The History of Artificial Intelligence

From Turing's 1950 thought experiment to the ChatGPT phenomenon, artificial intelligence has endured two winters, three hype cycles, and a deep learning revolution. This is the full story of how a summer workshop at Dartmouth became a technology reshaping civilization.

Before the Beginning

The dream of creating thinking machines is ancient. Greek myths describe Hephaestus forging mechanical servants. Medieval legends tell of alchemists building artificial beings. But the scientific study of artificial intelligence, the attempt to build machines that can perform tasks requiring human-level cognition, began in the middle of the twentieth century with a convergence of mathematics, philosophy, and early computing.

This is the story of AI's seven decades: its ambitions, its failures, its winters, and its eventual, extraordinary vindication.

The Foundations: 1940s and 1950s

The Mathematical Groundwork

Before anyone could build intelligent machines, they needed a theory of computation itself. Alan Turing provided the foundation in 1936 with his concept of the Turing machine, a mathematical model that formalized what it means to compute something. If a problem could be solved by following a finite set of rules, a Turing machine could solve it. This established the theoretical framework within which all future AI research would operate.

In 1943, Warren McCulloch and Walter Pitts published "A Logical Calculus of the Ideas Immanent in Nervous Activity," proposing a mathematical model of biological neurons. Their artificial neuron was a simple binary switch: it took weighted inputs, summed them, and fired (output 1) if the sum exceeded a threshold. This was a vast simplification of real neurons, but it showed that logical operations could be performed by networks of simple processing units. The paper planted the seed that would eventually grow into modern neural networks.

The Turing Test

In 1950, Turing published "Computing Machinery and Intelligence," which opened with the question: "Can machines think?" Rather than trying to define "thinking," Turing proposed an operational test. A human interrogator converses with two hidden interlocutors, one human and one machine, through text-only communication. If the interrogator cannot reliably distinguish the machine from the human, the machine has passed the test.

The Turing Test was not meant as a rigorous definition of intelligence. It was a pragmatic reframing: instead of debating what "thinking" means philosophically, Turing suggested focusing on what the machine can do. The test remains influential and controversial to this day. Modern LLMs can pass casual versions of it, but whether that constitutes genuine intelligence is still vigorously debated.

The Dartmouth Conference: 1956

The field of artificial intelligence was formally born in the summer of 1956, when John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon organized a two-month workshop at Dartmouth College. Their proposal stated the bold hypothesis: "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."

The workshop brought together the researchers who would define the field for decades. The attendees included Allen Newell and Herbert Simon, who had already created the Logic Theorist, a program that could prove mathematical theorems from Whitehead and Russell's Principia Mathematica. It was arguably the first AI program.

McCarthy coined the term "artificial intelligence" at this conference, choosing it partly to distinguish the field from the existing discipline of cybernetics. The name stuck, along with the optimistic assumption that human-level AI was perhaps twenty years away. That prediction would prove spectacularly wrong.

The Golden Age: Late 1950s to Early 1970s

The years following Dartmouth were marked by exuberant optimism. Early AI programs achieved results that seemed to confirm the founders' boldest claims.

Early Triumphs

In 1958, Frank Rosenblatt built the Perceptron, a physical machine (not just a program) that could learn to classify patterns. Rosenblatt's perceptron was a single-layer neural network that could be trained to recognize simple visual patterns. The New York Times reported that the Navy had built a machine that could "think." Rosenblatt himself predicted that perceptrons would eventually "be able to walk, talk, see, write, reproduce itself, and be conscious of its existence."

In 1964, Daniel Bobrow created STUDENT, a program that could solve algebra word problems stated in English. Joseph Weizenbaum created ELIZA in 1966, a chatbot that simulated a Rogerian psychotherapist by using pattern matching to reflect users' statements back to them. ELIZA was remarkably effective at creating the illusion of understanding, so effective that Weizenbaum's own secretary asked him to leave the room so she could have a private conversation with the program. Weizenbaum was disturbed, not encouraged, by this response.

Newell and Simon's General Problem Solver (GPS) attempted to create a universal problem-solving engine based on means-ends analysis. Their Physical Symbol System Hypothesis (1976) argued that a physical symbol system (like a computer program) has the necessary and sufficient means for intelligent action. This became the philosophical foundation of symbolic AI: intelligence is symbol manipulation.

The Limits of Optimism

In 1958, Simon predicted that within ten years, a computer would be chess champion of the world and would discover and prove an important new mathematical theorem. In 1965, Simon claimed that "machines will be capable, within twenty years, of doing any work a man can do." Minsky told Life magazine in 1970 that "in from three to eight years, we will have a machine with the general intelligence of an average human being."

These predictions reflected genuine excitement about early results, but they dramatically underestimated the difficulty of the problems ahead. Simple, well-defined tasks (theorem proving, chess in constrained settings, pattern matching) were achievable. Common sense reasoning, natural language understanding, and perception in the real world proved immeasurably harder.

The First AI Winter: 1974 to 1980

Perceptrons and the Neural Network Collapse

In 1969, Minsky and Seymour Papert published "Perceptrons," a mathematical analysis of the perceptron's limitations. They proved that a single-layer perceptron could not learn the XOR function (a simple logical operation). The book was technically correct but widely interpreted as a death sentence for neural network research. Funding for neural network research dried up almost overnight.

The irony is that multi-layer networks could learn XOR and much more. Minsky and Papert acknowledged this in their book but expressed skepticism that practical training algorithms for multi-layer networks would be found. This skepticism proved wrong, but it took nearly two decades for the correction to arrive.

The Lighthill Report

In 1973, the British mathematician Sir James Lighthill published a devastating report to the British Science Research Council evaluating the state of AI research. Lighthill concluded that AI had failed to achieve its ambitious goals, that progress had been grossly overstated, and that the field suffered from a "combinatorial explosion" problem: techniques that worked on toy problems fell apart when applied to real-world complexity.

The Lighthill Report led to the near-total elimination of AI funding in the United Kingdom. Combined with funding cuts in the United States (DARPA significantly reduced AI spending after the Vietnam War shifted defense priorities), the field entered its first winter. University AI labs shrank, researchers moved to other fields, and the word "artificial intelligence" became something of an embarrassment.

The Expert Systems Boom and Bust: 1980 to 1993

The Rise of Expert Systems

The second wave of AI enthusiasm was driven by expert systems: programs that encoded human expert knowledge as collections of if-then rules. MYCIN (1976) could diagnose bacterial infections with accuracy rivaling human specialists. R1/XCON (1980) configured computer systems for Digital Equipment Corporation, saving the company an estimated $40 million per year.

The commercial potential was obvious. Companies like Symbolics, Lisp Machines Inc., and IntelliCorp built specialized hardware and software for AI applications. Japan launched its ambitious Fifth Generation Computer Systems project in 1982, aiming to build massively parallel computers for AI. The United States and Britain responded with their own funding initiatives, worried about falling behind.

By 1985, the AI industry was worth over $1 billion annually. Corporate AI labs multiplied. Expert systems were deployed in manufacturing, finance, medicine, and oil exploration.

The Second Winter: 1987 to 1993

The collapse came swiftly. Expert systems proved brittle: they worked well within their narrow domains but failed catastrophically at the edges. They could not learn from experience, could not handle uncertainty gracefully, and required enormous effort to build and maintain. Every piece of knowledge had to be manually encoded as a rule, and real-world expertise turned out to be far harder to formalize than anyone expected.

The specialized Lisp machines that powered many AI applications were made obsolete by cheaper, faster general-purpose workstations from Sun Microsystems and others. The AI hardware companies collapsed. Japan's Fifth Generation project quietly wound down without achieving its goals.

By 1993, the term "artificial intelligence" had become toxic in funding proposals and corporate strategy. Researchers rebranded their work as "machine learning," "data mining," "pattern recognition," or "knowledge systems" to avoid the stigma. AI entered its second, and deeper, winter.

The Quiet Revolution: 1993 to 2011

Machine Learning Emerges

While the AI label fell out of favor, the underlying research continued and quietly produced genuine breakthroughs. The key shift was from hand-coded rules to learning from data.

In 1986, David Rumelhart, Geoffrey Hinton, and Ronald Williams published their landmark paper on backpropagation, providing an efficient algorithm for training multi-layer neural networks. This was the answer to Minsky and Papert's skepticism from 1969: multi-layer networks could be trained, and they could learn functions that single-layer perceptrons could not.

But neural networks remained a niche interest through the 1990s. They required more data and compute than was readily available, and other techniques, particularly Support Vector Machines (SVMs) and ensemble methods like Random Forests, delivered strong results with less computational overhead and better theoretical foundations.

Statistical approaches dominated natural language processing. Instead of trying to teach computers the rules of grammar, researchers trained statistical models on large text corpora. IBM's statistical machine translation systems, speech recognition based on Hidden Markov Models, and probabilistic information retrieval all achieved practical success without claiming to be "AI."

Milestones in the Wilderness

Even during the quiet years, landmark achievements accumulated. In 1997, IBM's Deep Blue defeated world chess champion Garry Kasparov. Though Deep Blue relied heavily on brute-force search rather than learning, it was a cultural milestone that demonstrated machines could exceed human performance in complex intellectual tasks.

In 2005, a Stanford autonomous vehicle completed the DARPA Grand Challenge, driving 132 miles through the Mojave Desert without human intervention. Self-driving cars went from science fiction to engineering problem.

In 2011, IBM's Watson defeated the two greatest Jeopardy champions, Ken Jennings and Brad Rutter, on national television. Watson combined natural language processing, information retrieval, and machine learning in a system that could understand and respond to natural language questions across arbitrary domains.

The Deep Learning Breakthrough: 2012 to 2019

The ImageNet Moment

Everything changed in September 2012. Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton submitted a deep convolutional neural network called AlexNet to the ImageNet Large Scale Visual Recognition Challenge. AlexNet achieved a top-5 error rate of 15.3%, compared to 26.2% for the second-place entry. It did not just win; it demolished the competition by a margin that made the entire computer vision community reconsider its assumptions.

AlexNet was not architecturally revolutionary. Convolutional neural networks had existed since Yann LeCun's LeNet in the late 1980s. What changed was the combination of three factors: much larger training datasets (ImageNet had 1.2 million labeled images), GPU computing (Krizhevsky trained the model on two NVIDIA GTX 580 GPUs), and a few key techniques like ReLU activations and dropout regularization.

The deep learning era had begun. Within two years, nearly every computer vision benchmark was dominated by deep neural networks. Researchers who had spent decades on hand-engineered features switched to learning features from data.

The Expansion

Deep learning expanded rapidly beyond image recognition. Recurrent neural networks and LSTMs advanced machine translation, speech recognition, and text generation. Generative Adversarial Networks (GANs), introduced by Ian Goodfellow in 2014, could generate realistic images by pitting two neural networks against each other. DeepMind's AlphaGo defeated world Go champion Lee Sedol in 2016, a feat that many experts had predicted was decades away, since Go's enormous branching factor made brute-force search infeasible.

Google, Facebook, Microsoft, Amazon, and other technology companies invested billions in AI research, hiring top academics and building massive GPU clusters. AI moved from academic curiosity to strategic corporate priority.

The Transformer

In 2017, researchers at Google published "Attention Is All You Need," introducing the transformer architecture. By replacing recurrence with self-attention, transformers could process sequences in parallel, making them dramatically faster to train and better at capturing long-range dependencies. The transformer became the foundation for the modern era of language models.

BERT (2018) showed that transformers pre-trained on large text corpora could be fine-tuned for virtually any NLP task with state-of-the-art results. GPT-2 (2019) demonstrated that simply scaling up a transformer trained on next-token prediction produced remarkably fluent and coherent text generation.

The GPT Era: 2020 to Present

GPT-3 and the Scaling Hypothesis

In June 2020, OpenAI released GPT-3, a transformer language model with 175 billion parameters trained on a large fraction of the internet. GPT-3 demonstrated something remarkable: with enough scale, a model trained purely on next-token prediction could perform translation, arithmetic, code generation, creative writing, and many other tasks it was never explicitly trained for. You could prompt it with a few examples of a task, and it would generalize, a capability called "in-context learning" or "few-shot learning."

GPT-3 was not the most accurate model for any specific task, but its generality was unprecedented. A single model could attempt virtually any language task with reasonable competence. This validated the scaling hypothesis: that simply making models larger, training them on more data, and using more compute would continue to produce qualitative improvements in capability.

ChatGPT and the Public Awakening

In November 2022, OpenAI released ChatGPT, a version of GPT-3.5 fine-tuned for conversation using RLHF. It reached 100 million users within two months, making it the fastest-growing consumer application in history. For the first time, the general public interacted directly with a capable AI system and experienced its strengths and limitations firsthand.

ChatGPT sparked a global conversation about AI's implications for education, employment, creativity, and truth. It also ignited a commercial arms race. Google rushed Bard (later Gemini) to market. Meta released the open-source Llama models. Anthropic launched Claude. Microsoft invested $10 billion in OpenAI and integrated AI into its products. Startups proliferated across every sector.

The Current Landscape

As of the mid-2020s, AI development is characterized by several simultaneous trends. Foundation models continue to grow in capability. Multimodal models process text, images, audio, and video through unified architectures. Reasoning capabilities are improving through techniques like chain-of-thought prompting and reinforcement learning on reasoning tasks. AI agents that can use tools, browse the web, and execute multi-step plans are becoming practical.

The economic impact is substantial and accelerating. AI is transforming software development (with code-generation tools), customer service (with chatbots grounded in company knowledge), scientific research (with protein structure prediction and drug discovery), and creative work (with image, video, and music generation).

At the same time, serious concerns about safety, alignment, job displacement, misinformation, and concentration of power are driving policy debates worldwide. The European Union's AI Act, the first comprehensive AI regulation, took effect in stages beginning in 2024. The United States, China, and other nations are developing their own regulatory frameworks.

What the History Teaches

Seven decades of AI history reveal several recurring patterns.

Hype cycles are real. The field has repeatedly promised more than it could deliver in the short term, leading to backlash and funding cuts. But each cycle left behind genuine advances that contributed to the next wave.

Compute and data matter as much as algorithms. Many of the ideas behind modern AI (neural networks, backpropagation, attention) were conceived decades before they became practical. What changed was the availability of sufficient computing power and training data.

Simple objectives at scale produce surprising capabilities. GPT-3's few-shot learning was not designed into the architecture. It emerged from the interaction of a simple training objective (predict the next token) with massive scale. Many of AI's most impressive capabilities have been emergent rather than engineered.

The field's center of gravity shifts. From symbolic AI to expert systems to statistical learning to deep learning to large language models, each era's dominant paradigm was different. Researchers who adapted thrived; those who clung to previous paradigms were left behind.

AI's impact is real but different from predictions. Early AI pioneers imagined general-purpose reasoning machines. What materialized instead were specialized systems of extraordinary capability: models that can translate languages, generate images, fold proteins, and carry on conversations, but that lack the common sense and general reasoning that a five-year-old possesses.

The history of artificial intelligence is a story of ambition, failure, persistence, and eventual, partial vindication. The field's founders were wrong about how quickly their dreams would be realized. But they were right that machines could be made to perform tasks that seem to require intelligence. The machines just turned out to be very different from what anyone expected.