The Turing test

Table of Contents

Turing looking appropriately mischievious.

The Turing test is the name given to Alan Turing's test for mechanical intelligence. He called it the "Imitation game," which is described below. We also look at variations suggested by Daniel Dennett to support his claim that the Turing test is actually very strong. Turing's original description can be found in his 1950 paper, "Computing machinery and intelligence."

The imitation game

The game is played by three people: a man (A), a woman (B), and an interrogator (C) who may be of either sex. The interrogator stays in a separate room. The object of the game is for the interrogator to determine who of A and B (known to the interrogator as X and Y) is the man and who is the woman.

The interrogator asks questions of A and B such as: Will X please tell me the length of his or her hair?

A wants C (the interrogator) to come to the wrong conclusion. Thus, A may say his hair is long (which, we are supposing, gives C reason to think that A is a woman, i.e. that A is B).

Clearly, some measures must be taken so that C cannot simply see or hear who is the man and who is the woman. So conversation happens through a typewriter.

B, unlike A, wants C to come to the right conclusion. So, B may say, "I am the woman!" But so might A.

What's happening?

Before we talk about replacing one of these players with a machine, we should try to understand what Turing is doing here. Let's take it for granted (as I suppose he did) that men and women are quite different; that there is some fundamental distinction and a sophisticated interrogator might have a chance of telling them apart. (This must be Turing's belief because the imitation game is pointless if there is no reasonable way to distinguish the players, for example distinguishing identical twins with similar life experiences, neither of whom you know closely.)

So, given that men and women are very different, person A would succeed in confusing the interrogator (making C believe that X is B rather than the true answer, A), and similarly, person B would fail in convincing C of the truth, if A is a better imitator of a woman than B is, and furthermore that B's arguments to the contrary fail.

Thus, B may win (expose A as a man) if A is a bad imitator or B has really strong arguments.

Suppose C gets the right answer only p% of the time (perhaps with different interrogators but the same A and B players; Turing is unclear here). This performance p% is important to keep in mind.

Replace the man with a machine!

Now we are ready to understand why Turing set up this game to test machine intelligence. If we replace A with a machine (perhaps B is still a woman, perhaps not, but of course is a human), then we have a test for machine intelligence.

Now, the machine is trying to convince the interrogator that it is a human, but the human is not. The human is trying to convince the interrogator that he or she is a human, but the machine is not.

Just as, supposedly, a man would find it difficult to imitate a woman, a machine would probably find it difficult to imitate a human. Yet, if the machine is not only able to act like a human but also (or alternatively?) to counteract the arguments produced by the human, then the machine is quite intelligent indeed.

Some interrogator may be naive. Thus, I believe this performance p% is relevant. If the same machine and human players (or perhaps different human players) can convince different interrogators more often than p% of the time, then the machine has passed the test. The machine need not achieve perfect performance (p = 100%), just has to beat human performance. (Because, presumably, humans are the most intelligent creatures that can be used in this test.)

I don't believe there is any reason to suppose a man can successfully convince a judge he is a woman and the woman is a man more than 50% of the time. If a man could do this, then he acts like a woman better than most other women, and furthermore can counteract their arguments to the contrary. Because the woman is arguing as well, maximum performance near 25% is probably more appropriate. Anyway, the machine's goal is to beat that figure, not achieve perfect convincing. The existence of a machine that can consistently convince a judge that it is human and the human is a machine, even though the human is begging for mercy, is an amazing thought!

[Note, I've since read that this understanding of the Turing Test, in which the machine is trying to convince the judge it is a human, is a "neutered" version. Originally, it is said, Turing expected the machine to be tasked with convincing the judge that it is a woman, which is an easier task since the questions from the judge would not revolve around what it means to be a human. Read "The Voice of the Turtle: Whatever Happened to AI?" (PDF) by Lenat (2008).]

A historical note

Descartes somewhat "predicted" the usefulness of such a test, even though he believed machines would never be able to pass it, in principle (because he thought machines could never respond to their environments):

If there were machines which bore a resemblance to our bodies and imitated our actions as closely as possible for all practical purposes, we should still have two very certain means of recognizing that they were not real men. The first is that they could never use words, or put together signs, as we do in order to declare our thoughts to others. For we can certainly conceive of a machine so constructed that it utters words, and even utters words that correspond to bodily actions causing a change in its organs. … But it is not conceivable that such a machine should produce different arrangements of words so as to give an appropriately meaningful answer to whatever is said in its presence, as the dullest of men can do. Secondly, even though some machines might do some things as well as we do them, or perhaps even better, they would inevitably fail in others, which would reveal that they are acting not from understanding, but only from the disposition of their organs. For whereas reason is a universal instrument, which can be used in all kinds of situations, these organs need some particular action; hence it is for all practical purposes impossible for a machine to have enough different organs to make it act in all the contingencies of life in the way in which our reason makes us act. — Source

The Turing test is quite strong

All quotes are from Dennett, "Can Machines Think?" (1985)

Remember, failure on the Turing test does not predict failure on those others, but success would surely predict success. His test was so severe, he thought, that nothing that could pass it fair and square would disappoint us in other quarters. Maybe it wouldn't do everything we hoped—maybe it wouldn't appreciate ballet, understand quantum physics, or have a good plan far world peace, but we'd all see that it was surely one of the intelligent, thinking entities in the neighborhood.


Terry Winograd, a leader in AI efforts to produce conversational ability in a computer, draws our attention to a pair of sentences. They differ in only one word. The first sentence is this: "The committee denied the group a parade permit because they advocated violence." Here's the second sentence: "The committee denied the group a parade permit because they feared violence."

The difference is just in the verb—"advocated" or "feared." As Winograd points out, the pronoun "they" in each sentence is officially ambiguous. Both readings of the pronoun are always legal. Thus, we can imagine a world in which governmental committees in charge of parade permits advocate violence in the streets and, for some strange reason, use this as their pretext for denying a parade permit. But the natural, reasonable, intelligent reading of the first sentence is that it's the group that advocated violence, and of the second, that it's the committee that feared the violence.

Now sentences like this are embedded in a conversation, the computer must figure out which reading of the pronoun is meant, it is to respond intelligently. But mere rules of grammar or vocabulary will not fix the right reading. What fixes the right reading for us is knowledge about politics, social circumstances, committees and their attitudes, groups that want to parade, how they tend to behave, and the like. One must know about the world, in short, to make sense of such a sentence.


The only way, it appears, for a computer to disambiguate that sentence and keep up its end of a conversation that uses that sentence would be for it to have a much more general ability to respond intelligently to information about social and political circumstances and many other topics. Thus, such sentences, by putting a demand on such abilities, are good quick probes. That is, they test for a wider competence.

More about "Winograd Schemas" can be found in the paper "The Winograd Schema Challenge" (PDF) by Levesque, Davis, and Morgenstern (2012).

Weaker varieties

Candidate 1

A computer is intelligent; it wins the World Chess Championship.

That's not a good test, it turns out. Chess prowess has proven to be an isolatable talent. There are programs today that can play fine chess but do nothing else. So the quick-probe assumption is false far the test of playing winning chess.

Candidate 2

The computer is intelligent; it solves the Arab-Israeli conflict.

This is surely a more severe test than Turing's. But it has some defects: passed once, it is unrepeatable; it is slow, no doubt; and it is not crisply clear what would count as passing it. Here's another prospect, then:

Candidate 3

A computer is intelligent; it succeeds in stealing the British crown jewels without the use of force or violence.

Now this is better. First, it could be repeated again and again, though of course each repeat test would presumably be harder, but this is a feature it shares with the Turing test. Second, the mark of success is clear: either you've got the jewels to show for your efforts or you don't. But it is expensive and slow, a socially dubious caper at best, and no doubt luck would play too great a role.

Bringsjord's strategies

Selmer Bringsjord, in an article titled "If I Were Judge," in the book Parsing the Turing Test (2009), described some strategies he would take if he believed the Turing test was valid (which he does not):

Strategy 1

Require that each contestant pass an array of standardized IQ tests.

Strategy 2

Give each contestant problems like this one:

  • If there is a king in the hand, then there is an ace, or else if there is not a king in the hand, then there is an ace.
  • There is a king in the hand.
  • Therefore, what?

The answer is that there is not an ace in the hand. (I bet you got that wrong.)

If the contestant gets the right answer, that probably tells us the contestant is a human. A computer program, on the other hand, can easily derive the logically correct answer.

But a very intelligent machine, so intelligent it can appear stupid, may therefore be able to solve very difficult problems like logical paradoxes.

Strategy 3

Give each contestant supposed logical paradoxes, and ask them to solve the puzzle. Certain paradoxes fool machine reasoning but (very intelligent) mathematicians can solve them.

Strategy 4

Finally, if both contestants solve the paradoxes, resort to asking each contestant to write short stories based on a prompt like, "As Gregor Samsa awoke one morning from uneasy dreams he found himself transformed in his bed into a gigantic insect."


Figure 1: Metamorphosis of Kafka by James LeGros. Acrylic on Watercolor Board.

We have a long way to go

Nobody actually practices with a real Turing test. It's too hard for the machine. Instead, the yearly Loebner Prize experiments with a reduced form of this test. The judges simply estimate how "closely" a computer imitates a human, and judges are not allowed to attempt trickery (such as typing random symbols) or delve into deep conversation.

"Computers take the Turing Test in Reading in October 2008 to see if judges believe the machines are human during instant message conversations." (Source)

The Turing test in culture

Seated where he could catch the readings on the two gauges of the Voigt-Kampff testing apparatus, Rick Deckard said, "I'm going to outline a number of social situations. You are to express your reaction to each as quickly as possible. You will be timed, of course."

"And of course," Rachael said distantly, "my verbal responses won't count. It's solely the eye-muscle and capillary reaction that you'll use as indices. But I'll answer; I want to go through this and — " She broke off. "Go ahead, Mr. Deckard."

Rick, selecting question three, said, "You are given a calfskin wallet on your birthday." Both gauges immediately registered past the green and onto the red; the needles swung violently and then subsided.

"I wouldn't accept it," Rachael said. "Also I'd report the person who gave it to me to the police."

After making a jot of notation Rick continued, turning to the eighth question of the Voigt- Kampff profile scale. "You have a little boy and he shows you his butterfly collection, including his killing jar."

"I'd take him to the doctor." Rachael's voice was low but firm. Again the twin gauges registered, but this time not so far. He made a note of that, too.

"You're sitting watching TV," he continued, "and suddenly you discover a wasp crawling on your wrist."

Rachael said, "I'd kill it." The gauges, this time, registered almost nothing: only a feeble and momentary tremor. He noted that and hunted cautiously for the next question.

— from Do Androids Dream of Electric Sheep?, Phillip K. Dick


Talking to machines (Radiolab, May 31, 2011)

The Turing Problem (Radiolab, March 19, 2012)

Intro to AI material by Joshua Eckroth is licensed under a Creative Commons Attribution-ShareAlike 3.0 Unported License. Source code for this website available at GitHub.