From Turing’s Question to the Age of Machine Learning

“Can a machine think?” Alan Turing opened his 1950 paper, Computing Machinery and Intelligence, with that single sentence, then spent the rest of it arguing the question itself was probably the wrong one to ask. What he proposed instead — a test based on whether a machine’s conversation could be told apart from a person’s — reframed artificial intelligence as a question about behavior rather than a question about the soul. That reframing is still the fault line the field sits on today, echoed in reference texts like Stuart Russell and Peter Norvig’s Artificial Intelligence: A Modern Approach, which reads less like a history than a working map of everything downstream of Turing’s opening line.

The field didn’t begin with him, though. It inherited centuries of groundwork from people who never heard the word computer. Aristotle’s formal logic gave the West its first rigorous account of valid reasoning, treating argument as something with a structure that could be checked rather than merely felt. Descartes went further, describing the mind itself as a system of operations, a move that made the notion of a thinking mechanism conceivable long before anyone could build one. Leibniz dreamed of a universal symbolic language precise enough to settle any dispute through calculation alone, a dream that looks, in hindsight, like an early sketch of programming. George Boole gave that dream a working grammar in the nineteenth century with Boolean algebra, which later became the literal substrate of digital circuits. And Turing’s own 1936 concept of an abstract computing machine proved something quietly enormous: that a single formal model could, in principle, represent any computational process at all. Everything that followed was, in a sense, implementation detail.

The field got its name and its founding optimism at the Dartmouth Workshop in 1956, organized by John McCarthy and attended by figures including Marvin Minsky and Herbert Simon. Their working assumption was almost audacious in its confidence: that intelligence, properly described, could be simulated by a machine in full. The years that followed produced real, if narrow, proof of concept — programs like the Logic Theorist and the General Problem Solver, which could work through symbolic and logical problems that had previously required a person.

What those early systems couldn’t do turned out to matter more than what they could. They handled clean logic well and ambiguity badly, and they had been oversold to funders who expected human-level reasoning on a timeline the technology had no way of meeting. The resulting contraction in research and funding, documented in reports like the UK’s 1973 Lighthill Report, became known as the AI Winter — less a single event than a prolonged reckoning with a field that had confused enthusiasm for evidence. The recovery, when it came in the 1990s, arrived by abandoning the founding premise rather than vindicating it: intelligence stopped being modeled through hand-built symbolic rules and started being modeled statistically, learned from data rather than programmed as logic. Turing’s original question was never really answered so much as sidestepped by an entire discipline that found it more productive to ask what a machine could learn to do than what it truly was.

The clearest illustration of that sidestep is IBM’s Deep Blue defeating world chess champion Garry Kasparov in 1997, a landmark that gets celebrated as a triumph of intelligence and is really closer to its opposite. Deep Blue won by brute-force search through millions of positions per second, guided by hand-tuned evaluation functions, not by anything resembling human insight or strategic understanding. It played chess better than any person alive while understanding nothing about the game the way a person does, which is exactly the gap Turing’s test was built to route around rather than resolve.

A fair complication is that machine learning’s own success story complicates the field’s founding optimism rather than vindicating it outright. Statistical, data-driven systems solved problems symbolic AI never cracked — image recognition, language translation, protein folding — but they did it by trading interpretability for performance, producing systems that work reliably without anyone, including their own engineers, being able to fully explain why a given output emerged from a given input. The field got machines that behave intelligently and lost, in the same move, much of its ability to say what intelligence inside them actually consists of.

Turing, writing in 1950, seemed to anticipate exactly this kind of settlement without being able to name it. He didn’t ask his famous question because he thought consciousness was unknowable in principle; he set it aside because he suspected the field would make faster progress arguing about behavior than about essence. Three-quarters of a century on, that bet looks close to correct. The discipline he helped found still has no agreed answer to whether a machine can think. It has simply gotten extraordinarily good at building machines nobody can any longer confidently say aren’t.