Computing Machinery and Intelligence

Turing opens the paper with a question that sounds simple but is actually a philosophical trap: "Can machines think?" The words "machine" and "think" are so loaded with assumptions that any direct answer invites endless debate.

Rather than define thinking, Turing pulls a brilliant move: he replaces the question with a game. If a machine can behave indistinguishably from a human in a specific, controlled setting, the question of whether it "really" thinks becomes as meaningless as asking whether another person "really" thinks — since we only have behavioral evidence in both cases.

The game has three participants: a human judge (the interrogator), a human foil, and a machine. The judge sits in a separate room and communicates with the other two only through typed text — no voice, no appearance. The judge's task is to figure out which respondent is human and which is the machine.

The machine's goal is to fool the judge into thinking it is human. The human foil's goal is to help the judge make the correct identification.

Critical design choices:

• Text-only — strips away appearance, voice, and physical form so the test focuses purely on intellectual behavior.
• Comparative — the machine is not measured in isolation but side-by-side with a human, which automatically calibrates the difficulty.
• Operational — the outcome is a measurable event (the judge's verdict), not a subjective philosophical judgment.

Turing predicted that by the year 2000, a machine with about 10⁹ bits of storage could fool an average interrogator 30% of the time in a five-minute conversation.

Turing specifies exactly which kind of machine he means: a digital computer — a device with a store (memory), an executive unit (processor), and a control (instructions). He explains that digital computers are discrete-state machines: they jump between definite states, unlike analog systems that vary continuously.

The crucial insight is universality: a single digital computer, given the right program, can imitate any other discrete-state machine. This means we do not need to build a new machine for each task — we only need to write a new program. The question "Can machines think?" narrows to: "Can a suitably programmed digital computer pass the Imitation Game?"

This was radical in 1950. The first stored-program computer (the Manchester Mark 1) was barely a year old, and most people had never heard of computers, let alone considered them capable of imitating human conversation.

The largest section of the paper is a systematic demolition of every argument Turing could imagine against machine intelligence. He lists nine objections and dismantles each one, sometimes with a single devastating sentence. This section is remarkable because nearly every debate about AI that has occurred since 1950 was anticipated here.

In the final section, Turing shifts from philosophy to engineering. He argues that instead of programming a machine with adult-level knowledge — an impossible task — we should build a "child machine" with a simple starting structure, then educate it.

The child machine would learn through a process strikingly similar to what we now call reinforcement learning: it receives rewards and punishments from a teacher, and gradually modifies its own behavior. Turing even anticipated the distinction between the teacher's direct influence and the machine's own autonomous exploration.

He also noted a crucial property that would later become central to machine learning: the teacher "will often be very largely ignorant of quite what is going on inside" the machine. The machine develops internal strategies that its creator did not explicitly program — what we now call emergent behavior.

Turing's paper has seven sections, each building on the last in a carefully constructed argument. The flow moves from posing the question, to reframing it as the Imitation Game, to defining the type of machine (digital computers), to defending the idea against nine objections, and finally to sketching how such machines might be built through learning.

Turing's paper did not solve the question of machine intelligence — but it did something more important: it made the question tractable. By shifting from metaphysics to measurement, from "what is thinking?" to "what would convince us?", he gave the field its first foothold. The Dartmouth Proposal built directly on this foundation, and ELIZA was the first system to demonstrate that Turing's game was not just a thought experiment but a practical challenge.