COMMENT ⊗   VALID 00005 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	%searle[f89,jmc]		Letter to Scientific American about Searle article
C00011 00003
C00018 00004	\smallskip\centerline{Copyright \copyright\ 1989\ by John McCarthy}
C00026 00005
C00034 ENDMK
C⊗;
%searle[f89,jmc]		Letter to Scientific American about Searle article
\input memo.tex[let,jmc]


	Perhaps it would be better for a computer scientist to
stay out of a ``debate'' among philosophers.  However, in his
``Is the Brain's Mind a Computer Program'' [SCIENTIFIC AMERICAN,
January, 1990], John Searle has inadvertently made a statement
capable of being empirically tested.  Moreover, it is directly
relevant to the main contention of his article.

	Let's assume that the Chinese room rules are capable
of passing a Chinese Turing test, i.e. are not just trivial
responses based on the presence of key words in the text.

	I favor the system reply and like Searle's
modified Chinese room in which the man has memorized the rules.
The man would then be doing what computers do all the time---
running an interpreter program that is interpreting another
program.  If humans ordinarily carried out several independent
mental processes, maybe we couldn't say Searle wrote that paper;
we'd have to say Searle3 wrote the paper, but Searle1 and Searle2
didn't agree with it and almost succeeded in preventing Searle3
from making the Searle body put it in the mail box.  Likewise,
let Man1 be the usual personality and Man2 be the personality of
the process that Man1 is carrying out by executing the rules.
Man2 knows Chinese and Man1 doesn't.

	Here's where Searle has made a commitment about the real
world.

	His axiom 3 is:

	{\it ``Syntax by itself is neither constitutive of
nor sufficient for semantics.''}

	The statement is vague, because
what he means by ``semantics'' and what he means by ``sufficient
for'' is unclear.  However, he illustrates what he means later on
by the following paragraph.

{\it ``Fine.  But now imagine that as I am sitting in the Chinese room
shuffling the Chinese symbols, I get bored with shuffling the---to
me---meaningless symbols.  So, suppose that I decide to interpret
the symbols as standing for moves in a chess game.  Which semantics
is the system giving off now?  Is it giving off a Chinese semantics
or a chess semantics, or both simultaneously?  Suppose there is a
third person looking in through the window, and she decides that
the symbol manipulations can all be interpreted as stock-market
predictions.  And so on.  There is no limit to the number of
interpretations that can be assigned to the symbols because,
to repeat, the symbols are purely formal.  They have no
intrinsic semantics.''}

	This claims that a Chinese text reporting a
conversation could be a report of a chess game or a stock-market
prediction in some other language.  Experience with cryptography,
the statistical theory of communication and the theory of
algorithmic complexity show that this is untrue for texts of
substantial length.

	Cryptography first.  If Searle were right, cryptograms
would often have ambiguous solutions.  In fact simple substitution
ciphers in English generally have unique solutions if they are
longer than 20 letters.  Chinese would have a longer {\it unicity
distance} but surely not longer than a few hundred letters.

	Here's approximately what happens.  Consider a Chinese
sentence involving the character for dog.  Suppose someone
proposes to interpret the sentence as a stock market prediction
with that character meaning IBM.  In order that the sentence make
sense, he has to come up with interpretations for the other
characters.  Suppose he succeeds.  The next time the character
for dog appears, it again has to mean IBM, but the
interpretations of the other characters are also partially
constrained by the first sentence.  Very soon the alternative
interpretation breaks down.

	Statistical communication theory shows how to measure the
information ccntent of natural languages.  For written English,
it's about one bit per letter.  Judging from the length of signs,
it's a little less for most other European languages.
I don't know about Chinese, but signs in Chinese are slightly
shorter than English signs, but the amount of detail per square
inch is larger.  

	The redundancy of natural languages seems to be necessary
in order to permit partial texts to be understood as they are read
or heard.

	Algorithmic complexity theory (treated by Gregory Chaitin
in a Scientific American article) tells us that it is indeed possible
to have a language in which the Chinese conversation has the interpretation
of a stock-market prediction.  However, it will be a very complicated
language, and its description will almost certainly be as long as
the text to be interpreted.  As the Chinese conversation goes on,
the length of the rules interpreting it as a stock-market prediction
will grow.  The resulting languages will be too complex to be learned
by people, let alone two-year old children.

	This isn't just an accidental phenomenon associated with
Searle's fantasy, but is an essential part of the relation between
syntax and semantics.  The fact that the syntax of natural languages
do not generate texts with many interpretations is part of what allows
a child to learn its native language.  If Searle were right, what a
child hears and the gestures it sees would admit equally good
alternative interpretations.

	Searle's article lends itself to a concrete challenge.
Take a page of Chinese conversation from an existing novel.
Devise a pseudo-Chinese language in which the page is interpreted
as a stock market prediction.

Sincerely,
\vfill\end

	It seems to me that Searle has claimed that people not
knowing Chinese cannot distinguish Chinese texts representing
conversations, chess games, and stock market predictions.
Here are are three texts taken from the December xx, 1989
issue of xx, a newspaper published in Taiwan.  Scientific
SCIENTIFIC AMERICAN readers are challenged to determine
which is a conversation taken from a story, which is a
transcription of a chess game (bridge?) and which is
a section of stock market quotations.

TEXTS

  Cryptographers,
given a half hour to think about the problem before seeing
the texts, get the correct classification in 30 seconds.
Stanford graduate students sometimes take an hour.  The
texts have been made more difficult by putting numbers
in Chinese.

	If we allow substitutions for latin letters, English
texts become unique at about 20 letters.  Simple substitution
cryptograms longer than that have essentially unique solutions.

This ``unicity distance'' is considerably longer for Chinese, but
I challenge Searle, with the aid of Chinese,
to make substitutions for Chinese characters in a page of
a Chinese novel and get another comprehensible Chinese text
that differs other than by changing a few words that aren't
much repeated.

	Returning to the Chinese room, suppose a scientist
can observe the brain states of the man carrying out the
rules or, more plausibly, can observe the memory of a computer
program carrying out the task.  Suppose he proposes correspondences
between Chinese characters and memory structures.  Suppose he
interprets other memory structures as representing statements
about the world, about the state of the Chinese conversation and
about conversational goals.
Finally, suppose that most of the behavior can accounted for by
a principle of rationality: {\it The system does what it believes
will achieve its goals.''

	The scientist has then ascribed knowledge of Chinese
to the system.  He has also ascribed knowledge of the world
distinct from the knowledge of Man1.  I claim that it is
very improbable, about like the probability of all the
air molecules in the room collecting at one end, for there
to be a substantially different ascription which also accounts
for the behavior.  This cannot be a current challenge to Searle,
because the present state of AI cannot produce systems that
will carry out a sensible general conversation.  Much will have
to be done before this is possible.  For example, the system
will have to ``introspect'', i.e. do something to which we would ascribe the
term introspection.   It will have to ``observe'' its own
``mental'' state and generate sentences about it.

	Assuming that there is an essentially unique ascription
of beliefs and goals to the system, shall we say that the system
``really has'' these beliefs and goals.  That's optional.  However,
we have no better basis for ascribing mental states to each other.

	Searle isn't the first philosopher to get empirical
conclusions from purely analytic considerations.  For example,
Immanuel Kant said it was impossible that we would ever know
the chemical constitution of the sun only 40 years before
the discovery of spectroscopy.  He also said that Euclidean
geometry was the way the human mind necessarily regarded the
world.  It was 100 years before general relativity required
non-Euclidean geometry.  Kant was safely dead before these
problems arose.  Things move faster these days, and we wish
Searle good luck in backing out of his inadvertent venture
into cryptography.

	As for the other side of the ``debate'', the Churchlands
make empirical statements about the lack of progress of
``conventional AI''.  These statements are mistaken; they just
pay no attention to the scientific literature of AI but
only to other philosophical speculations.
\smallskip\centerline{Copyright \copyright\ 1989\ by John McCarthy}
\smallskip\noindent{\fiverm This draft of SEARLE[F89,JMC]\ TEXed on \jmcdate\ at \theTime}
%File originated on 25-Dec-89
\vfill\eject\end

	Perhaps it would be better for a computer scientist to stay
out of a ``debate'' among philosophers.
However, in his ``Is the Brain's Mind a Computer Program''
[SCIENTIFIC AMERICAN, January, 1990], John Searle
has inadvertently made a statement capable of being empirically
tested.  Moreover, it is directly relevant to the main contention
of his article.

His Axiom 3 states, ``Syntax by itself is neither constitutive of
nor sufficient for semantics.''  This statement is vague, because
what he means by ``semantics'' and what he means by ``sufficient
for'' is unclear.  However, he illustrates what he means later on
by the following paragraph.

``Fine.  But now imagine that as I am sitting in the Chinese room
shuffling the Chinese symbols, I get bored with shuffling the---to
me---meaningless symbols.  So, suppose that I decide to interpret
the symbols as standing for moves in a chess game.  Which semantics
is the system giving off now?  Is it giving off a Chinese semantics
or a chess semantics, or both simultaneously?  Suppose there is a
third person looking in through the window, and she decides that
the symbol manipulations can all be interpreted as stock-market
predictions.  And so on.  There is no limit to the number of
interpretations that can be assigned to the symbols because,
to repeat, the symbols are purely formal.  They have no
intrinsic semantics.''

	It seems to me that Searle has claimed that people not
knowing Chinese cannot distinguish Chinese texts representing
conversations, chess games, and stock market predictions.
Here are are three texts taken from the December xx, 1989
issue of xx, a newspaper published in Taiwan.  Scientific
SCIENTIFIC AMERICAN readers are challenged to determine
which is a conversation taken from a story, which is a
transcription of a chess game (bridge?) and which is
a section of stock market quotations.  Cryptographers,
given a half hour to think about the problem before seeing
the texts, get the correct classification in 30 seconds.
Stanford graduate students sometimes take an hour.  The
texts have been made more difficult by putting numbers
in Chinese.

	Texts.

	Searle has made the problem unusually easy by
choosing text domains with such obvious formats.  These
formats are appropriate to the subject matter, i.e. to the
semantics, independent of language.  Consider a somewhat
more difficult problem.  Given a novel in Chinese, pick
out the dialogs.  I'll bet the alternations, the short
sentences, the pattern of repetitions of the characters
that turn out to represent names and other clues that
haven't occurred to me would make this a feasible task.

	My guess is that given a whole novel in Chinese
and no other information, linguist could puzzle out most
of the words in a year or two.  The amount of information
in a novel is vastly larger than in the tomb inscriptions
and inventory lists available for figuring out ancient
languages.

	However, the relation between syntax and semantics is
given better by another task.  Suppose Searle is given a chapter
from a Chinese novel and a Chinese-English dictionary.  Could he
modify the dictionary so the chapter would have a quite different
interpretation and still make sense?  Probably he could find some
words occurring only a few times whose occurrences would still
make sense with changed meanings.

	In English language simple substitution ciphers, the
unicity distance is about 20 letters.  A cryptogram longer than
20 letters is unlikely to have more than one solution.  Chinese
would have a somewhat longer unicity distance.  Do Chinese
do simple substitution ciphers using Chinese characters?

	Looked at more generally, here's Searle's mistake.
Suppose we have a robot and a human which interact with
the world and whose internal states we
can observe.  Suppose they perform similar tasks.
Suppose someone proposes a correspondence between the
certain aspects of the internal states and beliefs and
goals of the human and robot.  Suppose he accounts for
much of the behavior using these correspondences and a principle
of rationality: {\it It does what it thinks will achieve
its goals}.  What are the chances that someone else can
find a substantially different ascription of beliefs and
goals that will also account for the behavior.  I claim
that the experience with cryptography and with scientific
induction show that the chances are very small---if the
amount and variety of behavior is substantial---about
like the chance of all the air molecules collecting at
one end of the room.

	Now suppose we have a computer program, or Searle's
man, conducting a dialog in Chinese.  I claim there is
likely to be an essentially unique ascription of Chinese
characters, words and idioms to concepts and to aspects
of the state of the system.

	Shall we then say that the system knows Chinese?
I claim that we have no better evidence that Searle knows
English.