~ Artificial Intelligence Consciousness

The Turing Test Was Never the Point

Most people who reference the Turing test have not read the paper it comes from. This is not a criticism. The paper is from 1950 and reads like a different era of philosophy, which it is. But the gap between what Turing actually argued and what the "Turing test" became in popular culture is wide enough to matter.

The paper is called "Computing Machinery and Intelligence." It was published in the journal Mind. The first sentence is: "I propose to consider the question, 'Can machines think?'" And then Turing immediately argues that this question is almost meaningless. The word "think" is too vague to support a rigorous answer. It depends on definitions that we cannot agree on. Rather than trying to define thinking and then asking whether machines can do it, Turing proposed replacing the question with a different one that might actually be answerable.

That replacement is the imitation game. A human judge communicates via text with two hidden entities, one human and one machine. If the judge cannot reliably tell which is which, the machine has passed. Turing was not claiming this proved the machine could think. He was claiming that this question, unlike "can machines think," was at least well-defined enough to have an answer.

This is a much more careful and more modest claim than what the Turing test became.

What It Became

In popular culture and in much of AI research, the Turing test became a finish line. Build a machine that fools a human judge, and you have achieved artificial intelligence. The test became a goal rather than a methodological proposal. Entire research programs were organized around it. Chatbots were built specifically to game conversational expectations. The Loebner Prize offered annual competitions. The question shifted from "is this a useful way to frame the problem" to "has a machine passed yet."

This transformation stripped out everything interesting about Turing's original argument. Turing was engaging with a philosophical problem about the limits of language and definition. The popularized version turned it into an engineering challenge. Turing was asking "can we find a meaningful question to replace an ill-defined one." The popularized version assumed the replacement question was meaningful and started trying to answer it.

The result is that we now have systems that can convincingly imitate human conversation and the test tells us almost nothing about what we actually want to know. Large language models can pass the imitation game with ease. They produce text that is indistinguishable from human text in many contexts. By the criteria of the popularized Turing test, the question is settled. But nobody believes the question is settled, because the thing we actually care about was never captured by the test in the first place.

What Turing Actually Anticipated

The remarkable thing about the 1950 paper is how many objections Turing anticipated and addressed. He considered the theological objection (thinking is a function of the soul, which God gave only to humans). He considered the consciousness objection (a machine might simulate behavior without having experience). He considered the mathematical objection (Godel's incompleteness theorems limit what machines can prove). He considered the argument from disability (machines cannot enjoy strawberries, fall in love, learn from experience).

For most of these, Turing's response was essentially: you are smuggling assumptions about what thinking requires that you have not justified. You assume thinking requires consciousness, but you cannot prove that other humans are conscious either. You accept their consciousness based on behavior. Why not apply the same standard to machines?

This is not an argument that machines are conscious. It is an argument that the behavioral evidence we use to attribute consciousness to other humans is exactly the same kind of evidence we refuse to accept from machines. The inconsistency is ours, not the machine's.

Turing also predicted that by the year 2000, computers would be able to play the imitation game well enough to fool an average human judge about 30 percent of the time. He was roughly right about the timeline, though the methods that got us there were not the ones he probably imagined. He thought it would take about 10^9 bits of memory. He was off by several orders of magnitude, but in an era when 10^9 bits was an almost unimaginable amount of storage, the estimate shows he understood the scale of the problem.

What We Should Take From This

The useful lesson from Turing's paper is not about the test. It is about the method. When a question is too vague to answer, replace it with a more precise one and see whether that helps. When people argue about whether machines can "think" or "understand" or "feel," they are often arguing about definitions rather than about the world. Turing's move was to sidestep the definitional argument and ask a behavioral question instead.

The limitation of that move is real. Behavioral evidence for internal states is always ambiguous. A perfect imitation of pain is not the same as pain. A perfect imitation of understanding is not the same as understanding. Or maybe it is. We do not know, and behavioral tests alone will not resolve this.

But the deeper point stands. When you find yourself in a debate where both sides are confident and neither side can specify what would change their mind, the problem is probably with the question. Turing saw this in 1950. We are still catching up.

The test was never a measure of machine intelligence. It was a demonstration that we do not know what we mean by intelligence, and that this confusion, not the capability of machines, is the bottleneck. Seventy-five years later, we have built machines that pass the test and the confusion remains exactly where Turing found it.