Open Journal of Philosophy
2012. Vol.2, No.3, 189-194
Published Online August 2012 in SciRes (
Copyright © 2012 SciRes. 189
Updating the Turing Test. Wittgenstein,
Turing and Symbol Manipulation*
Carlo Penco
University of Genoa, Genoa, Italy
Email: penco@unige .it
Received April 12th, 2012; revised May 14th, 2012; acc epted May 30th, 2012
In this paper I present an argument against the feasibility of the Imitation Game as a test for thinking or lan g u a g e
understanding. The argument is different from the five objections presented by Turing in his original paper, al-
though it tries to maintain his original intention. I therefore call it “the Sixth Argument” or “the Argument from
Context”. I show that—although the argument works against the original version of the imitation game—it may
suggest a new version of the Turing Test, still coherent with the idea of thinking and understanding as symbol
manipulation. In a new form, the main idea which lies behind the original Imitation Game remains untouched by
the criticism of Searle’s Chinese room argument and suggest a possible implementation which avoids some of
the shortcomings of the orig inal Turing Test.
Keywords: Philosophy of Logic; Meaning; Context; Imitation Game; Turing; Wittgenstein
Wittgenstein’s Philosophical Investigations was published
just a few years after his pupil’s famous paper on Mind, “Com-
puting Machinery and Intelligence” (1950). In this paper Turing,
who had attended Wittgenstein’s lectures in 1939, was relying
on a vision of language as an essential feature of intelligence.
He was following the intuition of his teacher, according to
whom “thinking is essentially operating with signs”1. Wittgen-
stein’s main novelty in his book posthumously published in
1953 was a vision of language which partly challenged the
Turing Test. I don’t mean that Wittgenstein was against the
idea that machines can think; he made a few remarks on the
topic, such as the following:
Could a machine think?——Could it be in pain?—Well,
is the human body to be called such a machine? It surely
comes as close as possible t o being such a machine. But a
machine surely cannot think!—Is that an empirical state-
ment? No. We only say of a human being and what is like
one that it thinks. We also say it of dolls and no doubt of
spirits too. Look at the word “to think” as a tool (Wittgen-
stein, 1953: pp. 359-360).
In this quotation we have a very general remark on the use of
the verb “to think”: if something is sufficiently “like” a human
being it is reasonable to attribute to it the property of thinking.
And being sufficiently “like” a human being implies—among
other things—showing correct linguistic behavior. As Harnad
2000 (p. 429) remarks, “‘likeness’ can take two forms: likeness
in structure and likeness in function”; Turing chooses likeness
in function, particularly concerning the function tested by lin-
guistic behavior. Harnad criticizes the limitation of the original
Turing Test using only linguistic behavior, and proposes a hi-
erarchy of Turing tests, where the fundamental one is the 3rd
level of test, where robotics is central; this step is motivated by
the fact that “things that human beings can do go beyond mere
verbalising” and only with robotics we may map these intelli-
gent abilities The idea had already been presented in Harnad
1991 with the TTT (Total Turing Test), which is exactly like
the Turing test, but requires machines to answer all kinds of
input, not just verbal2. My proposal is slightly different, al-
though it shares with Harnad the idea that the meaning of words
cannot develop in isolation from action (“it is hard to imagine
how our words could have the meanings they have if they were
not first grounded in these nonverbal interactions with the
world”)3. The central idea of the present paper is connected
with a better and more adequate definition of what is meant by
“linguistic behavior”. With his conception of “language game”
Wittgenstein insisted that no linguistic expression has any
meaning unless considered inside a context of actions and goals.
Certainly the imitation game is a “language game” with clear
actions and goals, a game to which Wittgenstein could have
given his approval and interest.
In what follows I want to make two opposite remarks on the
Turing Test from the point of view of Wittgenstein’s extreme
contextualism. On the one hand Wittgenstein helps us in clari-
fying the limits of the imitation game with respect to what we
may call the core of our language use, the connection between
linguistic expressions and actions. On the other hand, some of
Wittgenstein’s ideas may respond to the concerns raised by
John Searle’s famous mental experiment of the Chinese room,
and offer a possible alternative to a renewed Turing Test to help
us better understand what we could be meant by saying that
machines could think.
2This idea has provoked the search for stronger and stronger kinds of Tur-
ing Tests, like the proposal by Schweizer, 1998 for a Truly Total Turing
Test, requiring machines able to develop languages and new ideas. Con-
temporary robotics is not so far from this conception. But Turing himsel
suggested putting sense organs in a machine and teaching it to learn a lan-
guage. A completely opposite view is the view of “restricted” Turing Tests,
devoted to single topics or abilities (see for instance Rajaraman, 1997).
3Harnad, 2000: p. 429.
*A previous version of this paper was presented at the IIT (Italian Institute
of Technology); I thank Giulio Sandini and all the participants for their
criticism and their patience in front of a philosophical discussion a bit far
from the direct commitments of the researchers. Thanks also to Marcello
Frixione for comments o n a previous version of th e paper.
1A well known recurrent m otto used in Wittgenstein 1958.
The Sixth Argument
In Loebner Prize, a competition based on the Turing Test, the
prize is given to the computer whose responses are indistin-
guishable from a human’s. After 10 years of competition for the
Loebner Prize, the Turing test has never been completely over-
come, and no programs have had much more success than the
original Eliza by Weizenmbaum4. Although there are still many
defences of the Turing test (e.g. Copeland, 2000; Moor, 2001),
and Turing himself suggested different variations5, according to
Luger, 2005, the main contemporary reactions to the Turing
Test are critical: 1) it deals with purely symbolic problem solv-
ing without any connection with perceptual skills or actions in
real world, given that the dialog is conducted by means of a
keyboard in a separate room; 2) it needlessly constrains ma-
chine intelligence to fit the mould of human intelligence. Partly
following this second attitude, Hayes & Ford, 1995 claim that
the test is a “distraction” from real artificial intelligence re-
search6. French 2000 suggested that the real problem is not how
to pass the Turing Test, but “why we can’t pass it”. Both reac-
tions are reasonable, but the Turing Test still has some appeal,
and I want to explore reaction a) from a particular point of view:
let us keep the Test as dealing with “symbol manipulation”;
even working inside the concept of symbol manipulation we are
forced to challenge the original setting of the test. And, as we
will see, a more difficult test could prove to be a more tractable
When Turing devised the original setting for the test he an-
ticipated many possible objections: 1) the theological objection;
2) the “head in the sand” objection; 3) the mathematical objec-
tion; 4) the argument from consciousness 5) the argument from
various disabilities. If we stick to the original setting and the
original idea that thinking is symbol manipulation, can we be
content with the answer Turing gave to these objections? I
don’t think so, and I think that Wittgenstein’s insistence on the
context dependence of meaning constitutes a sixth objection
which is apparently very difficult to overcome in principle with
the imitation game: I will call it “the Context Argument”.
The Context argument might be stated in a very simple defi-
nition, which the standard Turing test could not overcome:
CA: We cannot have language understanding unless there
is a processing of the context of utterance and the cogni-
tive context in which a sentence is used.
The context of utterance, as defined by Kaplan and Lewis, is
given basically by speaker, time and location of the utterance,
and requires a direct connection with action and perception in
order to interpret the proper use of indexicals and demonstra-
tives (the referential use of “I”, “you”, “now”, “here”, “today”
and so on). No amount of given information about the world,
no universal encyclopedia, like for instance CYC, can replace
this basic ability in language understanding: an enormous knowl-
edge base may connect a huge number of inferences about a
great deal of general and particular information about the world,
but cannot help in understanding “I have to meet her today” or
“I left all the beers in the fridge yesterday”. The cognitive con-
text includes general rules of conversation such as those in-
spired by the work of Grice dealing with conversational impli-
catures, and with presuppositions from the “common ground”
of a conversation as suggested by Stalnaker. Moreover, lan-
guage understanding cannot rely only on semantic networks
defining inferences running over the lexicon of a language:
linguistic competence deals not only with inferential aspects of
the lexicon, but also with referential aspects, the ability to tell a
cat from a dog, or to recognize the person now in front of you
as the same person you met yesterday. Proper language under-
standing requires the ability to use symbols in context, non only
to manage the inferential relations among symbols (which is
just one aspect of linguistic competence, as stressed by Marconi,
1997). Being organized in a closed room and devoted only to
linguistic exchange and reasoning without any kind of interac-
tion with the real world, the standard Turing test seems unable
to face the challenge of the Context Argument.
In the next paragraph I want to exemplify some particular
ways in which the topic of context dependence has been devel-
oped in philosophy and semantics. I claim that these develop-
ments may help us to suggest another form of the Turing test,
which could overcome not only the criticism raised by the
Context Argument, but also the radical challenge posed by
Searle’s Chinese Room argument (Searle, 1980).
The Context Argument: Details
The problem of contextual dependence is not only a problem
for the Turing test, but for semantics in general and for formal
semantics in particular. Standard model theoretical semantics
treats the meaning of a sentence as its truth conditions: under-
standing a sentence is knowing when the thought expressed by
the sentence is true. This is the fundamental step of traditional
philosophy of language, the step according to which we speak
of the “truth-conditional content” of a sentence. I understand
the meaning of “2 + 2 = 4” if I know that the sentence is true
only if 2 and 2 is 4; or, I know the meaning of “the Eiffel
Tower is located in Paris” only if I know to what “Eiffel
Tower”, “Paris”, “in” and “located” refer to, and I know that
the sentence is true only if the Eiffel Tower is located in Paris.
Since the beginning of the analytic tradition in philosophy of
language it has been apparent that there are problems for this
general project. Formal semantics works quite well with ex-
pressions where time and place are clearly expressed and we
know the meanings and referents of all the expressions com-
posing the sentence. Problems arise with indexicals and de-
monstrative expressions, that is with context dependent expres-
sions like “I”, “you”, “he”, “now”, “today”, “this”, “that” and
4The point is that Turing Test could have been c on si dered “successful” even
with Eliza, i f a ver y “naiv e” perso n might believ e Eliza i s a real p erson ; but
this is based on a “trick” and should be rejected (on the Turing test not
being a trick see also Harnad, 2000; on Eliza being too naïve to enter the
Loebner Prize, see Saygin, Cicekli, & Akman, 2000). The Loebner Prize
udges are a bit more sophisticated (even if sometimes ignorant). But the
asic chat b oxes are still made o f tricks, and unab le even to fulfi l the basic
anaphoric links between two sentences. On the debates on the Turing Test
see the collection edited by Moor, 2003, but it is also possible to find some-
thing interesting in documents linked in web pages like :
5Turing himself on the possibility of having an “unrestricted version” of his
test answered in a BBC broadcast: “Oh yes, at least 100 years, I should say”
This shows that also Turing could see a future for his test (See Moor, 2001:
p. 91).
6See also Hayes & Ford, 1997 reviewing a book of interviews on AI, with
opposite attitudes towards Turing Test. Minsky, not interviewed in the book
had an attitude not dissimilar from the one held by Hayes, and even offered
100 dollars to whoever could bring an end to the Loebner Prize. Given that
the Loebner Prize is supposed to stop when somebody succeeds in passing
the Turing Test, Loebner called his prize “Minsky-Loebner Prize”, assum-
ing that, if there is a winner—and the Prize stops—Minsky would enrich
the avera ge prize by 100 dollars.
Copyright © 2012 SciRes.
so on. As Perry 1997 makes evident, these expressions do not
represent a syntactic set (they have different syntactic roles,
from pronouns to adverbs), but they are a semantic set, charac-
terized by their dependence on the context of utterance, defined
as a limited set of parameters: “speaker, time, location”. Ap-
parently if I say:
(1) “On April, 11, 2012 Carlo is trying to finish a paper”
not only does everybody understand the sentence, but it is easy
to get the truth conditions: the sentence is true only if on April,
11, 2012 Carlo is trying to finish a paper. But things become
more problematic when I say:
(2) “Today I am trying to finish a paper”
Every English speaker understands the sentence, and the
sentence uttered at the time and place of the utterance has clear
truth conditions. But if I find sentence (2) written on paper,
there is no way to give a semantic evaluation (to evaluate at
what conditions it is true) because the sentence can be under-
stood only in the context of its utterance. Certainly there is
some kind of truth condition: the sentence is true if the person
speaking on the day to which “today” refers was trying to finish
a paper during that day. But this is only a schema of truth con-
ditions, it is just a set of rules waiting to be applied (Perry
would speak of “reflexive truth conditions”).
Therefore, while (1) has a determined meaning, represents a
set of procedures whose application permit us to check the truth
of the sentence, in (2)—unless we know the value of the pa-
rameters of the context of utterance (time, location and speaker)
—there is no way to evaluate the truth of the matter.
The general rules attached to the indexicals (“today” refers to
the day of the utterance; “I” refers to the speaker of the utter-
ance…) are procedures which need to be activated and filled
with the appropriate contents in order to evaluate the sentence.
But we need knowledge of the context to fill the gap. Probably
the traditional Turing test would have no problem with the term
“I” which is interpreted as “the individual writing on the screen
at the moment”. But problems arise when we deal with expres-
sions like “here”, “there”, “he”, “she” or “this” and “that”
(when used demonstratively and not anaphorically). The stan-
dard treatment of this kind of context dependent expressions—
indexicals and demonstratives—has been given by David Kap-
lan. Kaplan 1989 distinguishes what he calls “character” from
what he calls “content”. The character is the linguistic rule
attached to the indexical, and the content is the referent of the
ind exical. In the last couple of decades there has been much dis-
cussion on the logical and semantical treatment of “indexicals”.
Different classifications of indexi cals have been given, relying on
the presence/absence of intentions or gestures (according to
which only “now” and “I” are “pure indexicals”, given that the
other expressions seem to have a common, at least implicit, de-
monstrative or intentional aspect). Philosophers and logicians
have devoted much time and ef fort to find a suitable way to treat
such context dependent expressions semantically.
Now a problem arises: are these the only expressions whose
referents (and meanings) depend on the context? Some authors
(e.g. Cappelen & Lepore, 2004) assert that indexicals are the
only strictly context dependent expressions, and all other ex-
pressions contribute a “minimal content” to the thought ex-
pressed by an utterance. For instance to give a semantic evalua-
tion of “he is tall” we just need to apply the rule for “he” in the
context of utterance and we assert that the property of being tall
applies to the person referred to as “he”. Indexicals are the “ba-
sic set” of context dependent expressions7. But not everybody
agrees. Significant concerns have been expressed regarding
scalar adjectives, definite descriptions and shortened expres-
sions used in local settings. Let us consider these three cases.
In a very Wittgensteinian vein, many authors (e.g. Recanati,
2007, 2010) claim that the context of utterance intended in a
broad sense (including also presuppositions and beliefs, that is
including cognitive context) is constitutive of the meaning and
reference of many kinds of expressions. Take “tall”: I can say,
referring to the same person on different occasions,
(3) “John is tall”
because John is taller than the average height of his friends at
school, and at the same time I may say
(4) “John is not tall”
when asked about the possibility for him to become a basketball
player. If the meaning or the truth conditions of “John is tall”
do not depend on the context, then I would reach a contradic-
tion. The context of utterance in a narrow sense (time, speaker
and location) is not enough to solve the problem; to solve the
problem we need a wider conception of context, inclusive not
only of speaker, location and time, but also of shared presuppo-
sitions, implicatures and other kinds of assumptions.
Another much discussed case is that of referential uses of
definite descriptions like “the x that has the property F”. We
often use descriptions loosely because we want to refer to some
salient element of the scene, and the context helps us to pick the
right referent; think of
(5) “The book is on the table”
Which book? Which table? This is a case of incomplete de-
scription, where the context fills the gaps and helps to under-
stand what I am speaking about. Things get worse in case of
misdescriptions or inaccurate descriptions: at a party I may say
(6) “The man drinking a martini is a philosopher”
Let us assume that the man is not drinking a martini, but
sparkling water. Strictly speaking I am referring to the only
person in the room drinking a martini, therefore I say some-
thing false and by implicature I make myself understood (this is
briefly the classical stance held by Kripke 1975, and generally
shared by many contemporary authors). But certainly most of
you will understand whom I am referring to, without making a
step from the literal falsity to the implicature (see for instance
7A further case is the case of quantifiers; we use quantifiers in everyday
language with implicit rest riction of the domain of interpretation.
1) Every bottle of beer is in the frid ge
Apparently we do not mean that every bottle of beer in the world is in the
fridge; no fridge could be so big. We are just referring to a specific domain,
maybe the bottles we have bought for the party. But also something which
seems even simpler can be ambiguo us.
2) Everyone is at home
An utterance of 2) may have different interpretations: “everyone belong-
ing to a defined set of individuals is at the home of this set” or “everyone
belonging to a defined set of individuals is at the home of N (where N is the
relevant p er so n in the con t ext)”. In cas es l i k e t h ese we need t o fi nd a way t o
formalize something which seems implicit, and in a logical language repre-
sentation we need to postulate bound variables in the structure of quantified
hrases, whose values, relative to a context, generate a domain o
quantification (see Stanley,2005). However, restriction to a domain o
quantification is something which could be treated formally with less diffi-
culties than the other ca ses under discussion.
Copyright © 2012 SciRes. 191
Penco, 2010; Korta & Perry, 2011).
A third kind of examples concerns what is normally called
“synecdoche” (the part for the whole). Take for instance:
(7) “The ham roll ran away without paying”
said by the waiter to the owner of the bar: it is apparent that the
waiter is referring to the customer who ordered a ham roll, but
literally speaking the matter is different. How can we give a
treatment of the truth conditions of these and other kinds of
examples without referring to the context? And which specific
rules have to be applied in order to get the right truth conditions?
In the case of indexicals we have standard, systematic rules for
applying the parameters and get the content; is it possible to
find some standard rules for other kinds of context dependent
expressions in order to get their referent and build a correct
semantic representation of the sentence in context?
Summing up, we have to accept that in basic cases of lin-
guistic interaction, in order to understand the meaning of what
is said, we need perceptual awareness of the context of utter-
ance, together with the shared presuppositions of cognitive
context, and language alone is not enough. It is the basic prob-
lem raised many years ago by Carnap, who said that “pragmat-
ics is the basis for all linguistics”. This does not mean that syn-
tax and semantics have to deal with proper language use in
action, but means that pragmatics is what fills the content of
semantical and syntactical features. The inferential working of
the lexicon begins after pragmatic disambiguation.
I have given some space to this cursory presentation of what
has been the average discussion on context dependence in the
last twenty years in order to provide some examples of the
challenges of the Context Argument. As we have said before,
we need to take account not only of the “narrow context”, that
is the context of utterance, as for (1) and (2), but also of the
“broad context”, that is the intended domain of interpretation, the
intention of the speaker, the setting of the sce ne, the presupposi-
tions which emerge from the discourse, and many other features
of the situation in which a sentence is uttered. Is anything like the
imitation game a good guide in dealiong with these problems?
Not really. In fact contemporary natural language understanding
systems work fairly well for si mple automatic translation, but do
not work properly in understanding language-in-context. Natural
language understanding systems (like humans) need the capabil-
ity of understanding elements of context corresponding to in-
dexicals, demonstratives and othe r “context uals”, like quantifier s,
definite descriptions, adjectives and local abbreviations. These
abilities are the ground which permits semantics (the truth eva-
luation of sentences) to work, and—as far as natural language
processing is concerned—formalizing these abilities is one of
the main challenges of our century.
Updating the Turing Test
The attempts to overcome the limitations of standard formal
semantics are promising, especially when connected with re-
search programs in computational linguistics and artificial in-
telligence (the multi context theory developed by John M c C arthy
has been a first fundamental step in this direction)8. If, on the
one hand, they constitute a challenge to the original Turing Test,
they might also offer some hope of a new reconstruction of the
imitation game, as inspired by the fundamental tenets shared by
Turing and Wittgenstein: thinking or understanding is the capa-
bility of using signs or, in other words, “Language understand-
ing is symbol manipulation”.
Unfortunately the depth of this idea has been obscured by the
interpretation of the test given by John Searle in his mental
experiment of the Chinese Room. Searle’s challenge to artifi-
cial intelligence was exactly a critique of the concept of “sym-
bol manipulation”, considered literally as working with sym-
bols detached from any real interaction with the environment.
In Searle’s mental experiment an English speaker has some
instructions in English to take Chinese symbols as input and is
to give some other symbols as output; the rules in English per-
mit the English speaker to produce as output reasonable an-
swers to the questions in Chinese. A Chinese speaker therefore
would understand the answers to her questions produced by this
procedure, thinking that whoever is inside the room under-
stands Chinese. It is apparently a rhetorical presentation of the
Turing test with the aim of depriving the test of its significance.
In fact, Searle asks whether we can say that the man in the
room understands Chinese. Certainly not! He understands Eng-
lish, and is able to use rules (formal or not) to give as output
some patterns of Chinese symbols as answers to other patterns
of Chinese symbols taken as input, without having any idea of
what those symbols may mean. Symbol manipulation is not
understanding language! What is missing is the understanding
on the meaning of the Chinese symbols and the intentionality,
that is the ability to understand what a symbol refers to. The
English speaker inside the room has no idea what the Chinese
symbols refer to; he only knows how to manipulate symbols, he
is only using a syntactic ability without semantics.
However, the Chinese room is based on the traditional set-
ting of the Turing test: somebody writes a sentence in Chinese
and the Chinese room answers. One of the first reactions to the
mental experiment was that what answers is not the man inside
the room, who apparently knows only English, but the entire
system: what is endowed with an understanding of Chinese is
not the man inside the room—the linguistic program—but the
entire system composed of the man, the room and the Chinese
symbols. Another answer is to give the Chinese room some
possibility of movement and perception: the entire system be-
comes a robot, where the man inside with his rules is just a
small part, the syntactic manipulation; if the output symbols are
correct answers to the questions, this means that the system can
interact coherently with the environment. Searle’s argument
seems insufficiently robust to answer the argument of the sys-
tem and that of the robot combined9.
Thinking as symbol manipulations is not intended to be re-
stricted to actions “on” symbols alone (as was supposed in the
9These are two of the main objections to Searle’s argument. They may be
found in the original collection in Brain and Behavioral Sciences; we have
also a ni ce summary o f the d ifferen t reaction s in th e Stanford Encyclopedia
(Cole, 200 9).
10The argument is as follows: let us not fall prey to the fallacy according to
which in a symbolic language we can leave the world of symbols for a non
symbolic world which could give them meaning. If we have a simulation
we are inthe symbolic world. The problem of the conception of perceptual
primitives is a problem of robotics, which does not contribute in any man-
ner to what is intended as “meaning”, intended as a procedure for manipu-
lating symbols relative to a fixed domain. Providing sensors to a “thinking
machine” do es no t enr ich the pr oced ural as pect of meanin g . Worki ng in sid e
a simulated world is separated from working in robotics, and projects in
robotics cannot properly give more information about meaning than already
givenin natural language processing. Unless… meaning cannot be given a
urely inferential definition and procedures have to deal with real world
8See for instance McCarthy, 1990; Guha & McCarthy, 2003; Penco&
Vignolo, 2 005.
Copyright © 2012 SciRes.
eighties; e.g. Wilks, 1986)10, but to actions “with” symbols:
symbol manipulation is the ability to use symbols in context,
like using and understanding indexicals, demonstratives and
definite descriptions, shortened or not, on the ground not only
of shared presuppositions, but also of perceptual abilities of
recognizing individual objects and patterns. On the contrary, in
the original setting of the imitation game, the human is con-
ceived as an inference machine, and a program simulating a
human is thought to have a similar mastery of inferences, such
that a dialogue is possible though a computer screen and a
keyboard. This setting of the Imitation Game permits mental
experiments such as Searle’s where a system is supposed to
receive sentences as input and give sentences as output. But if
we want to implement the more basic features of language un-
derstanding (such as demonstratives and referential uses of
definite descriptions, quantifiers…) we need to rely on a dia-
logue with a shared environment. We need therefore to have a
form of interaction with real situation, mastering the use of
symbols to detect items in the environment. The challenge is to
give machine the ability to use perceptual information from the
context, and to mix it with background knowledge in order to
use the most difficult aspects of context-dependent language
use presented in the previous paragraph. The formal work done
on these new aspects of the boundary between semantics and
pragmatics is some of the most promising as regards this possi-
bility: updating the Turing test in real situations.
Actually, unnoticed by many, the first example of a possible
query of the interrogator to the unknown (man, woman or
computer) interlocutor in the original paper by Turing (1950) is:
“C: Will X please tell me the length of his or her hair?”
The question is typically considered as an application of
“his” or “her” anaphorically towards a previously restricted
domain (the supposed interlocutors). In this case the role of
indexicals is anaphoric, that is it picks the individual referred to
in the conversation (in this case X). But what will happen if the
question considers a referential use of “his” or “her”? In this
case, we need the ability to detect in the environment some-
thing which might match the use of the indexical adjective,
probably referring to a human salient in the context. We have
here the beginning of a different kind of imitation game: we
might for instance test a group of experts in front of a robot,
whose behavior could be either autonomous or directed by a
human at distance. The challenge would be to detect whether
the robot is an autonomous one or something human directed.
The ability to manipulate symbols here is the ability to interact
with symbols in a common environment. If the autonomous
robot is able to interact correctly, why not accept that, being
sufficiently “like” a human being in manipulating symbols in
context, it thinks?
In a very subtle analysis of the limitation of the Turing Test
and of other possible challenges to artificial intelligence, Cohen
2006 praises other kinds of tests, such as robot soccer competi-
tion, because of their feasibility and capacity of development in
more and more complex stages: “Turing’s test requires simul-
taneous achievement of many cognitive functions and doesn’t
offer partial credit to subsets of these functions. In contrast,
robot soccer presents a graduated series of challenges: it gets
harder each year but is never out of reach”. However a Turing
Test based on basic linguistic abilities in an actual situation
might ask for very simple symbol manipulation in understand-
ing basic actions like “pass me that ball” or “take the red can
you see near him”. Before arriving at a highly complex dia-
logue on Shakespeare (as happened at the Loebner Price where
judges were a bit more ignorant than machines on the specific
subject matter), a challenge on basic linguistic abilities in con-
text might be more easily transformed into an updated Turing
Summarizing: Searle, against Turing, suggested the idea that
understanding is not symbol manipulation. However, properly
understood, a new Turing test might be grounded on the idea
that thinking and language understanding is symbol manipula-
tion in context. Will we be able to invent an imitation game
which could constitute a new challenge for the present century?
An updated Imitation Game would consist of an interrogator
trying to understand whether she is interacting with a robot or
with a human in changing, real world situations. Given that so
many humans have stereotypical behaviors, it would be easy to
raise the doubt that one was meeting some kind of automatic
agent, and the test would be a true challenge for humans and
Cappelen, H., & Lepore, E. (2004). Insensitive semantics. A defence of
semantic minimalism and speech act p lu r a l i s m. Oxford: Blackwell.
Cohen, P. (2006). If not Turing’s test, then what? AI Magazine, 26,
Cole, D. (2009). The Chinese room argument. The Stanford Encyclope-
dia of Philosophy (Winter 2009 Ed it io n).
Copeland, B. J. (2001). The Turing test. Mind and Machines, 10, 519-
539. doi:10.1023/A:1011285919106
French, R. (2000). The Turing test: The frst fifty years. Trends in Cog-
nitive Sciences, 4, 115-121. doi:10.1016/S1364-6613(00)01453-4
Harnad, S. (1991). Other bodies, other minds: A machine incarnation of
an old philosophical problem. Mind and Machines, 1, 43-54.
Harnad, S. (2000). Mind, machines and Turing. The indistinguishability
of the indistinguishable. Mind and Mach ine s, 9, 425-445.
Hayes, P., & Ford, K. (1995). Turing test considered harmful. Pro-
ceedings of the 14th International Joint Conference on Artificial In-
telligence. San Francisco: Morgam Kaufman Publishers.
Hayes, P., & Ford, K. (1997). Talking heads—A review of speaking
minds: Interviews with twenty eminent cognitive scientists. AI Ma-
gazine, 18, 123-125.
Korta, K., & Perry, J. (2011). Critical pragmatics. Cambridge: Cam-
bridge University Press.
Guha, R., & McCarthy, J. (2003). Varieties of contexts. Lecture Notes
in Artificial Intelligence, 2116 , 290-303.
McCarthy, J. (1990). Formalizing commonsense. New York: Ablex.
Moor, J. H. (2001). The status and future of the Turing test. Mind and
Machines, 11, 73-93. doi:10.1023/A:1011218925467
Moor, J. H. (2003) The Turing test. The elusive standard of artificial
intelligence. Dordrecht: Kluwer.
Luger, G. F. (2005). Artificial intelligence: Structures and st r ategies for
complex problem solving (5th ed.) . Boston: Addison-Wesley.
Penco, C. (2010). Essentially incomplete descriptions. European Jour-
nal for Analytic Philosophy, 6, 47-66.
Penco, C., & Vignolo, M. (2005). Converging towards what? Pragmatic
and semantic competence. In P. Bouquet, & L. Serafini (Eds.), Con-
text representation and reasoning ( Vol. 136). CEUR-WS.
Perry, J. (1997). Indexicals and demonstratives. In R. Hale, & C.
Wright (Eds.), Companion to the philosophy of language (pp. 586-
612). Oxford: Blackwell.
Rajaraman, V. ( 1997). Turing test and after. Resonance, 2, 50-59.
Recanati, F. (2007). Perspectival thought. Oxford: Oxford University
Press. doi:10.1093/acprof:oso/9780199230532.001.0001
Recanati, F. (2010). Truth conditional pragmatics. Oxford: Clarendon
Copyright © 2012 SciRes. 193
Copyright © 2012 SciRes.
Press. doi:10.1093/acprof:oso/9780199226993.001.0001
Saygin, A. P., Cicekli, I., & Akman, V. (2000). Turing test: 50 years
later. Mind and Machines, 10, 463-518.
Searle, J. (1980). Minds, brains and programs. Behavioral and Brain
Sciences, 3, 417-457. doi:10.1017/S0140525X00005756
Schweizer, P. (1998). The truly total Turing test. Mind and Machines, 8,
263-272. doi:10.1023/A:1008229619541
Stanley, J. (2005). Language in context. Oxford: Oxford University
Turing, A. M. (1950). Computing machinery and intelligence. Mind,
Turing, A. M. (1951). Can digital computers think? BBC 3rd pro-
gramme. In B. J. Copeland (Ed.), The essential Turing. Oxford: Ox-
ford University Press, 2004.
Wilks, Y. (1986). Default reasoning and self-knowledge. Proceedings
of the IEEE, 74, 1399-1404. doi:10.1109/PROC.1986.13640
Wittgenstein, L. (1953). The philosophical investigations. Oxford:
Wittgenstein, L. (1958). The blue and brown books. Oxford: Blackwell.