Computing Machinery and Intelligence
Source: https://academic.oup.com/mind/article/LIX/236/433/986238 ↗
Turing's 1950 paper posed the question "Can machines think?" and then methodically dismantled every common objection, from theological arguments to Lady Lovelace's claim that machines can only do what they are told.
The "imitation game" he proposed — now called the Turing test — was not meant as a definitive criterion for intelligence but as a way to replace a vague philosophical question with a concrete experimental one.
The paper is far more subtle than the popular version suggests: Turing discussed learning machines, the role of randomness, and the limits of mathematical logic with remarkable prescience.
He anticipated objections that would dominate AI philosophy for decades, including Searle's Chinese Room argument, without needing to name them.
It remains the most important philosophical text on artificial intelligence, and most people who cite it have not actually read it.
Central argument
Turing argues that the question 'can machines think?' is too philosophically muddled to answer directly, so he proposes replacing it with a concrete operational test: can a machine sustain written conversation well enough to be indistinguishable from a human? He then systematically dismantles objections — theological, mathematical, and empirical — including the argument that machines can only do what they are programmed to do, which he counters by introducing the concept of learning machines capable of surprising their creators. His deeper claim is that intelligence should be judged by observable behaviour, not by substrate or origin.
Critique
The move from 'can machines think?' to 'can machines imitate?' sidesteps rather than resolves the hard problem: a system could pass the imitation game while having no understanding whatsoever, which is precisely the force of Searle's later Chinese Room argument that Turing anticipates but does not fully defeat. Turing's behaviourist framing treats the question of machine intelligence as settled once external performance criteria are met, but this leaves open whether the test is measuring intelligence or merely the capacity to model human linguistic convention — a distinction that matters enormously as language models now pass crude versions of the test routinely.
Why it matters for product
Product leaders building or integrating AI features face exactly the operational trap Turing identified: teams tend to debate what AI 'really' understands instead of defining measurable behavioural criteria for what constitutes acceptable performance in context, which stalls roadmaps and muddies success metrics. More pointedly, Turing's argument about learning machines — systems that improve through experience rather than explicit instruction — has direct implications for how product teams should design feedback loops and evaluate outputs iteratively rather than treating AI capability as a fixed input to a feature spec.