Can Machines Think?
How Turing made intelligence visible for the first time, and why visibility didn’t equal reproducibility.
Chapter 1 of A Brief History of Artificial Intelligence.
“The original question, ‘Can machines think?’ I believe to be too meaningless to deserve discussion.” — Alan Turing, “Computing Machinery and Intelligence” (1950)
In the autumn of 1950, five years after the first programmable computers came online and six years before anyone would coin the term “Artificial Intelligence,” Alan Turing had a problem. Machines existed, room-sized calculating engines with names like EDSAC and LEO, built from thousands of vacuum tubes and miles of wire. They could perform arithmetic at speeds no human could match. They could execute instructions precisely, tirelessly, without error.
But they couldn’t learn.
Turing was 38 years old, working at the University of Manchester, already famous among the handful of people who knew what he’d done during the war, which was almost no one. His wartime work breaking German codes at Bletchley Park remained classified, a secret that would stay buried for decades. To most of the world, he was simply a talented mathematician with unusual ideas about machines.
And he had a question on his mind: Can machines think?
He didn’t especially like the question. It struck him as the sort of thing philosophers might argue about forever without reaching any useful conclusion. The word “think” was slippery. The word “machine” had connotations. Put them together and you had a recipe for endless semantic quibbling. Does thinking require consciousness? Must it involve language? Can something made of metal and wires truly understand, or only simulate understanding?
These questions, Turing realized, had no good answers. Not because they were profound, but because they were poorly formed. Asking whether machines can think was like asking whether submarines can swim. The question presupposes a definition of swimming, or thinking, that we don’t actually have.
So Turing did something clever. He replaced the unanswerable question with a testable one.
The Imitation Game
Instead of asking “Can machines think?”, Turing proposed a game. Later it would be called the Turing Test, though he called it the Imitation Game.
The setup is simple. Imagine a human judge conducting text-based conversations with two hidden participants: one human, one machine. The judge asks questions, receives answers, and tries to determine which respondent is which. If the judge cannot reliably tell the difference, then for all practical purposes, the machine has demonstrated intelligence.
Not consciousness. Not understanding. Not whatever ineffable quality makes human thought human. Just the capacity to produce responses indistinguishable from those of a thinking being.
Turing put it plainly in his paper: “The original question, ‘Can machines think?’ I believe to be too meaningless to deserve discussion.” Instead, he asked: Can a machine play the imitation game successfully?
This wasn’t an evasion. It was a profound reframing.
What Turing understood, and this is the insight that makes his paper remarkable seventy-five years later, is that intelligence becomes visible only in its effects. You cannot see thinking. You cannot measure it directly. You can only observe what it produces: patterns of behavior that appear to come from something that understands.
Before Turing, intelligence was invisible. After Turing, it had a shape.
Intelligence Made Empirical
This might seem like dodging the question. But consider what happens when you try to answer it any other way. How do you know that I’m intelligent? You’ve never been inside my mind. You’ve never experienced my thoughts from the inside. All you have is my behavior: the words I write, the problems I solve, the way I respond to unexpected situations.
You infer my intelligence from these patterns. You assume that consistent, appropriate, contextual responses indicate an underlying capacity to think. But you cannot verify it directly. You take it on faith, based on evidence.
Turing was saying: apply the same standard to machines. If a machine can participate in a conversation indistinguishably from a human, what grounds would you have for denying it intelligence? You might say “but it’s just following rules” or “it doesn’t really understand.” But how do you know humans aren’t also following rules, just more complicated ones? How do you know I understand, rather than merely generating responses based on patterns I’ve internalized?
You can’t know. Not with certainty. You can only observe behavior and make inferences.
This is why Turing’s reframing matters. He made intelligence empirical. He gave us something we could test, measure, and build toward. Intelligence became visible for the first time. Not as some mysterious inner property, but as a recognizable pattern of behavior.
But visibility and reproducibility are different things. Turing showed us what intelligence looked like. He didn’t show us how to create it. That distinction would haunt AI research for the next fifty years.
A Paper Ahead of Its Time
The paper appeared in the philosophy journal Mind in October 1950. Most people didn’t read it. Those who did often found it strange, provocative, or both. Turing anticipated every obvious objection, from theology (”thinking is a function of man’s immortal soul”), from consciousness (”machines can never experience the world”), from mathematical logic (”machines are bound by formal systems”), and dispatched them with a mixture of rigorous argument and dry humor.
He even made a prediction: by the year 2000, machines would be good enough at the Imitation Game to fool a judge 30% of the time after five minutes of conversation.
He was wrong about the timeline. In 2000, chatbots were still transparently mechanical. But he was right about almost everything else.
What Turing had done was invent the first operational definition of artificial intelligence. Not a theory about how minds work. Not a blueprint for building thinking machines. Just a test: If it acts intelligent, treat it as intelligent.
And that operational definition would shape the next seventy years of AI research, though not in the way Turing might have expected. Because seeing intelligence and creating it turned out to require very different skills.
The Age Before Learning
Early AI researchers took Turing’s test seriously, but they misunderstood what it implied about how to proceed.
They thought: if intelligence is behavior that looks intelligent, we just need to program the right behaviors. Want a machine to seem smart? Teach it to play chess, everyone knows chess requires intelligence. Have it prove mathematical theorems. Make it diagnose diseases from symptoms. Build systems that could do the things intelligent humans do.
This was the age before learning. The age of programming everything by hand.
And it worked, sort of.
In 1956, Allen Newell and Herbert Simon created the Logic Theorist, a program that could prove mathematical theorems from Principia Mathematica. It was genuinely impressive: the program discovered a more elegant proof of one theorem than the original. Newell and Simon were so confident they’d cracked intelligence that they announced: “We have invented a computer program capable of thinking non-numerically, and thereby solved the venerable mind-body problem.”
They were wrong. But in 1956, it was hard to see why.
Chess programs got steadily better throughout the 1960s and 70s. By 1997, IBM’s Deep Blue defeated world champion Garry Kasparov. The machine could evaluate millions of positions per second, see far deeper into the game tree than any human. It played chess at superhuman levels.
Expert systems proliferated in the 1980s. MYCIN diagnosed blood infections as accurately as human specialists. DENDRAL identified molecular structures from mass spectrometry data. XCON configured computer systems for Digital Equipment Corporation, handling thousands of orders and saving millions of dollars.
These systems looked intelligent. They performed tasks that required expertise, pattern recognition, strategic thinking. They had the shape Turing described.
But something was off.
The Hollow Shape
The chess programs couldn’t explain their strategies in natural language. Ask Deep Blue why it made a particular move and you’d get nothing, or rather, you’d get a list of millions of evaluated positions, which isn’t an explanation any human could parse. The program played chess brilliantly but couldn’t talk about chess at all.
The theorem-provers couldn’t handle slight variations of problems they’d solved. Show the Logic Theorist a theorem from a different domain and it would fail completely. It had mastered one tiny corner of mathematical logic but understood nothing about reasoning in general.
The expert systems were even more brittle. MYCIN was exceptional at diagnosing blood infections but useless for anything else. It couldn’t learn from new cases. It couldn’t explain its reasoning beyond reciting the rules it followed. And if you asked it anything outside its narrow domain, say, whether a patient’s symptoms might indicate stress rather than infection, the system would collapse into nonsense.
These weren’t intelligent systems adapting to novel situations. They were elaborate lookup tables, matching patterns to pre-programmed responses. They had the shape of intelligence, they produced intelligent-looking behavior in specific contexts, but the shape was hollow.
The problem, it turned out, was fundamental. Intelligence isn’t a collection of behaviors you can program individually. It’s not knowing the answer to this question or that question. It’s not following decision trees or applying logical rules, no matter how sophisticated.
Intelligence is the capacity to generate appropriate behavior in novel situations. And that’s a very different thing.
The Gap
This distinction runs through the entire history we’re about to explore. It’s the difference between:
Knowing the answer and figuring it out
Following a recipe and understanding cooking
Translating word-by-word and speaking a language
Playing memorized chess lines and seeing the board
Turing’s test was designed to identify the second kind of intelligence: the generative kind, the adaptive kind, the kind that responds appropriately to things it’s never seen before. But for decades, AI researchers pursued the first kind, because it seemed more tractable. You could program chess rules. You could encode medical knowledge. You could build logical systems that manipulated symbols according to formal procedures.
What you couldn’t do, or at least, what nobody knew how to do, was build systems that learned their own rules from experience. Systems that could extract patterns from examples and generalize them to new situations. Systems that improved through practice rather than through programming.
The early researchers were trying to build intelligence by hand-coding it, rule by rule, behavior by behavior. But intelligence, it turns out, is not something you can specify in advance.
It emerges from learning. And in 1950, nobody knew how to teach machines to learn.
The Road Ahead
Today, seventy-five years after Turing wrote his paper, you can open your phone and have a conversation with an AI that would probably pass his test. Not perfectly, you can still trip them up with the right questions. But well enough that for many everyday interactions, you wouldn’t reliably know you weren’t talking to a human.
This happened gradually, then suddenly. For fifty years, chatbots were obviously fake. ELIZA in 1966 could do a passable impression of a Rogerian therapist, but only by reflecting your statements back at you with slight modifications. (”I’m feeling sad.” “Why are you feeling sad?”) Ask anything unexpected and the facade crumbled immediately.
Then, in the 2020s, something shifted. Not because someone solved the theoretical problem of consciousness. Not because we figured out how human brains work, we still barely understand that. But because researchers discovered how to train machines on enough examples that they learned to generate plausible responses to almost anything.
The machines hadn’t become conscious. They hadn’t acquired understanding in any deep philosophical sense. But they had learned to produce the shape of intelligence with startling fidelity.
What changed wasn’t our theory of intelligence. It was our approach to creating it. We stopped programming and started teaching. We stopped specifying behaviors and started showing examples. We moved from the age before learning to the age of learning machines.
Turing made intelligence visible in 1950. He showed us the destination. But the path there led through territory he couldn’t have mapped, through failure and winter and unexpected resurrection, through neural networks dismissed as hopeless that later proved essential, through decades of wandering before finding the light.
This book tells that story. The journey begins here, in 1950, with a question Turing couldn’t fully answer. Not just “Can machines think?” but the deeper question hiding beneath it: “How can machines learn to think?”
Intelligence was visible at last. But no one yet knew how to teach it.
That would require a revolution: one that would take decades to arrive.
Notes & Further Reading
Turing’s foundational paper:
Turing, A.M. (1950). “Computing Machinery and Intelligence.” Mind, Vol. 59, No. 236, pp. 433-460. Available online through various academic databases and open archives. Still remarkably readable today. Turing writes with unusual clarity for a technical philosopher.
On Turing’s life and work:
Hodges, Andrew (1983). Alan Turing: The Enigma. The definitive biography, detailed and deeply researched. Later adapted into the film The Imitation Game, though the film takes considerable liberties with the historical record.
Copeland, B. Jack, ed. (2004). The Essential Turing. A comprehensive collection of Turing’s major papers with helpful editorial commentary.
Early AI systems mentioned:
Newell, Allen and Herbert A. Simon (1956). “The Logic Theory Machine: A Complex Information Processing System.” Their bold claims about solving the mind-body problem appear in their subsequent papers and represent the supreme confidence of early AI: a confidence that would later prove premature.
Weizenbaum, Joseph (1966). “ELIZA. A Computer Program for the Study of Natural Language Communication Between Man and Machine.” Communications of the ACM, 9(1), pp. 36-45. Weizenbaum was disturbed by how readily people anthropomorphized his simple program, a concern that resonates even more strongly with today’s systems.
On the Turing Test in practice:
The Loebner Prize (1991-2019) held annual competitions based loosely on the Turing Test. No program convincingly passed under strict conditions. The competition highlighted both how far AI had to go and how ambiguous Turing’s original criterion was in practice.
The pre-learning era:
For more on why rule-based and symbolic AI ultimately failed to achieve general intelligence, see Chapter 2. For the breakthrough that changed everything, see Chapter 4.