Computing Machinery and intelligence (1950)

 

A. M. Turing

 

Heil, pp. 212-34

 

I.  The Imitation Game

A. Begins with the question, "Can machines think?"

1. to answer this question, would have to begin by defining "machine" and "think"

2. however, if we define these terms simply by describing how they're ordinarily used, then we end up merely reporting what most people think about the question (q.v.)

B. so instead, Turing proposes to replace that with another question:  could a computer play the imitation game?  

C. Imitation game consists of three players

1. two of opposite sex

2. a third who tries to guess which is which through question and answer

3. man attempts to deceive, woman to help

4. if third person guesses correctly, woman wins, if not, man does

5. if we can successfully substitute a computer for the man, the computer is intelligent (212-13, q.v.)

II. Critique of the New Problem (213)

A. one may argue that the odds are weighted against the computer or that the computer thinks differently than we do (q.v.)

B  Turing replies that if a machine can nevertheless play the imitation game satisfactorily, we need not worry about this objection (213-14)

III. The Machines Concerned in the Game (214)

A. Turing still needs to specify what he means by a machine

B. he makes it clear that

1. he means something produced by engineering

2. he does not mean a person "born in the usual manner" or a human clone

C. specifically, he permits only a digital computer

1. which of course were fairly new in 1950

2. but his question is not whether a 1950 machine could pass the test, but whether some imaginable machine could (214-15)

IV. Digital Computers (215)

a. For Turing, a digital computer is one that carries out operations according to fixed rules -- in other words, through algorithms

B. Turing goes on to explain concepts like "store," "executive unit," "control," and "programming," but in this day and age it hardly seems necessary for me to either explain these terms or to defend the notion that such machines are possible

c. One thing that you might not know is that Babbage attempted to build such a machine as early as the 1820s and 1830s (216-17)

V.  Universality of Digital Computers (217)

A. for Turing, a computer is a discrete state machine

1. although strictly speaking everything in the universe moves continuously

2. it is a convenient fiction to think of a computer as being in discrete states

3. if there are a finite number of such discrete states, a discrete state machine can be can be completely described in a table

4. Given this table, the initial state of the machine, and the inputs, it is possible to predict all future states of the machine (218)

B. Universal machines (219)

1. this predictability of a discrete state machine allows it to be mimicked by a digital computer

2. to say that a digital computer is a universal machine is to say that it can mimic any discrete state machine

3. that is, we can program a universal machine to mimic any other

C. The imitation game question then becomes:

1. given enough memory, speed, and the right program

2. can a digital computer successfully play the part of A in the imitation game, the part of B being taken by a man?  That is, tell man from machine, rather than man from woman as back on p. 212.

VI. Contrary Views on the Main Question (219)

A. predicts that in 50 years,

1. computers will play the imitation game so well that a human will be able to tell the computer from the person after 5 minutes only 70% of the time

2.  at that time, too, people will not find it strange to say that machines can think

3.  his prediction has turned out wrong:  no computer has yet passed his test

B. The Theological Objection: that thinking is a function of the soul and that machines and animals do not have souls (220)

Replies:

1. although he does not accept this theological view

2. he would rather draw the line between animals and machines than between people and animals

a. the line between animals and people is rather arbitrary

b. in fact, the line has been drawn at various times in such a way as to discriminate against women and minorities

3. also, isn't it to restrict God's power to say that he could not confer a soul on a machine or an animal? (cf. Locke)

4. Finally, as in the case of the Earth's motion, theological arguments have proved unsatisfactory in the past (220)

C. The "Heads in the Sand" Objection

1. this objection says that machine intelligence would have the dreadful consequence of robbing us of our special place in nature (220-21)

2. Turing finds this objection more worthy of consolation than refutation (221)

D. The Mathematical Objection

1. There are theorems such as Gödel's Theorem that say that in a logically consistent system there are (well-formed) statements that can be neither proved nor disproved within the system

2. Applied to computers, this says that there are certain things a machine can't do or questions that it cannot answer or will answer incorrectly, such as a question about a machine like itself

Reply

1.   nobody has ever proved that the same limitation does not apply to human beings (221-22)

2.   we too often give wrong answers to questions to feel superior to machines (222)

3. more seriously, it might be the case that for any given machine, we could tell that it gave the wrong answer, but it doesn't follow that we would therefore triumph over all possible machines (222, q.v.)

4. at least, people who raise this objection -- unlike the first two objections -- would be willing to discuss a criterion such as the imitation game

E. The Argument from Consciousness (222)

1. Basically says that machines don't feel anything:  see quotation

2. In reply, Turing says

a. this objection denies the validity of the imitation game test

b. suggests that this objection amounts to saying that the only way to know the machine is thinking is to be the machine and to feel it thinking

c. but this response risks solipsism

1.) which may be logically unassailable

2.) but makes communication with others difficult

3.) hence the “polite convention that everyone thinks”

d. when the risk of solipsism is pointed out to them, those who raise this objection would probably adopt some form of the Turing test

1.) something like the imitation game is used in oral examinations (viva voce) to see whether a person truly understands something:  see for example the conversation on pp. 222-23

2.) Turing thinks that if a machine gave answers this good, they could not be described as “an easy contrivance” (Jefferson’s objection)

3. He doesn't mean to deny that there is some problem about consciousness.  However, he does not think that we need to solve this problem in order to answer the question as to whether machines can play the imitation game (223)

F. Arguments from Various Disabilities (223)

1. this argument takes the form of saying that machines will never do X, when X could be:

a. enjoying ice cream

b. falling in love,

c. making mistakes,

d. display a diversity of behavior, etc. (see list)

2. typically this argument is based on induction, that is, is simply a generalization from past (limited) experience

3. As for making mistakes, we need to distinguish (224)

a. Errors of functioning

1.) when the machine breaks down in some mechanical or electrical fashion and which would probably give itself away

2.) an abstract computer is incapable of such errors. (224-25)

b. Errors of conclusion (225)

1.) this can happen only when we attach some meaning to the output

2.) if it says something false, we say it has committed an error of conclusion

3.) but there is no reason we could not program a computer to say something false

4.) alternatively, we could program a machine to draw conclusions by induction (225)

4. For Turing, the argument from disabilities is often just a disguised form of the argument from consciousness

G. Lady Lovelace's Objection (225)

1. she raised it against Babbage's analytical engine, claiming that it does not originate anything but only follows our orders (q.v.)

a. of course, even if this were true at the time of his machine, this doesn't show that we'll never make a machine that can think for itself (225-26)

b. what it shows is that there was no evidence available to Lady Lovelace that would lead her to believe machines could think (225)

c. Turing will consider this problem again under the heading of machine learning (226)

2. another version of this argument is to say that a machine can never really do anything new, but then who can say that we ever really do?

3. Still another version is to say that a machine will never surprise us

a. he replies that he's surprised all the time by machines

b. he anticipates the rejoinder that this surprise may be due to some creative mental act on his part, and reflects nothing about the machine

c. in reply,

1.) he suggests that once more we are faced with perhaps a hidden version of the argument from consciousness

2.) even to regard what a person does as surprising requires a “creative mental act”

3.) but he also suggests that it reflects a prejudice people have (226)

a.) to the effect that something that follows logically could never surprise us

b.) as if people can automatically see all the logical consequences of some fact, which is just not true

H. Argument from Continuity in the Nervous System

1. that is, that the nervous system is not a discrete state machine

2. Turing argues that if the conditions of the imitation game are adhered to, that the interrogator should not be able to tell the difference (227)

a. He makes an analogy with another sort of calculator, a "differential analyzer," which is not a discrete state machine

b. and argues that a digital machine could give fairly accurate predictions of the differential analyzer's answers

I. The Argument from Informality of Behavior

1. basically, this objection makes the following argument:

 

   If each person had a definite set of rules of conduct by which he/she regulated his or her life, he or she would be no better than a machine.

   There are no such rules.

    -------------------------------------------

   Therefore, people are not machines.

 

2. Now Turing agrees with the second premise: that it is not possible to find a set of rules of conduct that determines what somebody will do in every conceivable set of circumstances

3. nevertheless, the argument is illogical

a. as Turing says, the undistributed middle is glaring

b. actually, in the form in which it's presented, it looks like the fallacy of denying the antecedent

4. Turing also explains that there is some confusion over rules and laws (227)

a. we need to distinguish:

1.) rules of conduct, such as stopping at red lights, that you follow consciously

2.) laws of behavior, or laws of nature that apply to the human body

b. if we substitute laws for rules in the first premise, it’s true

c. however, if we make the substitution in the second premise, we don’t know that it’s true

5. whether or not human behavior follows laws is an open question that calls for scientific research: 

a. we're never going to be in a position to say we haven't found these laws and we've searched enough

b. nor can we ever be sure of finding such laws even when they do exist -- as he argues by defying the reader to figure out a program he's written (228, q.v.)

J. The Argument from Extra-Sensory Perception

1. Turing takes the possibility of things like telepathy seriously

2. Nor is he satisfied with the argument that science can safely ignore such things

3. Returning to the imitation game once again, he discusses the situation in which the interrogator is trying to decide between a computer and a telepathic or clairvoyant human (228)

a. for instance, the interrogator asks what suit of cards he is holding, and the human is right more often than the machine

b. however, if such powers existed, then we would simply have to make sure the test was carried out in a telepathy-proof room

c. in other words, it would be one more cue like tone of voice that we need to eliminate

VII. Learning Machines (229)

A. returns to Lady Lovelace's objection that a machine can only do what we tell it to do

B. For Turing, there seems to be no good argument other than to say that we need to wait and see what will happen in 50 years

C. In the mean time, he tends to see it as a question of programming.  He thinks it unlikely that the hardware of the future will be inadequate

1. it is already possible to increase memory

2. machines are already faster than nerve cells (230)

D. we also need to remind ourselves that we are comparing machines to adult humans.  Perhaps what we ought to try is to write a program to simulate the mind of a newborn

1. then, with the appropriate education, it may think like an adult (230)

2. the hope is that there is very little structure in the child's brain and it will be easy to program

E. Clearly, a machine won't learn in exactly the same way as a child (231)

1. punishment and reward won't have the same effect

2. but Turing suspects a lot of the learning that child does comes through "unemotional" channels

F. Turing also raises the issue of the complexity of the child-machine: specifically, should it include the ability to draw inferences from what it is taught?

1. Here the problem is that, from a logical point of view, there are an unlimited number of inferences we can draw from what we know (232)

2. real intelligence is shown in our ability to make choices about which inference to draw (q.v.)

G. Turing also argues intelligent behavior is not completely random behavior, either (233, q.v.)

H. Of course, this paper was written quite a while ago, and Turing could only speculate about such things as chess-playing programs, let alone ones that learn to play better chess