Nathaniel Rochester, and Claude Shannon. McCarthy coined the term “artificial intelligence” for the proposal. The premise was bold: “Every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.”

Attendees predicted that human-level AI would be achieved within a generation. They were wrong by decades — possibly by a century. But the conference established AI as a formal academic discipline, attracted government funding, and launched the careers of researchers who would lead the field for the next 40 years.

The years following Dartmouth were a period of extraordinary optimism. Early successes included programs that could solve algebra problems, prove geometric theorems, and play checkers. DARPA (then ARPA) poured funding into AI labs at MIT, Stanford, and Carnegie Mellon. The prevailing belief was that general intelligence was just a matter of scaling up existing approaches.

British mathematician James Lighthill published a devastating assessment of AI research for the UK Science Research Council. His verdict: “In no part of the field have the discoveries made so far produced the major impact that was then promised.” The core problem was combinatorial explosion — techniques that worked on small toy problems collapsed when applied to real-world complexity.

The UK government cut AI funding at all but two universities. DARPA, influenced by similar disappointments and by Minsky and Papert’s 1969 critique of neural networks in Perceptrons, dramatically reduced AI funding. Researchers scattered. The term “AI winter” was coined to describe this period of collapsed expectations and dried-up investment.

The fundamental issue was that early AI researchers underestimated the difficulty of common-sense reasoning. A computer could prove a theorem but couldn’t understand a children’s story. The gap between narrow problem-solving and general intelligence turned out to be far wider than anyone anticipated. Compute power and data were also orders of magnitude below what was needed.

Expert systems encoded human specialist knowledge as if-then rules. A medical diagnosis system, for example, would contain thousands of rules like: “If the patient has fever AND cough AND chest pain, then consider pneumonia.” These systems didn’t learn from data — they codified what human experts already knew.

R1/XCON at Digital Equipment Corporation configured computer orders and reportedly saved $40 million per year. By 1985, companies were spending over $1 billion annually on expert systems. Japan launched its ambitious Fifth Generation Computer project. The AI industry was back.

Expert systems were brittle. They worked well within their narrow domain but failed unpredictably at the edges. Maintaining the rule base was expensive — every new scenario required manual rule updates. They couldn’t learn, adapt, or handle ambiguity. And they required enormous effort to build: extracting knowledge from human experts and encoding it as rules was slow and error-prone.

The specialized hardware market for AI (Lisp machines) collapsed almost overnight as general-purpose desktop computers from Apple and IBM became powerful enough to run the same software at a fraction of the cost. Expert systems proved too expensive to maintain and too brittle for real-world deployment at scale. Japan’s Fifth Generation project failed to meet its goals.

DARPA’s Strategic Computing Initiative, which had invested hundreds of millions in AI, was wound down. Corporate AI labs were shuttered. The term “artificial intelligence” became so toxic that researchers began rebranding their work as “machine learning,” “knowledge systems,” or “computational intelligence” to avoid the stigma.

Beneath the surface, important foundational work continued. Statistical approaches to machine learning gained traction. Yann LeCun demonstrated convolutional neural networks for handwriting recognition in 1989. Support vector machines and other mathematical frameworks matured. The internet was beginning to generate the massive datasets that would eventually fuel the next wave.

IBM’s Deep Blue defeated world chess champion Garry Kasparov 3½–2½ in a six-game match. It evaluated 200 million positions per second using brute-force computation. This was not “intelligence” in any human sense — it was raw computational power applied to a well-defined problem. But it captured the public imagination and proved machines could outperform humans in complex cognitive tasks.

IBM’s Watson defeated champions Ken Jennings and Brad Rutter on Jeopardy!, demonstrating that AI could process natural language, handle ambiguity, and retrieve knowledge in real time. Unlike Deep Blue, Watson dealt with the messiness of human language — puns, wordplay, and implicit context.

While public milestones grabbed headlines, the real revolution was happening quietly. Google was using machine learning to improve search results. Amazon was building recommendation engines. Netflix launched its $1 million Prize competition in 2006 to improve its recommendation algorithm. Statistical ML was being embedded into products used by hundreds of millions of people — without anyone calling it “AI.”

Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton at the University of Toronto entered a deep neural network called AlexNet into the ImageNet competition. It won with an error rate less than half that of the next best system. This is widely considered the “Big Bang” of the deep learning era — the moment the field realized that neural networks, given enough data and GPU compute, could dramatically outperform all other approaches.

Google DeepMind’s AlphaGo defeated world Go champion Lee Sedol. Go has more possible board positions than atoms in the observable universe — brute force was impossible. AlphaGo used deep reinforcement learning, teaching itself through millions of self-play games. It represented a qualitative leap: machines learning strategy, not just calculation.

Eight Google researchers published “Attention Is All You Need,” introducing the transformer architecture. Unlike previous models that processed text sequentially (word by word), transformers process entire sequences in parallel using a mechanism called self-attention. This made it possible to train far larger models on far more data — and it became the foundation for GPT, BERT, and every major language model that followed.

OpenAI released GPT-3 with 175 billion parameters, demonstrating that scaling transformer models produced emergent capabilities — abilities that appeared without being explicitly trained. GPT-3 could write essays, generate code, translate languages, and answer questions, all from a single model trained on internet text. The “scaling hypothesis” — that bigger models with more data produce qualitatively better results — reshaped the industry’s investment thesis.

OpenAI released ChatGPT, built on GPT-3.5 with reinforcement learning from human feedback (RLHF). It reached 100 million users in two months — the fastest-growing consumer application in history at the time. For the first time, the general public could interact with a capable AI system through natural conversation. This was AI’s “iPhone moment.”

The field has exploded: GPT-4, Claude, Gemini, Llama, Mistral — multiple frontier models from competing organizations. Image generation (DALL-E, Midjourney), video (Sora), and code assistants (GitHub Copilot, Cursor) have entered mainstream use. AI agents that plan and use tools are the current frontier. Gartner projects $2.52 trillion in global AI spending by 2026.