The decade and a half following the Dartmouth workshop was a period of extraordinary optimism and genuine achievement. Researchers built programs that could prove theorems, play games, understand simple English sentences, and manipulate objects in simulated worlds. Funding flowed freely. Predictions were extravagant. For a brief, heady period, it seemed like human-level artificial intelligence was just around the corner.
It was not. But the work done during this golden age laid foundations that would prove essential decades later — and the failures were as instructive as the successes.
The Logic Theorist and GPS
Allen Newell and Herbert Simon's Logic Theorist, demonstrated at the Dartmouth workshop, was just the beginning. In 1957, they unveiled a far more ambitious program: the General Problem Solver (GPS).
GPS was designed to be, as its name implied, a general-purpose problem-solving system. Newell and Simon had studied how humans solved problems — giving subjects puzzles and asking them to think aloud — and tried to replicate those strategies in software. The core technique was called "means-ends analysis": compare where you are with where you want to be, identify the biggest difference, and find an action that reduces that difference. Repeat until you reach your goal.
GPS could solve logic puzzles, prove theorems, and work through simple algebra problems. It was not programmed with specific solutions to specific problems. Instead, it was given a general strategy for problem-solving and let loose.
Simon was so excited by these results that in 1957 he made a series of predictions that have become infamous in AI history. He predicted that within ten years, a computer would be chess champion of the world, that a computer would discover and prove an important new mathematical theorem, and that most theories in psychology would take the form of computer programs.
None of these predictions came true within ten years. The chess prediction was off by forty years, the theorem prediction arguably has still not been fulfilled in the way Simon meant, and psychology did not become a branch of computer science. But Simon's confidence was representative of the era. The early successes were so surprising that extrapolating them to human-level intelligence seemed reasonable.
ELIZA: The First Chatbot
In 1966, Joseph Weizenbaum at MIT created ELIZA, a program that simulated conversation by applying simple pattern-matching rules to the user's input. ELIZA's most famous mode was DOCTOR, which mimicked a Rogerian psychotherapist — a school of therapy that involves reflecting the patient's statements back to them in question form.
If you typed "I am unhappy," ELIZA might respond "Why are you unhappy?" If you wrote "My mother hates me," it might ask "Who else in your family hates you?" The program had no understanding of language, emotions, or psychology. It simply matched patterns in text and applied transformation rules.
What shocked Weizenbaum was not the program's cleverness but people's reaction to it. Users who knew ELIZA was a program would still find themselves confiding in it, developing emotional attachment to it, and insisting on privacy during their "sessions." Weizenbaum's own secretary, who had watched him build the program and understood exactly how it worked, asked him to leave the room so she could talk to ELIZA privately.
Weizenbaum was disturbed by this reaction. He had intended ELIZA as a demonstration of how superficial human-computer interaction was — how easily people could be fooled by a simple trick. Instead, he discovered that people desperately wanted to believe the machine understood them. They projected intelligence and empathy onto a program that had neither.
This phenomenon — humans attributing understanding to machines that merely simulate the appearance of understanding — would recur throughout AI history. It is perhaps more relevant today than ever, as millions of people interact daily with language models that are vastly more sophisticated than ELIZA but still, arguably, engage in a more complex version of the same pattern-matching game.
SHRDLU: Understanding Language
If ELIZA showed how easy it was to fake understanding, Terry Winograd's SHRDLU, created at MIT in 1970, showed how hard genuine understanding actually was — even in the simplest possible world.
SHRDLU could understand and respond to natural English commands about a simulated world of colored blocks on a table. You could type "Pick up a big red block" and the system would identify the correct block, plan a sequence of movements, and execute them. You could ask "What is sitting on the blue pyramid?" and SHRDLU would examine its world and answer correctly. You could even engage in surprisingly natural dialogue:
Human: "Can the table pick up blocks?" SHRDLU: "No." Human: "Can a block be bigger than a pyramid?" SHRDLU: "Yes." Human: "Why did you clear off that block?" SHRDLU: "To put it on a larger green block."
This was genuinely impressive. SHRDLU could parse English sentences, resolve ambiguities, track context across a conversation, plan actions, and explain its reasoning. Within its tiny world, it appeared to understand language.
The catch was "within its tiny world." The blocks world was so simple that SHRDLU could maintain a complete model of every object and relationship. The real world is infinitely more complex. SHRDLU's approach — manually encoding knowledge about objects, properties, and relationships — simply did not scale. Every new concept required new programming. The dream of extending SHRDLU's approach to open-ended language understanding turned out to be a dead end.
This pattern — impressive demonstrations in restricted domains that fail to generalize — would become one of the defining features of AI research for decades.
Shakey the Robot
At the Stanford Research Institute (SRI), a team led by Charles Rosen built Shakey, the first mobile robot capable of reasoning about its own actions. Shakey could navigate rooms, recognize objects, push boxes, and plan multi-step sequences of actions to achieve goals.
Shakey was a remarkable engineering achievement, but it was also painfully slow. It would stop and "think" for extended periods before each action, as its planning algorithms worked through the consequences of possible moves. A task that would take a human a few seconds could take Shakey hours.
Despite its limitations, Shakey pioneered techniques that remain fundamental to robotics and AI. Its planning system, STRIPS (Stanford Research Institute Problem Solver), became one of the most influential formalisms in AI. The representation it used — describing the world as a set of logical propositions and actions as operations that change those propositions — is still widely used in AI planning today.
The Perceptron Controversy
While symbolic AI was enjoying its golden age, an alternative approach was developing in parallel — and heading toward a collision.
Frank Rosenblatt's Perceptron, developed at Cornell in 1958, was a simple neural network that could learn to classify patterns. It took inputs, multiplied them by adjustable weights, summed the results, and produced an output. If the output was wrong, the weights were adjusted to make it more correct next time. Through many rounds of this process, the Perceptron could learn to distinguish simple patterns — horizontal lines from vertical lines, for example.
Rosenblatt was enthusiastic about his invention — perhaps too enthusiastic. He made grand claims about the Perceptron's potential, telling reporters it was "the embryo of an electronic computer that [the Navy] expects will be able to walk, talk, see, write, reproduce itself and be conscious of its existence."
This did not sit well with the symbolic AI establishment, particularly Marvin Minsky, who viewed neural networks as a competing paradigm. In 1969, Minsky and his colleague Seymour Papert published Perceptrons, a mathematical analysis of what single-layer perceptrons could and could not do.
Their key result was devastating: single-layer perceptrons could not learn certain simple functions, including XOR (exclusive or). This was a genuine limitation. But Minsky and Papert's book was widely interpreted as proving that neural networks in general were a dead end — an interpretation that went far beyond their actual mathematical claims.
The impact was enormous. Funding for neural network research dried up almost overnight. Researchers who worked on neural networks found it difficult to get grants or publish papers. The field would not recover for nearly two decades.
Whether Minsky and Papert intended this outcome is debated. They later said they had merely pointed out limitations of a specific architecture, not condemned the entire approach. But the damage was done. The road not taken at Dartmouth was now blocked by a fallen bridge, and AI research would proceed almost exclusively along symbolic lines for the next fifteen years.
The Unfulfilled Promises
By the early 1970s, the gap between AI's promises and its achievements was becoming impossible to ignore.
Simon's prediction that computers would be world chess champions within ten years had not materialized. Natural language understanding remained confined to tiny toy worlds like SHRDLU's blocks. Machine translation — one of the field's earliest goals — had produced embarrassing failures. (A possibly apocryphal story has an early translation system rendering "the spirit is willing but the flesh is weak" into Russian and back as "the vodka is good but the meat is rotten.")
The fundamental problem was becoming clear: intelligence required knowledge — vast amounts of knowledge about the world, about language, about how things relate to each other. The programs of the golden age worked in carefully constrained domains where all necessary knowledge could be hand-coded. But the real world does not come neatly packaged. The challenge was not building clever algorithms. The challenge was representing and managing the enormous body of knowledge that even a child brings to every interaction.
This problem — sometimes called the "knowledge bottleneck" — would haunt AI research for decades. It was not until the rise of machine learning from massive datasets, beginning in the 2000s, that AI found a way around it: instead of programming knowledge in by hand, learn it from data.
But that breakthrough was still far in the future. First, AI had to survive its first winter.