Mind Association

Computing Machinery and Intelligence Author(s): A. M. Turing Reviewed work(s): Source: Mind, New Series, Vol. 59, No. 236 (Oct., 1950), pp. 433-460 Published by: Oxford University Press on behalf of the Mind Association Stable URL: http://www.jstor.org/stable/2251299 . Accessed: 24/11/2011 09:18
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Oxford University Press and Mind Association are collaborating with JSTOR to digitize, preserve and extend access to Mind.

http://www.jstor.org

VOL.

LIX.

No. 236.]

1950 [October,

M IND
A QUARTERLY
OF

REVIEW

PSYCHOLOGY

AND PHILOSOPHY
AND

MACHINERY I.-COMPUTING INTELLIGENCE

BY A. M. TURING

Game. 1. The Imitation I PROPOSE tO considerthe question, 'Can machinesthink? of the meaningof the terms This shouldbeginwith definitions be framed 'machine' and 'think'. The definitions so as to might reflect so far as possiblethe normaluse of the words,but this attitudeis dangerous. If the meaningof the words'machine' and 'think'are to be found howtheyare commonly by examining used it is difficult to escape the conclusionthat the meaning to tlie question,' Can machines and the answer think? ' is to be such as a Gallup poll. But thisis soughtin a statistical survey sucha definition I shallreplace absurd. Insteadofattempting the whichis closely relatedto it and is expressed question by another, in relatively words. unambiguous The new formof the problemcan be describedin termsof a game whichwe call the 'imitationgame'. It is playedwith a man (A), a woman(B), and an interrogator threepeople, (C) who may be of eithersex. The interrogator stays in a room apart from two. The objectofthe game forthe interrogator the other is to determine whichof the othertwo is the man and whichis the woman. He knowsthemby labels X and Y, and at the end ' X is A and Y is B ' or ' X is B and Y ofthe gamehe says either is A'. The interrogator is allowedto put questions to A and B thus: C: Will X pleasetell me the length ofhis or herhair 2 Now suppose X is aetually A, then A must answer. It is A's
28 433

434

A. M. TURING:

objectin the game to tryand cause C to make the wrong identification. His answermight therefore be 'My hair is shingled, and the longeststrandsare about nine incheslong.' In orderthat tones of voice may not help the interrogator the answersshouldbe written, or betterstill,typewritten.The ideal arrangeimaent is to have a teleprinter communicating between the two rooms. Alternatively the questionand answerscan be repeated byan intermediary.The objectofthegamefor thethird player(B) is to help the interrogator.The best strategy forher is probablyto give truthful answers. She can add such things as I am the woman,don't listento him! ' to heranswers, but it willavail nothing as the man can makesimilar remarks. We now ask the question,'What willhappenwhena machine takes thepartof A in thisgame ? ' Will the interrogator decide wrongly as oftenwhen the game is played like this as he does whenthe game is played between a man and a woman? These questions replaceour original, 'Can machines think? 2. Critique oftheNew Problem. As well as asking,'What is the answerto thisnewform ofthe question', one may ask, 'Is this new questiona worthy one to investigate ? ' This latter question we investigatewithout further shortan infinite ado, thereby cutting regress. The new problem has the advantageof drawing a fairly sharp line between and theintellectual thephysical capacitiesofa man. No engineer or chemistclaims to be able to producea material fromthe human skin. It is possible whichis indistinguishable be done,but evensupposing that at sometimethismight thisinventionavailable we shouldfeeltherewas littlepointin trying it up in to make a 'thinking machine' morehumanby dressing in whichwe, have set the problem suchartificial flesh. The form whichprevents the interrogator reflects thisfactin the condition or hearingtheir fromseeingor touching the othercompetitors, voices. Some otheradvantagesofthe proposed criterion may be shownup by specimen questionsand answers. Thus: Q: Please write me a sonnet on the subject of the Forth Bridge. A: Countme out on thisone. I nevercouldwrite poetry. Q: Add 34957to 70764 A: (Pause about 30 secondsand thengive as answer)105621. Q: Do you play chess? A: Yes..

COMPUTING

MACHINERY

AND

INTELLIGENCE

435

Q: I have K at my Ki, and no otherpieces. You have only K at K6 and R at RI. It is yourmove. What do you play ? A: (After a pause of 15 seconds)R-R8 mate. The question and answer method seems to be suitable for introducing almostany one ofthe fields ofhumanendeavour that we wish to include. We do not wish to penalisethe machine for its inability to shinein beautycompetitions, norto penalise a man forlosingin a race againstan aeroplane. The conditions of our game make thesedisabilities irrelevant. The 'witnesses' can brag, if they consider it advisable,as much as theyplease about their charms,strength or heroism, but the interrogator cannotdemandpracticaldemonstrations. The game may perhapsbe criticised on the groundthat the odds are weighted too heavilyagainstthe machine. If the man wereto tryand pretend to be the machine he wouldclearly make a verypoorshowing. He wouldbe given awayat oncebyslowness and inaccuracy in arithmetic. May notmachines carry out somethingwhichoughtto be described as thinking but whichis very different from whata man does ? This objection is a verystrong one, but at least we can say that if,nevertheless, a machinecan be constructed to play the imitation game satisfactorily, we need not be troubled by thisobjection. It mightbe urged that when playingthe 'imitationgame' the best strategy for the machinemay possiblybe something other thanimitation ofthebehaviour ofa man. Thismaybe,but I thinkit is unlikely that thereis any greateffect of this kind. In any case thereis no intention to investigate herethe theory of the game, and it will be assumed that the best strategy is to tryto provide be givenby a man. answers thatwouldnaturally 3. The Mlachines in theGame. concerned The questionwhichwe put in ? 1 will not be quite definite untilwe have specified what we mean by the word 'machine'. wish It is natural thatweshould to permit kindofengineering every to be used in ourmachines. We also wishto allowthe technique or team of engineers than an engineer possibility may construct a machinewhichworks, but whosemannerof operation cannot be satisfactorily described becausetheyhave by its constructors applied a methodwhich is largelyexperimental.Finally, we menbornin theusual manner. wishto excludefrom themachines It is difficult the definitions so as to satisfy to frame thesethree conditions. One might for instance insist that the team of

436

A. M. TURING:

engineers should be all of one sex, but this would not really be satisfactory, for it is probablypossible to rear a complete individualfrom a singlecell of the skin (say) of a man. To do so would be a feat of biologicaltechniquedeserving of the very highestpraise,but we would not be inclinedto regardit as a a thinking case of 'constructing us to machine'. This prompts abandon the requirement that everykind of techniqueshould be permitted. We are the moreready to do so in view of the in 'thinkingmachines' has been fact that the presentinterest aroused by a particularkind of machine,usually called an ' electroniccomputer' or 'digital computer'. Followingthis suggestion we onlypermit digitalcomputers to take part in our game. This restriction appears at first sightto be a verydrasticone. I shall attempt to showthat it is not so in reality. To do this necessitates a shortaccountofthe natureand properties ofthese computers. of machineswith It may also be said that this identification like our criterion for 'thinking', will only digital computers, if (contrary be unsatisfactory to my belief),it turnsout that are unableto give a good showing in the game. digitalcomputers in working There are already a numberof digital computers and it maybe asked,'Why nottrythe experiment order, straight of the game. the conditions away ? It would be easy to satisfy of interrogators could be used, and statistics A number compiled theright identification was given.' Theshort to showhow often all digitalcomputers answeris that we are not askingwhether the computers at present woulddo wellin the game norwhether thereare imaginablecomavailable would do well,but whether answer. puterswhichwould do well. But this is onlythe short' in a diflerent We shall see thisquestion lightlater. 4. DigitalComputers. The idea behinddigitalcomputers maybe explained by saying that these machinesare intendedto carryout any operations whichcouldbe donebya human computer. Thehumancomputer fixedrules; he has no authority is supposedto be following themin any detail. We may supposethat these to deviatefrom rulesare suppliedin a book,whichis alteredwhenever he is put on to a new job. He has also an unlimited supplyof paper on which he doeshiscalculations. He mayalso do hismultiplications on a 'desk machine',but thisis not important. and additions as a definition If we use the above explanation we shall be in

COMPUTING

MACHINERY

AND INTELLIGENCE

437

danger of circularity of argument. We avoid this by giving an outlineof the means by whichthe desiredeffect is achieved. A digitalcomputer can usuallybe regarded as consisting ofthree parts: (i) Store. (ii) Executiveunit. (iii) Control. The storeis a storeofinformation, and corresponds to thehuman computer's paper,whether thisis the paper on whichhe does his calculations or that on whichhis book of rulesis printed. In so faras the humancomputer does calculations in his head a partof the storewillcorrespond to his memory. The executiveunit is the part whichcarriesout the various individual operationsinvolved in a calculation. What these individualoperationsare will vary frommachineto machine. Usually fairly lengthy operations can be done such as 'Multiply 3540675445by 7076345687' but in some machinesonly very simpleonessuchas 'Write down0 ' are possible. We have mentioned that the ' book of rules-'suppliedto the is replacedin the machineby a part of the store. It computer is then called the 'table of instructions'. It is the dutyof the control to see that theseinstructions are obeyedcorrectly and in theright order. The control is so constructed thatthisnecessarily happens. The information in the storeis usuallybroken up intopackets of moderately small size. In one machine, forinstance, a packet often decimaldigits. Numbers mightconsist are assigned to the parts of the storein whichthe various packets of information are stored,in some systematic manner. A typical instruction mightsay'Add the number storedin position6809 to that in 4302 and put the resultback into the latterstorageposition'. Needless to say it would not occurin the machineexpressed in English. It would more likelybe coded in a formsuch as 6809430217. Here 17 says whichof variouspossibleoperations is to be performed on the two numbers. In this case the operation is that describedabove, viz. 'Add the number.. . .' It will be noticedthat the instruction takes up 10 digits and so forms one packetofinformation, veryconveniently.The control will normally take the instructions to be obeyedin the orderof the positionsin whichthey are stored,but occasionallyan instruction such as

438

A. M. TURING:

' Now obey the instruction storedin position5606, and continuefrom there' or again may be encountered, stored 'If position4505 contains0 obey next the instruction in 6707,otherwise on.' continue straight becausethey Instructions oftheselattertypesare veryimportant make it possiblefora sequenceof operations to be repeatedover is fulfilled, but in doingso and over again until some condition on each repetition, instructions but the same to obey,not fresh onesoverand overagain. To take a domestic analogy. Suppose on MotherwantsTommyto call at the cobbler'severymorning his way to schoolto see if her shoes are done, she can ask him she can stick up a notice afresheverymorning. Alternatively once and forall in the hall whichhe will see whenhe leaves for theshoes, and also to destroy schooland which tellshimto call for the noticewhenhe comesback ifhe has the shoeswithhim. can The readermustacceptit as a factthat digitalcomputers be constructed, and indeed have been constructed, according and that they can in fact we have described, to the principles veryclosely. mimic the actionsof a humancomputer we have described ourhumancomputer The bookofruleswhich as usingis of coursea convenient fiction. Actual human comremember puters really whattheyhavegotto do. If onewantsto make a machinemimicthe behaviourof the human computer in somecomplex operation one has to ask himhowit is done,and then translate the answerinto the formof an instruction table. Constructing instruction tables is usually describedas 'proto carryout the operagramming'. To 'programme a machine table into the tion A' meansto put the appropriate instruction machineso thatit willdo A. is a An interesting varianton the idea of a digital computer element'.Thesehaveinstructions 'digitalcomputer witha random involving the throwing of a die or some equivalentelectronic for instance be,' Throw thedie process; onesuchinstruction might such and put the resulting number into store1000'. Sometimes a machineis described as havingfreewill (thoughI would not use thisphrasemyself). It is not normally possibleto determine fromobserving it has a random element, a machinewhether fora similareffect can be producedby such devicesas making the choicesdependon the digitsof the decimalforw. have onlya finite store. There Most actual digitalcomputers is no theoretical in the idea of a computer withan undifficulty part can have been used limitedstore. Of courseonly a finite amountcan have been at any one time. Likewiseonly a finite

COMPUTING

MACHINERY

AND INTELLIGENCE

439

constructed, but we can imaginemoreand morebeingadded as required. Such computers have special theoretical interest and will be called infinitive capacitycomputers. The idea ofa digitalcomputer is an old one. Charles Babbage, Lucasian Professor of Mathematics at Cambiidgefrom1828 to 1839, planned such a machine,called the AnalyticalEngine, but it was never completed. AlthoughBabbage had all the essentialideas, his machinewas not at that time such a very attractive prospect. The speedwhichwouldhave been available wouldbe definitely faster than a humancomputer but something like 100 timesslowerthan the Manchester machine, itselfone of the slower of the modernmachines. The storage was to be purelymechanical, usingwheelsand cards. The factthat Babbage's AnalyticalEnginewas to be entirely willhelpus to ridourselves mechanical ofa supers,tition. Importance is often attachedto the factthat modern digitalcomputers are electrical, and thatthenervous system also is electrical. Sinc', Babbage's machinewas not electrical, and sinceall digitalcomputersare in a senseequivalent, we see thatthisuse ofelectricity cannotbe oftheoretical importance. Ofcourseelectricity usually is concerned, comes in wherefast signalling so that it is not surprising that we find it in both these connections. In the nervous systemchemicalphenomenaare at least as important as electrical. In certaincomputers the storagesystem is mainly acoustic. The featureof using electricity is thus seen to be only a very superficial similarity. If we wish to find such we should look rather similarities formathematical analogiesof function. 5. Universality ofDigitalComputers. in the last sectionmay be considered The digital computers '. These are the state machines classified the ' discrete amongst one quite machineswhichmove by suddenjumps or clicksfrom for different definite stateto another. Thesestatesare sufficiently between themto be ignored. Strictly the possibility ofconfusion speakingthereare no such machines. Everything reallymoves continuously.But thereare many kindsof machinewhichcan be thought profitably of as being discretestate machines. For for a lighting the switches instancein considering systemit is on or a convenient fiction that each switchmust be definitely off. There must be intermediate definitely positions,but for about them. As an example of a most purposeswe can forget discretestate machinewe mightconsidera wheel whichclicks

440

A.

M.

TURING:

round through120? once a second,but may be stopped by a leverwhichcan be operatedfrom outside; in additiona lamp is to lightin one ofthe positions ofthe wheel. This machine could be describedabstractlyas follows. The internalstate of the machine(whichis described by the position of thewheel)maybe of lever). ql, q2 or q3. Thereis an inputsignali0 or i1 (position The internal stateat any moment is determined by the last state and inputsignalaccording to the table Last State
q1 i0 q2 q2 q3 q2 q3 q1 q3

Input

?i

q1

The output signals, the only externallyvisible indicationof the internal state (the light)are described by the table

State q1
Output
00

q2 0?

q3

01

This exampleis typicalof discrete state machines. They can be number described by suchtablesprovided theyhave onlya finite ofpossiblestates. It will seem that given the initial state of the machineand all future the inputsignalsit is alwayspossibleto predict states. This is reminiscent of Laplace's view that fromthe complete state of the universe at one moment oftime,as described by the it should be possibleto positionsand velocitiesof all particles, we are considering predict all future which states. The predictioil is, however, rather than that considered nearerto practicability by Laplace. The systemof the 'universe as a whole' is such that quite small errorsin the initial conditionscan have an of a overwhelming effectat a later time. The displacement single electronby a billionthof a centimetre at one moment mightmake the difference betweena man being killed by an avalanche a year later,or escaping. It is an essentialproperty of the mechanicalsystems whichwe have called 'discrete state machines' that thisphenomenon does notoccur. Even whenwe considerthe actual physicalmachinesinstead of the idealised machines,reasonablyaccurate knowledgeof the state at one momentyields reasonablyaccurate knowledgeany numberof stepslater.

COMPUTING

MACHINERY

AND

INTELLIGENCE

441

fall within the class As we have mentioned, digitalcomputers of discretestate machines. But the numberof states of which large. For such a machine is capable is usually enormously at Manchester instance, the number forthemachine nowworking thiswithourexample it about 2165,000, i.e. about 1505?.??. Compare of the clicking wheel describedabove, which had three states. to see whythe number of states shouldbe so It is not difficult includesa store corresponding to the immense. The computer paper used by a humancomputer. It mustbe possibleto write of symbolswhich into the store any one of the combinations on the paper. For simplicity suppose mighthave been written 0 to 9 are used as symbols. Variations in that onlydigitsfrom are ignored. Suppose the computer is allowed 100 handwriting 50 lines each withroomfor30 sheetsof paper each containing digits. Then the numberof states is 10100X5OX30, i.e. 10150,000. of statesof threeManchester This is about the number machines to the base two of the number put together. The logarithm of statesis usuallycalled the 'storage capacity' of the machine. machinehas a storagecapacity of about Thus the Manchester 165,000 and the wheel machineof our example about 1-6. If their capacitiesmust be added two machinesare put together to obtain the capacityof the resultant machine. This leads to of statements such as 'The Manchester machine the possibility contains64 magnetic trackseach witha capacityof 2560, eight storage electronic tubes with a capacityof 1280. Miscellaneous a total of 174,380.' to about 300 making amounts to a discretestate machineit Given the table corresponding is possibleto predictwhat it will do. There is no reason why this calculationshouldnot be carriedout by means of a digital quickly computer. Providedit could be carriedout sufficiently could mimicthe behaviour of any discrete the digitalcomputer the statemachine. The imitation gamecouldthenbe playedwith machinein question(as B) and the mimicking digitalcomputer wouldbe unable to distinguish them. (as A) and the interrogator must have an adequate storage Of coursethe digital computer fast. Moreover, it must capacity as well as working sufficiently foreach new machinewhichit is desired afresh be programmed to mimic. This special propertyof digital computers, that they can mimic any discrete state machine, is described by saying machines. The existenceof machines that they are universal has the important with this property consequencethat, consiof speed apart,it is unnecessary to designvariousnew derations machinesto do various computing processes. They can all. be

442

A. M. TURING:

done with one digital computer, suitablyprogrammed foreach case. It willbe seenthatas a consequence ofthisall digitalcomputersare in a senseequivalent. We may now consideragain the pointraisedat the end of ?3. It was suggested tentatively that the question,'Can machines think? ' should be replaced by 'Are there imaginabledigital computerswhich would do well in the imitationgame ? ' If we wish we can make this superficially more generaland ask 'Are there discrete state machines which would do well? ' But in view of the universality property we see that eitherof these questionsis equivalentto this, 'Let us fix our attention on oneparticular digitalcomputer C. Is it truethatbymodifying this computer to have an adequate storage, its suitably increasing speedofaction, and providing it withan appropriate programme, C can be madeto play satisfactorily thepartofA in theimitation game,the partof B beingtakenby a man ? 6. Contrary Viewson theMain Question. We may now consider the groundto have been clearedanldwe are readyto proceed to thedebateon ourquestion, 'Can machines think? ' and thevariantofit quotedat the endofthelast section. We cannotaltogether abandonthe original form ofthe problem, for opinionswill differ as to the appropriateness of the substitution and we must at least listento what has to be said in this connexion. It willsimplify matters forthe readerif I explainfirst my own in thematter. Consider beliefs form first themoreaccurate ofthe question. I believe thatin a-bout fifty years'timeit willbe possible to programme computers, witha storagecapacityof about 109, to make themplay the imitation game so well that an average interrogator willnothave more than70 percent.chanceofmaking the rightidentification afterfiveminutesof questioning. The originalquestion,'Can machinesthink ? ' I believe to be too to deserve I believethat at meaningless discussion. Nevertheless theendofthecentury theuseofwords andgeneral educated opinion willhavealtered so much thatonewillbe able to speakofmachines without thinking expecting to be contradicted. I believefurther that no useful purpose is served by concealingthese beliefs. The popularview that scientists proceedinexorably fromwellestablishedfact to well-established fact,neverbeing influenced is quite mistaken. Providedit is by any unprovedconjecture, made clear whichare proved facts and whichare conjectures, no harm can result. Conjectures are of greatimportance since linesof research. theysuggest useful

COMPUTING

MACHINERY

AND INTELLIGENCE

443

I now proceedto consider opinions opposedto my own. (1) The Theological Objection. Thinking is a function of man's immortal soul. Godhas givenan immortal soul to every manand woman,but not to any otheranimal or to machines. Hence no animal or machine can think. I am unable to accept any part of this, but will attemptto reply in theological terms. I should findthe argument more ifanimalswereclassedwithmen,forthereis a greater convincing to my mind, betweenthe typical animate and the difference, inanimatethan there is betweenman and the other animals. The arbitrary character of the orthodox view becomesclearerif we consider how it mightappear to a memberof some other religious community.How do Christians regard theMoslemview that womenhave no souls? But let us leave this point aside and return to the main argument. It appears to me that the argument quotedabove impliesa seriousrestriction of the omnipotenceof the Almighty. It is admittedthat thereare certain thingsthat He cannotdo such as makingone equal to two,but should we not believethat He has freedom a soul on to confer an elephantif He sees fit? We mightexpectthat He would only exercise this powerin conjunction with a mutationwhich providedthe elephantwith an appropriately improvedbrainto minister to theneedsofthissoul. An argument ofexactlysimilar form maybe madefor thecase ofmachines. It mayseemdifferent because it is moredifficult to " swallow". But this reallyonly means that we thinkit would be less likelythat He would consider the circumstances suitableforconferring a soul. The cirin questionare discussedin the rest of this paper. cumstances In attempting to constructsuch machineswe should not be irreverently usurping His powerofcreating souls,anymorethan we are in the procreation of children: ratherwe a-re, in either case, instruments ofHis willproviding mansions for the souls that He creates. However,this is mere speculation. I am not veryimpressed withtheological arguments whatever theymaybe usedto support. Such arguments in the past. have often beenfound unsatisfactory In the time of Galileo it was arguedthat the texts, " And the sun stood still . . . and hasted not to go down about a whole day " (Joshuax. 13) and " He laid the foundations of the earth,
1Possibly thisviewis heretical.St. Thomas Aquinas(Summa Theologica, quoted by Bertrand Russell,p. 480) statesthat God cannotmake a man to have no soul. But thismaynotbe a realrestriction onHis powers, but only a resultof the fact that men's souls are immortal, and therefore indestructible.

444

A. M. TURING:

that it should not move at any time" (Psalm cv. 5) were an theory. Withour present ofthe Copernican adequate refutation appears futile. When that knowsuch an argument knowledge impression. ledgewas notavailableit made a quitedifferent (2) The 'Heads in theSand' Objection. " The consequences would be too dreadful. Let us hope and of machinesthinking believethattheycannotdo so." is seldomexpressedquite so openlyas in the This argument mostof us who thinkabout it at all. form above. But it affects to the We liketo believethatMan is in somesubtleway superior rest of creation. It is best if he can be shownto be necessarily his commanding ofhimlosing is no danger for thenthere superior, is clearly argument position. The popalarityof the theological in inwiththisfeeling. It is likelyto be quite strong connected more tellectualpeople, since they value the power of thinking highlythan others,and are more inclinedto base theirbelief ofMan on thispower. in the superiority substantial is sufficiently I do not thinkthat this argument would be moreappropriate: to requirerefutation. Consolation ofsouls. in the transmigration perhapsthisshouldbe sought ofresults Objection. Thereare a number (3) The Mathematical of mathematical logic which can be used to show that there machines. The to the powersof discrete-state are limitations and best knownof these resultsis knownas G6del's theorem,' statements logicalsystem powerful showsthat in any sufficiently be provednor disproved whichcan neither can be formulated is inconsistent. itself thesystem unlesspossibly within the system, resultsdue to Church, similar, in some respects Thereare other, Kleene,Rosser,and Turing. The latterresultis the most conwhereas to machines, directly sinceit refers venientto consider, indirect argument: can onlybe used in a comparatively the others is to be used we need in addition if G6del'stheorem forinstance logical systemsin termsof to have some means of describing in oflogicalsystems. The result and machines interms machines, a digital whichis essentially refers to a typeof machine question capacity. It states that there are computerwith an infinite cannotdo. If it is rigged up to thatsucha machine certain things game,therewill be as in the imitation to questions giveanswers answer,or fail givea wrong it willeither to which somequestions muchtimeis allowedfora reply. to givean answer at all however Theremay, of course,be many such questions,and questions whichcannotbe answered by one machinemay be satisfactorily
1Author's to the Bibliography. namesin italicsrefer

COMPUTING

MACHINEkY

AND INTELLIGENCE

445

for thepresent answered by another. We are ofcoursesupposing that the questionsare of the kind to whichan answer 'Yes ' rather than questionssuch as 'What do or 'No' is appropriate, you think of Picasso ?-' The questions that we know the the machine machinesmust fail on are of this type," Consider specified as follows.. . . Will this machineever answer'Yes' to any question? " The dots are to be replaced by a deswhich could be criptionof some machinein a standardform, like that used in ? 5. When the machinedescribed something simple relationto the machine bears a certaincomparatively it can be shownthat the answer whichis underinterrogation, This is the mathematical is eitherwrongor not forthcoming. to result: it is arguedthat it proves a disabilityof machines is not subject. whichthe humanintellect The short answer to this argumentis that although it is establishedthat there are limitationsto the powers of any particularmachine,it has only been stated, withoutany sort of proof, apply to the human intellect. thatno such limitations quite so lightly. But I do not thinkthis view can be dismissed Whenever one of these machines is asked the appropriate we knowthat this answer, and givesa definite criticalquestion, answer must be wrong,and this gives us a certainfeelingof superiority.Is this feelingillusory? It is no doubt quite genuine,buitI do not thinktoo much importanceshould be attached to it. We too oftengive wronganswersto questions in beingverypleased at suchevidenceof ourselves to be justified our superiority fallibility on the part ofthe machines. Further, to theonemachine can onlybe felt on suchan occasionin relation over whichwe have scored our petty triumph. There would over all machines. be no questionof triumphing simultaneously In short,then,there mightbe men clevererthan any given but thenagain theremightbe othermachinescleverer machine, again, and so on. I think, wouLld, argument Those whoholdto the mathematical game as a basis for mostlybe willingto accept the imitation discussion. Those who believe in the two previousobjections in any criteria. would probablynot be interested is very Consciousness.Thisargument (4) The Aryumentfrom Jefferson's Lister Orationfor 1949, well expressedin Professor from whichI quote. " Not untila machinecan writea sonnet or composea concertobecause of thoughtsand emotionsfelt, and not by the chance fall of symbols,could we agree that machineequals brain-that is, not onlywriteit but Inow that could feel (and not merely it. No mechanism it had written

446

A. M. TURING:

artificially signal,an easy contrivance) pleasureat its successes, griefwhen its valves fuse, be warmed by flattery, be made miserableby its mistakes, be charmedby sex, be angryor depressedwhenit cannot -getwhat it wants." This argument appears to be a denial of the validityof our test. According to the mostextreme form ofthisview the only way by whichone could be surethat a machinethinksis to be the machineand to feel oneself thinking. One could then describe these feelings to the world,but of course no one would be justified in taking any notice. Likewise accordingto this view the only way to know that a man thinksis to be that particular man. It is in factthe solipsist pointofview. It may be the mostlogicalview to hold but it makes communication of ideas difficult. A is liable to believe 'A thinksbut B does not' whilst B believes' B thinks but A does not '. Insteadofarguing continually over this point it is usual to have the polite conventionthat everyone thinks. Jefferson does not wish to adopt I am sure that Professor and solipsist the extreme pointof view. Probablyhe would be to accept the imitation quite willing game as a test. The game (with the player B omitted) is frequently used in practice someone really underthe name of viva voceto discover whether or has ' learntit parrotfashion'. Let understands something us listenin to a partofsucha vivavoce: Interrogator: In the firstline of your sonnet which reads ' Shall I comparethee to a summer's day ', would not ' a spring day ' do as well or better? Witness: It wouldn'tscan. Interrogator: How about ' a winter'sday' That would scan all right. to a winter's Witness: Yes, but nobodywantsto be compared day. Would you say Mr. Pickwickremindedyou of Interroga.tor: ? Christmas Witness: In a way. is a winter'sday, and I do not Interrogator: Yet Christmas thinkMr. Pickwickwould mindthe comparison. Witness: I don't thinkyou're serious. By a winter'sday one than a special one like means a typicalwinter's day, rather Christmas. Jefferson And so on. What wouldProfessor say ifthe sonnet? I likethis in the vivavoce was able to answer machine writing he would regardthe machineas 'merely do not knowwhether

COMPUTING

MACHINERY

AND

INTELLIGENCE

447

' theseanswers, but if the answerswere as artificially signalling and sustainedas in the above passage I do not satisfactory think he would describeit as 'an easy contrivance'. This intended to coversuchdevicesas theinclusion phraseis, I think, in the machineof a recordof someonereadinga sonnet,with to turnit on from timeto time. appropriate switching In shortthen,I thinkthat most of those who supportthe could be persuadedto abandon it from consciousness argument rather thanbe forced intothe solipsist position. They willthen probably be willing to accept ourtest. I do not wishto give the impression that I thinkthereis no about consciousness. Thereis, forinstance, mystery something of a paradox connected withany attemptto localiseit. But I do notthink thesemysteries need to be solved before necessarily we can answerthe question with which we are concernedin this paper. (5) Argumentsfrom Disabilities. Thesearguments Various take " I grantyou that you can make machines the form, do all the but you will neverbe able to make thingsyou have mentioned one to do X". NumerousfeaturesX are suggestedin this a selection: connexion. I offer Be kind, resourceful, beautiful, friendly (p. 448),have initiative, have a sense of humour, tell rightfromwrong, make mistakes (p. 448), fallin love,enjoystrawberries and cream(p. 448),make some one fall in love with it, learnfrom experience (pp. 456 f.), use words properly, be the subjectof its own thought(p. 449), have as much diversity of behaviouras a man, do something reallynew (p. 450). (Someofthesedisabilities are givenspecial as indicatedby the page numbers.) consideration No supportis usually offered for these statements. I believe founded on the principle theyare mostly of scientific induction. A manhas seenthousands ofmachines inhislifetime. Fromwhat he sees ofthemhe drawsa number ofgeneral conclusions. They are ugly, each is designedfor a very limitedpurpose,when requiredfora minutely different purposethey are useless,the varietyof behaviour of any one of themis verysmall,etc., etc. Naturallyhe concludesthat these are necessaryproperties of machinesin general. Many of these limitations are associated withthe verysmall storagecapacityof most machines. (I am assumingthat the idea of storagecapacityis extendedin some way to cover machines other than discrete-state machines.

448

A. M. TURING:

accuracy as no mathematical doesnotmatter The exactdefinition discussion.) A few years ago, when is claimed in the piresent it was possible verylittlehad been heard of digitalcomputers, them,if one mentioned concerning to elicit much incredulity their propertieswithout describingtheir construction.That due to a similarapplicationof the principle was presumably are induction. These applicationsof the principle of scientific of course largelyunconscious. When a burnt child fears the fireand showsthat he fearsit by avoidingit, I shouldsay that induction. (I could of course also he was applyingscientific describehis behaviourin many otherways.) The worksand customsof mankinddo not seem to be very suitable material induction. A very large part of to which to apply scientific if reliable resultsare to be space-timemust be investigated, we may (as most English childrendo) obtained. Otherwise speaks English, and that it is silly to decide that everybody learnFrench. to be made about many special remarks Thereare, however, to that have been mentioned. The inability of the disabilities and cream may have struckthe reader as enjoy strawberries frivolous. Possibly a machine might be made to enjoy this deliciousdish, but any attemptto make one do so would be is that it conabout this disability idiotic. What is important ofthe e.g.to thedifficulty disabilities, to someoftheother tributes as man and machine between occurring same kindoffriendliness betweenwhiteman and whiteman, or betweenblack man and black man. " seems a The claim that " machinescannot make mistakes " Aretheyany theworse to retort, curiousone. One is tempted and attitude, forthat ? " But let us adopt a moresympathetic can be try to see what is really meant. I thinkthis criticism ofthe imitation gamne. It is claimedthatthe explainedin terms the man simply the machinefrom could distinguish interrogator by settingthem a numberof problemsin arithmetic. The machine would be unmaskedbecause of its deadly accuracy. for The reply to this is simple. The machine (programmed answers playingthe game) would not attemptto give the right problems. It would deliberatelyintroduce to the arithmetic the interrogator.A calculatedto confuse mistakesin a manner an unsuitshow itselfthrough faultwouldprobablymechanical able decision as to what sort of a mistaketo make in tbe of the criticismis not arithmetic. Even this interpretation the space to go sympathetic. But we cannotafford sufficiently depends intoit muchfurther. It seemsto me thatthiscriticism

COMPUTING

MACHINERY

AND INTELLIGENCE

449

betweentwo kinds of mistake. We may call on a confusion ' and ' errors '. Errors ofconclusion offunctioning them' errors are due to some mechanicalor electricalfault of functioning which causes the machine to behave otherwisethan it was designed to do. In philosophical discussions one likes to discussing of such errors; one is therefore ignore the possibility are mathematical 'abstract machines'. Theseabstractmachines they are fictions ratherthan physical objects. By definition of functioning.In this sense we can truly incapable of errors say that 'machinescan nevermake mistakes'. Errorsof conclusion can only arise when some meaningis attached to the output signals from the machine. The machine might,for instance,type out mathematicalequations, or sentences in English. When a false propositionis typed we say that the of conclusion. Thereis clearly an error machinehas committed no reason at all for saying that a machinecannot make this but type out repeatedly kind of mistake. It mightdo nothing ' 0 - 1 '. To take a less perverse example,it mighthave some induction. We method for drawingconclusionsby scientific must expect such a methodto lead occasionallyto erroneous results. The claim that a machinecannot be the subject of its own if it can be shownthat thoughtcan of courseonlybe answered withsomesubjectmatter. Neverthe machine has some thought ' does theless, 'the subject matterof a machine's operations seemto meansomething, at least to the peoplewho deal withit. If, for instance,the machinewas tryingto finda solutionof the equationx2 - 40x - 11 0 one woald be temptedto describethis equationas part of the machine'ssubject matterat can that moment. In this sortof sense a machineundoubtedly be its own subject matter. It may be used to help in making in of alterations the effect or to predict up its own programmes, ofits ownbehaviour theresults its ownstructure. By observing so as to achievesomepurpose its ownprogrammes it can modify These are possibilitiesof the near future, more effectively. ratherthan Utopiandreams. -cannot have much diversity The criticism that a machine of behaviouris just a way of sayingthat it cannothave much storage capacity. Until fairlyrecentlya storage capacity of even a thousanddigitswas veryrare. hereare often The criticisms disguised that we are considering consciousness. Usuallyifone mainfrom forms ofthe argument the and describes can do one ofthesethings, tainsthat a machine kind of methodthat the machinecould use, one will not make 29

450

A. M. TURING:

muchofan impression. It is thought thatthemethod(whatever it may be, for it must be mechanical)is really ratherbase. theparenthesis in Jefferson's Compare statement quotedon p. 21. (6) Lady Lovelace's Objection. Our mostdetailedinformation of Babbage's Analytical -Enginecomesfrom a memoirby Lady Lovelace. In it she states," The AnalyticalEngine has no pretensionsto originate anything. It can do whatever we know how " (her italics). This statement toorderit to perform is quoted by Hartree(p. 70) who adds: " This does not imply that it may not be possible to constructelectronic equipmentwhich will 'thinkforitself ', or in which, in biological terms, one could set up a conditioned reflex, whichwould serve as a basis for 'learning'. Whether this is possiblein principleor not is a stimulating and excitingquestion,suggested by some of these recentdevelopments. But it did not seem that the machines or projectedat the timehad thisproperty constructed ". I am in thorough agreement withHartreeover this. It will be noticedthat he does not assertthat the machines in question had not got the property, but rather that the evidenceavailable to Lady Lovelacedidnotencourage herto believethattheyhad it. It is quite possiblethat the machinesin questionhad in a sense gotthisproperty. For supposethatsomediscrete-state machine has the property. The Analytical Engine was a universal digitalcomputer, so that,if its storagecapacityand speed were adequate, it could by suitableprogramming be made to mimic the machine in question. Probably this argumentdid not occurto the Countessor to Babbage. In anycase therewas no on themto claim all that could be claimed. obligation This wholequestion willbe considered again undertheheading of learning machines. A variantof Lady Lovelace's objectionstatesthat a machine can ' neverdo anything reallynew'. This may be parried fora momentwith the saw, 'There is nothingnew under the sun'. Who can be certainthat 'original work' that he has done was not simplythe growth of the seed plantedin him by teaching, or the effectof following well-known general principles. A bettervariantof the objectionsays that a machinecan never 'take us by surprise'. This statement is a moredirectchallenge and can be met directly. Machinestake me by surprise with greatfrequency. This is largelybecause I do not do sufficient calculation to decidewhatto expectthem to do,orrather because, I do a calculation, I do it in a hurried, although slipshod fashion, ' I suppose the voltage takingrisks. PerhapsI say to myself, hereoughtto be the same as there: anywaylet's assumeit is '.

COMPUTING

MACHINERY

AND INTELLIGENCE

451

forme for is a surprise and theresult wrong, I am often Naturally have been is done theseassumptions by the timethe experiment forgotten.These admissionslay me open to lectureson the subjectof my viciousways,but do not throwany doubt on my I experience. to the surprises whenI testify credibility I do not expectthis replyto silencemy critic. He will promentalact are due to somecreative bablysay that suchsurprises no crediton the machine. This leads us on mypart,and reflect the idea and farfrom consciousness, from back to the argument we must considerclosed, of surprise. It is a line of argument of somethatthe appreciation remarking but it is perhapsworth requiresas much of a ' creativementalact ' thingas surprising froma man, a book, a event originates the surprising, whether else. machineor anything is due, The view that machinescannotgive rise to surprises and mathematicians I believe,to a fallacyto whichphilosophers that as soon as subject. This is the assumption are particularly a fact is presentedto a mind all consequencesof that fact withit. It is a veryusespringinto the mind simultaneously but one too easily ful assumptionunder many circumstances, ofdoingso is that forgets thatit is false. A naturalconsequence one then assumesthat thereis no virtuein the mere working principles. data and general from out of consequences in the Nervous System. The from Contitnuity (7) Argument machine. A nervous systemis certainlynot a discrete-state impulse about the size ofa nervous in theinformation smallerror to the size may make a large difference on a neuron, impinging of the outgoing impulse. It may be arguedthat,this beingso, one cannot expect to be able to mimicthe behaviourof the system. witha discrete-state nervoussystem from machinemustbe different It is truethat a discrete-state of the machine. But ifwe adhereto the conditions a continuous will not be able to take any imitationgame, the interrogator can be made clearer advantageofthisdifference.The situation if we consider some other simpler continuousmachine. A analyser differential analyserwill do very well. (A differential type used is a certainkind of machinenot of the discrete-state for some kinds of calculation.) Some of these provide their answersin a typed form,and so are suitable for taking part in the game. It would not be possible for a digital computer analyser to predict exactly what answers the differential would give to a problem,but it would be quite capable of if asked to give givingthe rightsort of answer. For instance, the value of r (actually about 3.1416) it would be reasonable

452

A. M. TURING:

to choose at randombetweenthe values 3*12,3*13,3*14,3-15, 0.55, 019, 006 (say). of 005, 0O15, 3 16 with the probabilities for the it would be very difficult Under these circumstances analyser fromthe the differential to distinguish interrogator digital computer. of Behaviour. It is not fromInformality (8) The Argument to describewhat possible to producea set of rules purporting a man should do in every conceivableset of circumstances. One mightforinstancehave a rule that one is to stop when light,and to go if one sees a greenone, one sees a red traffic ? One may but what if by some fault both appear together perhaps decide that it is safest to stop. But some further to thisdecisionlater. To attempt may well arisefrom difficulty even those providerulesof conductto cover everyeventuality, lights,appears to be impossible. With all arisingfromtraffic this I agree. From this it is arguedthat we cannot be machines. I shall but I fear I shall hardlydo it the argument, try to reproduce like this. 'If each man justice. It seems to run something his set of rulesof conductby whichhe regulated had a definite life he would be no betterthan a machine. But thereare no such rules, so men cannot be machines.' The undistributed is everput quite middleis glaring. I do notthink the argument used nevertheless. like this, but I believe this is the argument between'rules of There may howeverbe a certainconfusion ' to cloudtheissue. By ' rules conduct' and ' laws ofbehaviour suchas ' Stop ifyou see redlights', of conduct' I meanprecepts on whichone can act, and of whichone can be conscious. By 'laws of behaviour' I mean laws ofnatureas appliedto a man's bodysuchas ' ifyoupinchhimhe willsqueak '. If we substitute whichregulatehis life' for'laws of conduct 'laws of behaviour his life' in the argument quoted the unby whichhe regulates middle is no longer insuperable. For we believe distributed by laws ofbehaviour thatit is not onlytruethat beingregulated a impliesbeing some sort of machine (thoughnot necessarily butthatconversely discrete-state beingsucha machine machine), implies being regulatedby such laws. However,we cannot laws of so easily convinceourselvesof the absence of complete behaviouras of completerules of conduct. The only way we such laws is scientific and we observation, know of forfinding underwhichwe could say, know of no circumstances certainly 'We have searchedenough. Thereare no such laws.' moreforcibly that any such statement We can demonstrate would be unjustified. For supposewe could be sure of;finding

COMPUTING

MACHINERY

AND INTELLIGENCE

453

machine such laws if theyexisted. Then givena discrete-s,,ate sufficent to discover by observation be possible it shouldcertainly a reasonand thiswithin its future behaviour, about it to predict able time,say a thousandyears. But this does not seemto be a small computer the case. I have set up on the Manchester themachine whereby usingonly1000unitsofstorage, programme supplied with one sixteenfigurenumberreplieswith another these withintwo seconds. I would defyanyoneto learn from to be able to predictany about the programme repliessufficient repliesto untriedvalues. Perception. I assume fromExtra-Sensory (9) The Argument perthat the readeris familiarwith the idea of extra-sensory ception,and the meaningof the fouritemsof it,viz.telepathy, and psycho-kinesis.These disturbprecognition clairvoyance, ideas. ing phenomenaseem to deny all our usual scientific the How we should like to discreditthem! Unfortunately It is is overwhelming. at least fortelepathy, evidence, statistical one's ideas so as to fitthesenew facts to rearrange verydifficult in. Onceone has acceptedthemit does not seema verybig step to believein ghostsand bogies. The idea that our bodiesmove with simplyaccordingto the knownlaws of physics,together some othersnot yet discoveredbut somewhatsimilar,would be one of the first to go. one. One can say This argument is to mymindquite a strong in replythat many scientific theoriesseem to remainworkable in practice,in spite of clashingwith E.S.P.; that in fact one about it. This is rather can get alongverynicelyif one forgets is just the kind of and one fearsthat thinking cold comfort, relevant. whereE.S.P. may be especially phenomenon runas follows: based on E.S.P. might A morespecific argument a man who game,usingas witnesses "Let us play the imitation and a digitalcomputer. The is good as a telepathicreceiver, can ask such questionsas ' What suitdoes the card interrogator or clairin my right hand belongto ? ' The man by telepathy voyancegivesthe rightanswer130 timesout of400 cards. The machinecan onlyguess at random,and perhapsgets 104 right, identification."Thereis an makesthe right so the interrogator whichopenshere. Supposethedigitalcompossibility interesting puter containsa random numbergenerator. Then it will be naturalto use thisto decidewhatanswerto give. But thenthe will be subjectto the psycho-kinetic randomnumber generator might powersof the interrogator.Perhaps this psycho-kinesis cause the machineto guess rightmore oftenthan would be calculation,so that the interrogator expectedon a probability

454

A. M. TURING:

mightstill be unable to make the rightidentification.On the without anyquestionbe able to guessright other hand,he might may happen. ing, by clairvoyance. With E.S.P. anything If telepathyis admittedit will be necessaryto tightenour test up. The situationcould be regardedas analogousto that weretalkingto himself which would occur if the interrogator withhis ear to the wall. was listening and one ofthe competitors room' would into a 'telep-thy-proof To put the competitors satisfyall requirements. 7. LearningMachines. thatI have no veryconvincing The readerwillhave anticipated myviews. If I had I of a positivenatureto support arguments should not have taken such pains to point out the fallaciesin contrary views. Such evidenceas I have I shall now give. Let us returnfor a momentto Lady Lovelace's objection, whichstatedthat the machinecan onlydo whatwe tell it to do. One could say that a man can ' inject' an idea intothe machine, and that it will respond to a certain extent and then drop by a hammer. Another struck likea piano string intoquiescence, simile would be an atomic pile of less than criticalsize : an the pile to a neutronentering injected idea is to correspond willcausea certain disturbance from without. Each suchneutron the size of the pile is dies away. If, however, whicheventually causedbysuchan incoming the disturbance sufficiently increased, untilthe whole neutron will verylikelygo on and on increasing for phenomenon pile is destroyed. Is there a corresponding ? There does seem to minds,and is there one for machines of themseem to be be one forthe humanmind. The majority ' sub-critical',i.e. to correspond in this analogyto piles of subto such a mindwill on average criticalsize. An idea presented proportion give rise to less than one idea in reply. A smallish are super-critical.An idea presented to such a mindmay give and tertiary of secondary, rise to a whole 'theory' consisting moreremoteideas. Animalsmindsseem to be very definitely to this analogywe ask, 'Can a machine sub-critical. Adhering ?' be made to be super-critical The 'skin ofan onion' analogyis also helpful. In considering the functions ofthe mindor the brainwe findcertainoperations terms. This we say whichwe can explainin purelymechanical does not correspond to the real mind: it is a sortof skinwhich we muststripoff if we are to findthe real mind. But then in off, and so on. whatremains we find skinto be stripped a further

COMPUTING

MACHINERY

AND INTELLIGENCE

455

in thisway do we ever come to the 'real' mind,or Proceeding do we eventually cometo the skinwhichhas nothing in it ? In the latter case the whole mind is mechanical. (It would not be a discrete-state machinehowever. We have discussedthis.) These last two paragraphsdo not claim to be convincing arguments. They should rather be describedas 'recitations tendingto producebelief'. The onlyreallysatisfactory support that can be givenforthe view expressed at the beginningr of ? 6, will be thatprovided by forthe end ofthe century and thendoingthe experiment waiting described. But what can we say -in the meantime ? What steps should be taken now if the experimentis to be successful ? is mainly theproblem oneofprogramming. As I have explained, willhave to be made too, but it seems Advances' in engineering unlikelythat these will not be adeqaate forthe requirements. Estimatesof the storagecapacity of the brain vary from1010 to 1015binarydigits. I inclineto the lowervalues and believe is used forthe highertypes of that only a very small fraction used forthe retention thinking. Most of it is probably of visual ifmorethan 109was required impressions. I shouldbe surprised oftheimitation forsatisfactory playing game,at anyrateagainst a blind man. (Note-The capacity of the Encyclopaedia is 2 X 109.) A storagecapacityof 107 Britannica,11thedition, even by presenttechwould be a very practicablepossibility niques. It is probablynot necessary to increasethe speed of operationsof the machinesat all. Parts of modernmachines which can be regardedas analogues of nerve cells workaboat a thousandtimesfasterthan the latter. This shouldprovidea ' which could cover losses of speed arising ' marginof sa.fety in manyways. Ourproblem thenis to find outhowto programme thesemachines to playthegame. At mypresent rateofworking I produceabout a thousanddigitsof programme a day, so that about sixty workers, working steadilythroughthe fifty years the job, if nothing wentinto the waste-paper mightaccomplish basket. Some more expeditiousmethod seems desirable. In the processof trying to imitatean adult hiuman mindwe are bound to thinka good deal about the process whichhas broughtit to the state that it is in. We may notice three components, (a) The initialstate ofthe mind,say at birth, to whichit has been subjected, (b) The education not to be describedas education,to (e) Other experience, whichit has been subjected.

456

A. M. TURING:

to simulatethe Instead of tryingto produce a programme tryto produceone whichsimulates adult mind,why not rather the child's? If this were then subjected to an appropriate course of education one would obtain the adult brain. Prelike a note-bookas one is something sumablythe child-brain and lots the stationers. Rather littlemechanism, buys it from our pointof are from and writing of blank sheets. (Mechanism view almost synonymous.) Our hope is that thereis so little likeit can be easily that something mechanism in the child-brain programmed. The amount of work in the education we can to be muchthe same as forthe approximation, assume,as a first human child. We have thus divided our problem into two parts. The and the educationprocess. These two remain child-programme verycloselyconnected. We cannotexpectto finda good childwith teaching attempt. Onemustexperiment thefirst machine-at one such machineand see how well it learns. One can thentry anotherand see if it is betteror worse. There is an obvious by the identifibetweenthis processand evolution, connection cations material Structure ofthe childmachine= Hereditary = Mutations ,, Changes ,, = ofthe experimenter Judgment Natural selection willbe moreexpeditious thatthisprocess Onemayhope,however, is a slow methodfor than evolution. The survivalof the fittest by the exerciseof measuringadvantages. The experimenter, is shouldbe able to speedit up. Equally important intelligence, to randommutations. If he the fact that he is not restricted thinkofthe he can probably can tracea cause forsomeweakness whichwillimprove kindofmutation it. It will not be possible to apply exactly the same teaching processto the machineas to a normalchild. It willnot, for instance,be providedwith legs, so that it could not be asked to go out and fillthe coal scuttle. Possiblyit mightnot have mightbe overcome eyes. But howeverwell these deficiencies to school one could not send the creature by cleverengineering, makingexcessivefunof it. It must without the otherchildren be given some tuition. We need not be too concernedabout the legs, eyes, etc. The example of Miss Helen Keller shows that education can take place provided that communication in both directions betweenteacherand pupil can take place by some means or other.

COMPUTING

AMACHINERY AND INTELLIGENCE

457

We normallyassociate punishments and rewardswith the teaching process. Some simple child-machines can be constructed or programmed on this sortof principle. The machine has to be so constructed that eventswhichshortly preceded the occurrence of a punishment-signal are unlikelyto be repeated, whereas a reward-signal increasedthe probability of repetition of the eventswhichled up to it. These definitions do not presuppose any feelings on the part of the machine. I have done some experiments with one such child-machine, and succeeded in teachingit a few things,but the teachingmethodwas too unorthodox forthe experiment to be considered reallysuccessful. The use of punishments and rewardscan at best be a part of the teachingprocess. Roughlyspeaking,if the teacherhas no other means of communicating to the pupil, the amount of information whichcan reachhimdoes not exceedthe total number of rewardsand punishments applied. By the time a child has learntto repeat 'Casabianca' he would probablyfeel very sore indeed,if the text could only be discovered by a 'Twenty Questions' technique,every 'NO' takingthe formof a blow. It is necessarytherefore to have some other 'unemotional' channelsof communication.If theseare available it is possible to teach a machineby punishments and rewards to obey orders givenin somelanguage,e.g. a symbolic language. These orders are to be transmitted throughthe 'unemotional' channels. The use of this language will diminish greatlythe numberof and rewardsrequired. punishments Opinionsmay vary as to the complexity whichis suitablein the child machine. One mighttry'to make it as simple as possibleconsistently withthe generalprinciples. Alternatively ' builtin '., one might havea complete system oflogicalinference In the lattercase the storewould be largelyoccupiedwithdefinitions and propositions. The propositions wouldhave various kinds of status,e.g. well-established facts,conjectures, mathematicallyprovedtheorems, statements given by an authority, expressions havingthe logicalform ofproposition but not beliefvalue. Certain propositions may be described as 'imperatives'. The machineshould be so constructed that as soon as an im' the appropriate perativeis classed as 'well-established action automatically takesplace. To illustrate this,supposetheteacher says to the machine,'Do your homework now'. This may cause " Teachersays 'Do yourhomework now ' " to be included amongst the well-established facts. Another suchfactmight be,
1Or rather 'programmed in' forourchild-machine willbe programmed in a digitalcomputer. But the logicalsystem willnothave to be learnt.

458

A. M. TURING:

"Everythingthat teachersays is true". Combining thesemay eventuallylead to the imperative, 'Do yourhomework now', being included amongst the well-established facts, and this, by the construction ofthemachine, willmeanthatthehomework actually gets started,but the effect is very satisfactory.The processesof inference used by the machineneed not be such as would satisfy the most exactinglogicians. There mightfor instancebe no hierarchy of types. But thisneed not mean that type fallacieswill occur, any more than we are bound to fall over unfencedcliffs. Suitable imperatives(expressed within the systems, not forming part of the rules of the system)such as ' Do not use a class unless it is a subclass of one whichhas been mentioned by teacher' can have a similar effect to 'Do not go too near the edge'. The imperatives that can be obeyed by a machinethat has no limbsare boundto be of a rather intellectual character, as in the example(doing homework) givenabove. Important amongst such imperatives will be ones whichregulatethe orderin which the rules of the logical systemconcernedare to be applied. For at each stage whenone is usinga logical system, thereis a very large numberof alternativesteps, any of which one is to apply,so faras obedienceto the rulesofthe logical permitted is concerned. Thesechoicesmakethe difference system between a brilliant and a footling reasoner, not the difference betweena soundand a fallacious one. Propositions leadingto imperatives of this kind mightbe " When Socrates is mentioned, use the syllogism in Barbara " or " If one method has been provedto be ". Some quickerthan another,do not use the slowermethod of these may be 'given by authority',but othersmay be produced by the machineitself, e.g. by scientific induction. The idea of a learningmachinemay appear paradoxical to some readers. How can the rules of operationof the machine how the machine change? They should describe completely will react whateverits historymight be, whatever changes it might undergo. The rules are thus quite time-invariant. This is quite true. The explanationof the paradox is that the rules whichget changedin the learning processare of a rather less pretentious kind,claiming onlyan ephemeral validity. The readermay draw a parallelwiththe Constitution of the United States. feature machineis that its teacher Alnimportant of a learning will oftenbe very largelyignorantof quite what is going on inside,althoughhe may still be able to some extentto predict his pupil's behaviour. This should apply most strongly to the

COMPUTING

MACHINERY

AND INTELLIGENCE

459

later educationof a machinearisingfroma child-machine of well-trieddesign (or programme). This is in clear contrast whenusinga machine withnormal procedure to do computations: one's object is thento have a clearmentalpicture ofthe state of in the computation. This object the machineat each moment can onlybe achievedwith a struggle. The viewthat'themachine can only do what we know how to orderit to do ',1 appears whichwe can strangein face of this. Most of the programmes put into the machinewill resultin its doingsomething that we cannot make sense of at all, or whichwe regardas completely random behaviour. Intelligentbehaviourpresumably consists in a departurefromthe completely disciplinedbehaviourinvolved in computation, but a ratherslightone, whichdoes not loops. give rise to randombehaviour,or to pointlessrepetitive ourmachine forits partin Another important resultofpreparing is that the imitation and learning game by a processof teaching ' is likelyto be omittedin a rathernatural 'human fallibility way, i.e. without special'coaching'. (The readershouldreconcile this withthe point of view on pp. 24, 25.) Processesthat are learnt do not producea hundredper cent. certainty of result; iftheydid theycould notbe unlearnt. wise to includea randomelement It is probably in a learning machine(see p. 438). A randomelement is ratherusefulwhen we are searching fora solutionof some problem. Suppose for instancewe wantedto finda number between50 and 200 which was equal to the square of the sum of its digits,we might start at 51 thentry52 and go on untilwe got a number that worked. we might Alternatively choosenumbers at random untilwe got a a good one. Thismethod has theadvantagethatit is unnecessary to keep track of the valaes that have been tried,but the disadvantagethat one may trythe same one twice,but this is not if there are several solutions. The systematic very important methodhas the disadvantagethat theremay be an enormous block withoutany solutionsin the regionwhichhas to be investigated first. Now the learningprocess may be regarded as a searchfora form of behaviour whichwillsatisfy theteacher (or some othercriterion). Since thereis probablya verylarge number of satisfactory solutions the random method seems to be better thanthe systematic. It shouldbe noticedthat it is used in the analogous process of evolution. But there the systematic methodis not possible. How could one keep track
1 Compare Lady Lovelace's statement (p. 450),whichdoes not contain the word'only'.

460

A. M. TURING:

COMPUTING

MACHINERY

AND INTELLIGENCE

of the different geneticalcombinations that had been tried,so as to avoid trying themagain ? We mayhopethat machines will eventually compete withmen in all purelyintellectual fields. But whichare the best ones to start with? Even this is a difficult decision. Many people thinkthat a very abstract activity,like the playing of chess, would be best. It can also be maintainedthat it is best to providethe machinewiththe best sense organsthat moneycan buy, and then teach it to understand and speak English. This process could followthe normalteaching of a child. Things would be pointed out and named, etc. Again I do not know what the rightansweris, but I thinkbothapproachesshouldbe tried. We can onlysee a shortdistanceahead, but we can see plenty therethat needs to be done.
BIBLIOGRAPHY Samuel Butler, Erewhon, London, 1865. Chapters 23, 24, 25, The Book of the Machines. Alonzo Church," An Unsolvable Problem of ElementaryNumber Theory ", American J. of Math., 58 (1936), 345-363. K. G6del, " Vber formalunentscheidbareSatze der Principia Mathematica und verwandter Systeme, I ", Monatsheftefuir Math. und Phys., (1931), 173-189. D. R. Hartree, Calculating Instrumentsand Machines, New York, 1949. S. C. Kleene, " General Recursive Functions of Natural Numbers ", American J. of Math., 57 (1935), 153-173 and 219-244. " The Mind of Mechanical Man ". Lister Oration for 1949. G. Jefferson, BritishMedical Journal, vol. i (1949), 1105-1121. Countess of Lovelace, ' Translator's notes to an article on Babbage's Analytical Engire', Scientific Memoirs (ed. by R. Taylor), vol. 3 (1842), 691-731. Philosophy,London, 1940. Bertrand Russell, History of Western A. M. Turing, " On Computable Numbers, with an Application to the Entscheidungsproblem", Proc. London Math. Soc. (2), 42 (1937), 230-265.

V7ictoria Universitty of Manchester.