Turing's Question, Seventy-Five Years On
What a 1950 paper about imitation got right about meaning, machines, and the limits of knowing
In October 1950, Alan Turing published a paper in Mind — a philosophy journal, not a computing one — titled "Computing Machinery and Intelligence." It opened with a sentence that discarded its own question:
"I propose to consider the question, 'Can machines think?'" — and then, immediately: "The original question, 'Can machines think?' I believe to be too meaningless to deserve discussion."
— Turing, A.M. (1950). "Computing Machinery and Intelligence." Mind, 59(236), 433–460.
What he offered instead was a test — the Imitation Game — and a set of objections, anticipated and answered. The paper is twenty-eight pages. It is now seventy-five years old. Almost nothing in it has dated.
The replacement question
Turing's move was philosophical, not technical. He refused to define "think" — because, he argued, the question "can machines think?" collapses into a dispute about definitions rather than about reality. Instead he proposed a behavioural criterion: if a machine can sustain a conversation indistinguishable from a human's, then whatever it is doing is, for all practical purposes, thinking. The question of whether it really thinks — in some inner, private, metaphysical sense — he set aside as undecidable and unproductive.
This is exactly how large language models are now evaluated by millions of users daily. Not: "is it conscious?" but: "did that response help me? Did it understand what I meant? Could I tell it apart from a knowledgeable person?" Turing's framing bypassed seventy-five years of philosophical dispute and landed, with precision, on the criterion people actually use.
The objections, pre-empted
The bulk of the paper is a catalogue of nine objections to machine intelligence — each stated fairly, then answered. Reading them today is uncanny, because they are the same objections raised in every contemporary debate about AI. Turing heard them coming.
Lady Lovelace's Objection: "A machine can only do what we tell it to. It cannot originate anything."
Turing's response: we cannot fully predict the behaviour of a sufficiently complex system, even one we built. A machine that learns from data will produce outputs its programmers did not foresee and could not have specified. The gap between "we wrote the code" and "we know what it will say" grows with complexity until it becomes, in practice, indistinguishable from origination.
Modern neural networks are the embodiment of this response. No engineer at OpenAI or Google can predict what a large model will say to a novel prompt. The weights were trained, not authored. The system was made, but its outputs are not dictated. Lovelace's objection — the strongest of the nine, and the one most often restated — has been empirically overcome.
The Mathematical Objection (from Gödel): Gödel's incompleteness theorem shows that any formal system has statements it cannot prove. Therefore, the argument runs, there are truths a machine cannot reach.
Turing's answer is characteristically precise: the same limitation applies to humans. We, too, are formal systems of a kind — or if we are not, we have no evidence of this. The Gödelian limit is not a difference between humans and machines; it is a feature of all finite systems. To wield it against machines alone is to claim a privilege for human cognition that has not been demonstrated.
The Argument from Consciousness: "A machine does not feel. It may behave as if it understands, but there is no experience behind the behaviour."
Turing's response is the most philosophically radical in the paper. He points out that the only consciousness you can be certain of is your own. You infer consciousness in others from their behaviour — from what they say and do. You have never had direct access to another person's experience. If you accept behavioural evidence as sufficient for inferring consciousness in humans, you cannot refuse it for machines without invoking a double standard. The alternative — solipsism — is available, but few wish to occupy it.
This argument has not been refuted. It has only been repeated, with increasing discomfort, as machines have become more convincingly conversational.
The learning machine
Section 7 of the paper — often overlooked — is where Turing lays out what he actually thinks will work. Not programming intelligence directly. Not encoding rules. Instead:
"Instead of trying to produce a programme to simulate the adult mind, why not rather try to produce one which simulates the child's? If this were then subjected to an appropriate course of education one would obtain the adult brain."
He is describing, in 1950, the paradigm that produced GPT, BERT, and every modern language model: begin with a general-purpose architecture (the "child-machine"), expose it to vast quantities of structured experience (training data), and let the patterns emerge from exposure rather than specification.
He even anticipates the role of reward and punishment in shaping the system — what we now call reinforcement learning:
"The use of punishments and rewards can at best be a part of the teaching process."
The entire modern ML pipeline — pre-training, fine-tuning, reinforcement learning from feedback — is outlined here, in a philosophy paper, in 1950. Not as engineering (the hardware did not exist), but as conceptual architecture. The engineering took seventy years to catch up with the concept.
What he did not foresee
One thing. The sheer scale required. Turing imagined this being achievable with modest resources — perhaps a few decades of work, running on the kind of hardware available in his lifetime. He could not have known that the threshold for convincing language use would require billions of parameters, terabytes of text, and planetary-scale computation. The idea was correct and complete. The cost was unimaginable.
This is itself a philosophical point. Some truths are conceptually accessible long before they are practically realisable. Turing could see the shape of the solution clearly — as clearly as Prabhākara could see, twelve centuries before the transformer, that meaning is relational. The structure of the problem was available to thought. The resources to instantiate the structure were not.
Why this paper matters now, more than in 1950
In 1950, the paper was speculative. It described machines that did not exist, making arguments about capabilities no one had witnessed. It could be — and was — dismissed as clever but untethered philosophy.
Today, every argument in the paper can be evaluated against working systems. The Imitation Game is no longer hypothetical; it is played millions of times daily. The Lady Lovelace objection is no longer abstract; it is contradicted by every unexpected model output. The learning-machine proposal is no longer speculative; it is the trillion-dollar architecture of the AI industry.
What has changed is not the arguments — they were already right. What has changed is that they can no longer be ignored.
The deeper point
Turing was a mathematician. His most famous result — the halting problem — is a theorem in mathematical logic. His wartime work was in cryptanalysis. Yet his most lasting intellectual contribution may be this philosophy paper, published in a philosophy journal, addressing a philosophical question.
He is another instance of the pattern this practice is built on: that the deepest workers in any discipline eventually arrive at philosophy. Not because they wander there by accident, but because the questions at the bottom of every field — what is computation? what is meaning? what is it to understand? — are philosophical questions wearing technical clothing.
Turing saw this. His paper is the proof.
References
Turing, A.M. (1950). "Computing Machinery and Intelligence." Mind, 59(236), 433–460. doi:10.1093/mind/LIX.236.433