3

Mirlc:J, .. .
John Haugeland
I Cognitive Science
"Reasoning is but reckoning," said Hobbes (1651,
ch. V), in the earliest expression of the computa-
tional view of thought. Three centuries later, with
the development of electronic "computers," his
idea finally began to catch on; and now, in three
decades, it has become the single most important
theoretical hypothesis in psychology (and several
allied disciplines), and also the basis of an exciting
new research field, called "artificial intelligence."
Recently, the expression cognitive science has been
introduced to cover all these varied enterprises, in
recognition of their common conceptual founda-
tion. This term, therefore, does not apply to every
scientific theory of cognition, but only to those
sharing a certain broad outlook - which is some-
times called the "information processing" or
"symbol manipulation" approach. Perhaps, at
last, Hobbes's philosophical insight has found its
home in a proper scientific paradigm (Kuhn,
1970). [ ... ]
Often, discussion of cognitive science focuses on
artificial intelligence - "AI," among friends -
because it amounts to a kind of distilled essence
of cognitive science. But again, it is important to
realize that" AI" (like "cognitive science") is more
specific in its meaning than the words themselves
might suggest. Crudely, we can put the point in
terms of different technologies: a project at IBM to
Haugeland, J., "Semantic engines: An introduction to
mind design;' from J. Haugeland (ed.), Mind Design
(MIT Press, Cambridge, MA,1981).
wire and program an intelligent robot would prob-
ably be AI, whereas a project at DuPont to brew
and mold a synthetic-organic android probably
would not. But this can be misleading; the crucial
issue is not protoplasm versus semiconductor
("wetware" versus "hardware"), but rather
whether the product is designed and specified in
terms of a computational structure. If it is, then a
working model could probably be manufactured
much more easily by means of electronics and
programming; and that's the only relevance of the
technology. Indeed, the guiding inspiration of
cognitive science is that, at a suitable level of
abstraction, a theory of "natural" intelligence
should have the same basic form as the theories
that explain sophisticated computer systems. It is
this idea which makes artificial intelligence seem
not only possible, but also a central and pure form
of psychological research.
A better perspective on all the excitement can be
gained by asking why it took three hundred years
for Hobbes's original proposal to be appreciated.
Mainly, three famous philosophical dilemmas
stood in the way: (i) the metaphysical problem of
mind interacting with matter; (ii) the theoretical
problem of explaining the relevance of meanings,
without appealing to a question-begging homun-
culus; and (iii) the methodological issue over the
empirical testability (and, hence, respectability) of
"mentalistic" explanations. The computational
idea can be seen as slicing through all three
dilemmas at a stroke; and this is what gives it,
I think, the bulk of its tremendous gut-level
appeal.
Descartes, a contemporary of Hobbes, gave the
. J/matter problem its modern form in his doc-
mm f metaphvsical dualism. Mind and body, he
tnne 0 - . .
,,'J are twO entirely different kmds of substance:
: ~ ~ ~ n e can have (as distinguishing characteristics)
. us thoughts and feelings, whereas the other
\'arlO
can have shapes, motions, and the causal inter-
actions described by physical laws (and not vice
\ersa). Intuitively, this is much more appealing
than materialism (the main alternative to dualism),
according to which everything, including minds, is
reallY just matter, in 6ne form or another. Not only
are ~ ' e reluctant to ascribe thought and feeling to
"mere" matter, but we also find it very hard to
ascribe shape and location to minds or ideas. There
is, however, one basic problem, which no dualist
has ever really solved: how can mind and body
Interact? On the one hand, they certainly seem to
interact, as when a mental decision leads to a
physical action, or when a physical stimulus leads
to a mental perception; indeed, it's not clear how
perception and action could be possible at all with-
out mind/body interaction. On the other hand,
however, physical laws are supposed to describe
all motions of all bodies completely in terms of their
interactions with one another. I In other words, phy-
sics leaves no room for causal intervention by the
mental; hence the price of mind/body interaction is
violation of the laws of physics - a price that few
philosophers (or scientists) are willing to pay.
Thought itself (quite apart from matter) is not
static and not random: it progresses and develops
in ways that obey (at least much of the time)
various rules of inference and reason. Super-
ficially, this suggests an analogy with material par-
ticles obeying the laws of physics. But the analogy
breaks down at a crucial point: particles have
neither choice nor difficulty in "obeying" physics
- it happens infallibly and automatically. People,
on the other hand, often have to work to be reason-
able; following the rules of reason is hardly infall-
ible and can be very difficult. But this means there
cannot be an explanatory dynamics of thought,
which is at all comparable to physical dynamic
theories; the respective roles of rules and laws in
the two cases are deeply different. In particular,
since correct application of the rules of reason to
particular thoughts depends on what those
thoughts mean, it seems that there must be some
active rule-applier, which (or: who) understands
the thoughts (and the rules), and which applies
the rules to the thoughts as well as it can. If the
activity of this rule-applier, following the rules of
Semantic Engines
reason, is to explain the rationality of our thought
processes, then it must be regarded as a complete
little person - or homunculus (in Latin) - inside the
head, directing the thoughts like a traffic cop. The
trouble is: a theory that invokes an homunculus to
explain thinking has begged its own question,
because the homunculus itself has to think, and
that thinking has not been explained.
Finally, there is the question of how the psy-
chology of thought could ever be properly scient-
ific. Thoughts, it seems, cannot be observed; and
the difficulty is not that, like electrons or distant
galaxies, they are too small or too far away. Rather,
they are somehow essentially subjective - we don't
even know what it would be like to observe (or
measure) them objectively. So all that science has
to go on, even in principle, is objective behanor
(which might include internal "physiological beha-
vior"); hence, thoughts can enter the picture only
as inferred or hypothesized intermediates. Unfor-
tunately, in any given case, invoking beliefs and
desires to explain an action is just too easy. What-
ever an organism does, one can always trump up a
million different ideas and motives that would
explain it. If there can be, in principle, no inde-
pendent, empirical check on which of these
hypotheses is the right one, it seems scientifically
disreputable to accept any of them, even tentat-
ively. This roughly, is the stance taken by beha-
viorism. The other half of the story, however, is
that explaining behavior without invoking inter-
vening mental processes is just too hard. In gen-
eral, the subtle regularities, connection, and
nonrandom variations in real-life behavior cannot
so much as be described in nonmentalist (pure
behaviorist) terms, let alone explained - hence,
the overwhelming mechanical ness and stupidity
of all the phenomena that the behaviorists were
ever really able to account for.
2
The upshot is
another philosophical standoff, with mentalists
and behaviorists gleefully accusing one another of
scientific bankruptcy.
Cognitive scientists can be materialists (nondual-
ists) and mentalists (nonbehaviorists) at the same
time; and they can offer explanations in terms of
meaning and rule-following, without presupposing
any unexplained homunculus. It all depends on a
marvellously rich analogy with computers - the
outlines of which we can see with a quick look at
everybody's favorite example: a chess-playing
machine. It would be very awkward and peculiar
to start assigning geometrical shapes and locations
to the internal program routines and operations
J. Haugeland
(decision processes and data structures, say) of
such a system; yet we are quite confident it has
no immaterial soul. These same decisions clearly
cause physical behavior (e.g., in teletypewriters or
TV screens), yet no one is worried that the laws of
physics are being violated. When the machine
plays, it follows rules in at least two senses: it
always abides by the rules of the game, and it
employs various reasonable rules of thumb to
select plausible moves. Though these rules are in
no way laws of nature, the machine's behavior is
explained (in part) by citing them - and yet, no
unexplained "compunculus" is presupposed.
Finally, this explanation will necessarily invoke
the system's internal reasoning processes; yet it is
far from easy to figure out (or design) processes
that will consistently lead to the observed (or
desired) behavioral responses. Moreover, for any
given machine, on any given occasion, there seems
to be a determinate right answer about which
reasonings it in fact went through.
That may have been a bit swift, but still, what an
inspiration! If there are no philosophical dilemmas
about chess-playing computers, then why should
there be any about chess-playing people - or,
indeed, about human intelligence in any other
form? To put it coldly: why not suppose that people
just are computers (and send philosophy packing)?
Well ... nothing very interesting is ever that sim-
ple. Various questions come up, among the first of
which is: What exactly is being proposed, anyway?
II Formal Systems
To start at the beginning, we must first say a little
bit more carefully what a computer is. It is an
automatic formal system. To see what this
means, we first consider what a formal system is,
and then what it is to automate one.
Aformal system is like a game in which tokens are
manipulated according to rules, in order to see what
configurations can be obtained. Basically, to define
such a game, three things have to be specified:
1 what the tokens are;
2 what the starting position is; and
3 what moves are allowed in any given position.
Implicit in (2) and (3) is a specification of what
positions are possible (for instance, what the board
is, if it's a board game). Also, there is sometimes a
specified goal position, which the player (or each
player) is trying to achieve - such as a "winning
position. "
For example, there is a familiar solitaire game in
which the tokens are pegs, arranged as follows in
the starting position:



0



The solid dots are the tokens (pegs), and the circle
in the middle is an empty space, into which a token
could fit. The only move allowed by the rules is
jumping one token over an adjacent one into an
empty space, and then removing the token jumped
over. This game has a goal: to perform such jumps
until only one token remains, and it is in the center
space.
Three points should be noticed about this (sim-
ple-minded) formal system. First, it is entirely self-
contained. Only its own tokens, positions, and
moves make any difference to it, and these only
insofar as they matter to the application of the
rules. In other words, the "outside world" (the
weather, the state of the economy, whether
the building is on fire, and so on) makes no differ-
ence whatsoever in the game. And, further, any
aspects of the tokens and positions themselves
which are irrelevant to determining which moves
are legal - e.g., (in this game) color, size, weight,
market value - are equally outside the game. Poli-
tics and courtship, by contrast, are not at all self-
contained (even though they are sometimes called
games) because just about anything could be rele-
vant in some situation or other. Second, every rele-
vant feature of the game is perfectly definite; that is,
barring outright mistakes or breakdowns, there are
no ambiguities, approximations, or "judgment
calls" in determining what the position is, or
whether a certain move is legal. For each peg and
slot, that peg is either definitely (obviously and 100
percent) in that slot, or definitely (obviously and
100 percent) not in that slot - there are no in-
between or borderline cases. Third, the moves are
finitely checkable, in the sense that for each position
and each candidate move, only a finite number of
things has to be checked to see whether that move
would be legal in that position. This is pretty trivial
for our example, but it's nontrivial and very
ortant for more complicated formal systems.
being self-containe.d, perfectly.definite,
and finitely checkable go nicely hand-In-hand;
we will say that a game or system that has all
three properties is digital. All formal systems are
digital in this sense.
The digitalness of formal systems has the fol-
lowing important consequence: two systems that
seem to be quite different may nevertheless be
essentially the same. Clearly, the peg-jumping
O"ame would be essentially unchanged if the pegs
:ere replaced by marbles, or even by helicopters
(given a big enough board) - so long as the same
rules were followed. But the differences can be
more dramatic. Imagine a game played with two
baskets and thirty-three dominoes, each with one
letter and one numeral written on it. At the begin-
ning, all the dominoes are in the start basket,
except the one marked 04, which is in the finish
basket; and the object of the game is to reverse that
situation, by a process of "switching triads." A
triad is three dominoes which have the same letter
and sequential numerals, or the same numeral and
sequential letters - so B4, C4, and 04 form a triad,
because they have the same numeral and sequen-
tial letters. Switching a triad is just moving each of
its members to the opposite basket; and this is legal
whenever the middle member and one other are in
the start basket and the third is in the finish basket.
Though one would hardly suspect it at first, it
turns out that this game (played with a certain
domino set) is essentially the same as the peg-
jumping game. It is easy to see why as soon as
the members of that domino set are listed in the
following revealing order:
A3, A4, AS,
B3, B4, BS,
CI, C2, C3, C4, CS, C6, C7,
01, 02, 03, OS, 06, 07, 04,
E1, E2, E3, E4, ES, E6, E7,
F3, F4, FS,
G3, G4, GS.
Thus switching the 02-03-04 triad in the start-
ing position would be equivalent to jumping the
peg in slot 02 over the peg in slot 03, thereby
emptying both of those slots and filling slot 04.
This kind of essential sameness among formal
systems is called formal equivalence. Two formal
systems are formally equivalent if they can be
translated back and forth in roughly the following
sense:
Semantic Engines
(I) for each position in one system, there is a
unique corresponding position in the other
system;
(2) the two starting positions correspond; and
(3) whenever you can get from one position to
another in one system, you can get from the
corresponding position to the corresponding
position in the other system.
Actually, this definition is a little more stringent
than necessary; but it gives the right idea. In
particular, it leaves room for equivalent systems
to be very different on the surface, so long as
appropriate correspondences can be found.
There are, of course, an unlimited number of
formal systems and most of the interesting ones are
significantly more complicated than our peg-
jumping example (or its domino-switching equiva-
lent). Two forms of complication are especially
widespread and important; we introduce them by
considering chess and algebra. The first is that
there can be different types of tokens, such that
what the rules permit depends on the type(s) of
token(s) involved. Thus, in chess, each side begins
with sixteen tokens of six different types; and
whether it would be legal to move a certain token
to a certain square always depends (among other
things) on what type it is - what would be legal for
a rook would not be legal for a bishop, and so on.
In fact, whether two tokens are treated equally by
the rules (in the same circumstances) is what
determines whether they are tokens of the same
type. For example, in some fancy chess sets, the
pawns are little figurines, each one different from
the next; but they are all the same type (namely,
pawns), because the rules specify the same moves
(namely, pawn moves) for all of them. To put it
another way, tokens of the same type are formally
interchangeable. Note, by the way, that the type of
each token has to be perfectly definite (and also
finitely checkable, and independent of the outside
world) if the overall system is to remain digital.
3
The second complication is that the positions of
one formal system can function as the tokens in
another ("higher level") system. We can see both
how this works and why it is important, by con-
sidering high-school algebra as a formal system.
Though one would not usually call algebra a game,
the rules have actually been formalized to the point
where it can be played like one. In the algebra
game the tokens are equations or formulae, and
the rules specify various transformations that can
be made in these formulae - or (what comes to the
J. Haugeland
same thing) various new formulae that can be
written down, given some that are already written
down. The starting position consists of a collection
of formulae that are "given" (including, of course,
all the axioms); and making a "move" is adding a
formula (theorem) to this list, by following the rules
- that is, giving a formal deduction or proof(or, at
least, a step in one). Now, the difficulty is that each
different algebraic formula is a different type of
token in this system, and there are indefinitely
many of them - so how can the moves be finitely
checkable? And the answer, of course, is that all
these different tokens are built up in a systematic
way out of a comparatively small collection of
letters and standard symbols.
More specifically, each algebra-game token is a
legal position in another game, which we might
call the "well-formed-formula game." The tokens
of the latter (i.e., the letters and symbols) are
usually called "simple" or "atomic" tokens, to
distinguish them from the tokens of the algebra
game (i.e., the well-formed equations), which are
"compound" or "molecular" tokens. The point is
that the rules of the algebra game can apply to the
various (molecular) tokens in terms of their struc-
ture; thus for any equation that has the same
addend on both sides, you can "cancel" that
addend. So, the same rule can apply in a parallel
fashion to tokens of indefinitely many different
types, so long as they have the specified structural
features; hence, finitely many rules can suffice for
the algebra game, even though there are indefin-
itely many types of well-formed algebraic formu-
lae. This general strategy of using the positions of
one formal system as the tokens of another makes
large complicated systems much easier to deal
with, and it is found throughout mathematics,
formal logic, and virtually everywhere else that
formal methods are used.
It might seem odd to include mathematics and
logic here in the same category with chess and the
peg-jumping game - because, roughly, their
tokens mean something (and thus might be true
or false, say). But their inclusion is perfectly seri-
ous and very important. Most mathematical and
logical systems are formal in exactly the same sense
that chess, checkers, and the like are formal: they
have starting positions, and rules for making
moves in any given position, and they are digital,
in the sense explained above. From this point of
view, any meanings that their tokens might have
are utterly irrelevant; meaning has to do with the
outside world, and is in no way part of any self-
contained formal system, as such. There are, of
course, other points of view, in which meaning is
very important (we will discuss some of these
when we come to interpretation and semantics).
But, considered only as .lilrmal systems, games,
logic, and mathematics are all equally meaningless,
and entirely on a par.
III Automatic Formal Systems (Turing
Machines and Computers)
An automatic formal system is a physical device
(such as a machine) which automatically manipu-
lates the tokens of some formal system according
to the rules of that system. It is like a chess set that
sits there and plays chess by itself; without any
intervention from the players, or an axiomatic
system that writes out its own proofs and theo-
rems, without any help from the mathematician.
The exciting and astonishing fact is that such
systems can be built. Looked at in the right way,
this is exactly what computers are. There are two
fundamental problems in building an automatic
formal system. The first is getting the device to
obey the rules (in principle, this problem has
already been solved, as will be explained in
the remainder of this section). The second is
the "control" problem - how the device selects
which move to make when there are several legal
options. We will consider this briefly in the next
section.
The theoretical ancestor of all automatic formal
systems is a class of devices invented (in the
abstract) by the mathematician Alan Turing and
now called Turing machines. A Turing machine
has:
an unlimited number of storage bins;
2 a finite number of execution units; and
3 one indicator unit.
The indicator unit always indicates one execution
unit (the "active" unit), and two storage bins (the
"in" and "out" bins, respectively). Each storage
bin can contain one formal token (any token, but
only one at a time). Each execution unit has its
own particular rule, which it obeys whenever it is
the active unit. What that rule specifies will
depend on what token is in the current in-bin;
and in each case it will specify two things: first,
what token to put in the current out-bin (discard-
ing the previous contents, if any), and second,
what the indicator unit should indicate next. The
machine proceeds by steps: the active execution
unit checks the in-bin, and then, according to what
it finds there and what its rule is, it refills the out-
bin and resets the indicator unit; then the next step
begins. Usually there is one execution unit which
nothing; so if it ever gets activated, the
. -l
mach me stops.
Clearly, any Turing machine is an automated
\ersion of some formal system or other. The start-
in" position is the initial contents of the storage
o .
bins, the moves are the machme steps, and the
rules are those which the execution units obey
(the control problem is handled in the rules for
resetting the indicator unit). Not so obvious, but
Jlmost certainly true, is the converse: any auto-
mJtic formal system can be formally imitated by
some Turing machine.
s
"Formal imitation" is like
formal equivalence, except for two things. First,
since we are talking about automatic systems that
is, svstems that actually "choose" which move to
make an imitating system has not merely to offer
corresponding legal options in each corresponding
position, but also to make the corresponding
choice in each case. Let's call systems which are
formally equivalent, and which make equivalent
choices in each position, equivalent.
Second, the imitating system is often divided
into two parts: the part which directly corresponds
to the system being imitated, and another
part which works behind the scenes, making
everything come out right. The first part is called
the rirtual machine, and the second part the
program.
For example, suppose some automatic formal
system, A, is being formally imitated by some
Turing machine, T. Then there is some virtual
machine, V, which is both a part of T, and dynam-
ically equivalent to A. So some portion (say half) of
T's storage bins will be allocated to V, as V's
storage bins; the tokens that appear in these bins
will constitute V's positions (which, in turn, cor-
respond to A's positions). The rest of T's storage
bins contain the program (and perhaps some
"scratchpad" workspace). The reason a program
is necessary is that in general V's rules will be
different from T's rules. The program is a (finite)
set of tokens so contrived that when T obeys its
own rules with respect to all of its storage bins, it
will, in effect, obey V's rules with respect to the
tokens in those bins that have been allocated to V.
Intuitively, we can think of the program as "trans-
lating" V's rules into T's rules, or even as "telling"
Semantic Engines
T what V's rules are, so that T can follow them in
moving the tokens in V's bins. In fact, it's not
quite that straightforward, and T's rules have to
haw a certain versatility in the first place, in order
to make such an imitation possible. But the point
is: for any formal system, there is a Turing
machine that can be programmed to imitate it.
But the fundamental importance of Turing
machines rests on yet another truly amazing fact
a theorem first proved by Turing which has,
perhaps more than any other single result, shaped
modern computer science. It is that there are spe-
cial Turing machines, called un/ursal Turing
machines, which can be programmed to imitate
any other Turing machine. In particular, one
could imitate a Turing machine that was itself
imitating some other automatic formal system
which means that, indirectly, the universal
machine is also imitating that other automatic for-
mal system. So, combining Church's thesis and
Turing's theorem, a universal Turing machine
can (suitably programmed) imitate any automatic
formal system whatsoever! To put it another way:
if you have just one universal Turing machine, and
you are prepared to do some programming, you
can have (a formal imitation of) any automatic
formal system you care to specify.
It was soon discovered that there are a number
of different kinds of universal machine, which are
not (in the strict sense) Turing machines. But
since they are universal, they can formally imitate
any Turing machine; and, of course, any universal
Turing machine can formally imitate any of them.
In principle, therefore, it doesn't matter which one
you have any universal machine will do as well as
any other (except for differences in efficiency,
elegance, and the like). The reason this is import-
ant is that, with one qualification, universal
machines can be built; that is what digital com-
puters are. The one qualification is that a true
universal machine must have unlimited storage,
whereas any actual machine will have only a cer-
tain fixed amount (though it can be very large). So,
aside from limitations of memory size, any stand-
ard digital computer can, with appropriate pro-
gramming, formally imitate any automatic formal
system yet discovered; that, basically, is why com-
puters are so powerful.
Interestingly, computer programmers almost
never program any of the machines that are actu-
ally constructed out of transistors, wires, and so
on; rather, most programs are written for rirtual
machines, which are themselves merely being
J. Haugeland
imitated by the hardware machine. The reason is
that some universal machines are cheaper to build,
while others are more convenient to use. The
sensible course for a manufacturer, therefore, is
to build the cheaper one, hire a few experts to
write complicated programs (called "interpreters"
or "compilers") which will make it imitate the
more convenient ones, and then sell the package.
The computer "languages" that you hear about -
BASIC, FORTRAN, PASCAL, LISP, and so on
- are really just some of these more convenient
universal machines, which are widely imitated by
various hardware machines. In fact, there are often
more layers than that: thus the hardware directly
imitates some intermediate virtual machine,
which, in turn, is programmed to imitate the
higher level machines, like those mentioned
above. And, of course, when a programmer pro-
grams the latter, he or she is really designing yet
another (virtual) machine, which the FORTRAN
or LISP machine, or whatever, will formally imit-
ate. This last machine, incidentally, will typically
be designed for some special purpose, and so prob-
ably will not be a universal machine (though it
could be).
A fundamental point should now be obvious: a
particular physical object can be, at one and the
same time, any number of different machines. There
is no single correct answer to the question: which
machine is that (really)? Of course, at the hardware
level it is some particular model from some man-
ufacturer; but at the same time it is (just as
"really") all the other machines that are being
imitated at various other levels. So a particular
object is a particular machine (or automatic formal
system) only at a particular level of description - at
other levels of description it is other machines.
Once we see this point, we see how foolish it is
to say that computers are nothing but great big
number crunchers, or that all they do is shuffle
millions of "ones" and "zeros." Some machines
are basically numerical calculators or "bit" manip-
ulators, but most of the interesting ones are noth-
ing like that. And the fact that most actual
commercial equipment can be described as such
machines (e.g., most can be described as bit
manipulators, on some level) is of no theoretical
consequence. The machine one cares about - per-
haps several levels of imitation up from the hard-
ware - may have nothing at all to do with bits or
numbers; and that is the only level that matters.
For instance, an automatic system that played the
peg-jumping game would probably not refer to
numbers. Part of the reason for using the expres-
sion "automatic formal system" in place of "com-
puter" is that the latter suggests a device which
only "computes," and that is just wrong.
IV The Control Problem
In the last section we considered systems in which
each move is completely determined by the rules
and the current position. This is essential in the
design of an automatic system, because each step
has to be made "automatically" - that is, it must
be completely fixed by the state of the machine at
the time.
6
But in most positions in most formal
systems, anyone of various moves would be legal;
usually, that is what makes them interesting, as
games or whatever. Does it follow that such inter-
esting formal systems cannot be automated? No; it
only follows that some device for deciding among
the several legal options at any point must be
automated as well.
The easiest way to think about it is to consider
the machine as divided into two parts or "subma-
chines": one to generate a number of legal options,
and another to choose from among them. This, of
course, is just an extension of the basic point that a
given device can be various machines, depending
on how you look at it; only now we are looking at it
as two separate (virtual) machines, interacting with
each other on the same level. The advantage is that
we can see how the above dilemma is resolved: the
move-generating submachine automates an "inter-
esting" system, in which a variety of moves might
be legal at any point; but the combined machine,
with both parts together, satisfies the requirement
that some particular next move be determined by
the overall state of the device at each step. Design-
ing the second submachine - the one that makes
the choices, given the options - is the control
problem. In most cases, control design turns out
to be the hardest part of automating an interesting
formal system.
In an average chess position, for example, a
player will have 30 or 35 legal possibilities to
choose from. A beginning chess player could dis-
cover and list all of them without too much trou-
ble; and designing a machine to do the same would
not be terribly difficult either. The hard part - the
entire difference between amateurs and world
champions - is deciding which move to make,
given these few possibilities; and that is the hard
part to program as well. At first it might seem that
b
modern computers, with their tremendous
Ig .
speed and memory, could Just ahead to
every possible outcome and see which moves lead
to victory. In principle, this would be
ossible, since chess is technically finite; and
a machine would be literally invincible. In
practical terms, however, such a computation is
nowhere near possible. Assuming an average of
31.6 options per play gives a thousand
(31. 6 x 31. 6) possible combinations per full move
(each side having a turn). Thus looking ahead five
moves would involve a quadrillion (10
1
;) possibil-
ities; forty moves (a typical game) would involve
10
120
possibilities. (For comparison, there have
been fewer than 10
18
seconds since the beginning
of the universe.) These numbers are just prepos-
terously large for any physically conceivable com-
puter. They get that big because the number of
choices at each additional step multiplies the total
number of possible combinations so far. For
understandable reasons, this is called the combina-
tOrial explosion; it plagues control design for all but
the most straightforward problems.
Obviously, human chess-players don't make
that many calculations either; in fact, the available
evidence indicates that they make rather few. The
trick is to consider only the relevant possibilities
and ignore the rest. Thus most of those 30 or 35
options in a typical chess position would be totally
pointless or manifestly stupid; hence, it would be a
waste of time to consider all the possible develop-
ments that could follow after them. If the number
of relevant alternatives could be reduced, say, to
three at each stage, then looking ahead five com-
plete moves would involve only 59,049 possible
combinations - still too many for a person to
consider (consciously, anyway), but well within
reach of computers. So, the approach to the con-
trol problem in this case will almost certainly con-
centrate on determining which possible moves are
the relevant ones, deserving further consideration.
Unfortunately, there is no fail-safe way to tell
what is and isn't relevant. Everybody knows how a
seemingly pointless or terrible move can turn out
to be a brilliant stroke (once the opponent takes the
bait). Any method that systematically bypasses
certain moves as not worth pursuing will inevitably
overlook some of these brilliancies. What we want
is a method that is efficient at bypassing moves
that really are worthless, but not easily fooled into
ignoring moves that only seem worthless. Such
methods in general are called "heuristics" (from
the Greek word for "discover"), in contrast to
Semantic Engines
"algorithms" (from the Latin word for the Arabic
system of numerals, named after an Arabic math-
ematician). In current usage, the essential differ-
ence is this: an algorithm is a rule or procedure that
is guaranteed to give a result meeting certain con-
ditions (you just turn the crank and out it pops); a
heuristic, on the other hand, is a rule or procedure
that is more or less reliable, but not infallible - just
a rule of thumb. We have algorithms for multi-
plying integers, arranging words in alphabetical
order, and finding a checkmate in certain chess
end-games (king and rook against king, for
example). But there are no feasible algorithms
for finding the best move in most other chess
positions, or arranging words into poetry, or dis-
covering proofs of arbitrary theorems in number
theory. In these cases people proceed by intuition,
inspiration, and a few explicit heuristic rules of
thumb. (Intuition and inspiration might just be
unconscious heuristics, of course - that's a
disputed point among psychologists.)
Automatic systems, in any case, must proceed
according to explicit rules (explicit, at least, in the
program of the virtual machine). And in one sense
these rules have to be algorithms - they have to
determine the next move of the machine defini-
tively, at each step. But often, whether a given rule
is an algorithm or a heuristic depends on how the
desired result is specified. Thus a reasonable rule of
thumb for winning chess games is: never trade
your queen for a pawn. Occasionally, of course, a
queen sacrifice is a brilliant play; so this rule is
only a heuristic when the result is specified as
"Find the best move." But if the specified result
is "Avoid swapping a queen for a pawn," then this
rule is (trivially) infallible. The point is that the
rules followed by the machine only have to be
algorithms in this trivial sense. That is, it can
perfectly well follow a bunch of inconclusive
rules of thumb, relative to the interesting result
specification (winning, say), so long as there are
algorithms defining the heuristics themselves. The
machine can infallibly follow quite fallible rules.
This shows, by the way, what's wrong with the
idea that a (properly functioning) computer never
makes a mistake. It just depends on what counts as
a mistake - i.e., relative to which result specifica-
tion. A chess-playing computer can function per-
fectly, never making a single mistake in following
its internal heuristics but making lots of mistakes
in the game, because its heuristics are rather
clumsy. It is only at special tasks (like arithmetic
and sorting), where there are algorithms for the
J. Haugeland
interesting results, that a (perfect) computer can be
infallible.
If we construe "heuristics" broadly to include
any methods that improve a system's chances of
making "correct" decisions while lea\'ing some
possibility of "mistakes," then any techniques for
providing default assignments (standard assump-
tions), jumping to conclusions, or reading between
the lines will be heuristic procedures. In this broad
sense, a major part of the effort in artificial intelli-
gence research goes toward finding better heur-
istics - ones that are more efficient and harder to
fool - and better ways of implementing them on
the machines (i.e., better algorithms for defining
the heuristics). Indeed, in certain specialized areas,
like game playing, theorem proving, and most
tasks in "micro-worlds," work on heuristics is
most of the problem. These areas are distinguished
by the fact that the desired result is already known,
and easy to specify in formal terms; for instance, in
chess the goal is winning, and winning positions
(checkmates) can be defined precisely. Many other
forms of intelligent behavior, however, like writing
good poetry (to take an extreme case) or carrying
on a normal conversation (to take a case that does
not seem so extreme) are not well defined in this
way. In these areas, a major effort is required even
to characterize the results toward which the heur-
istics are supposed to guide the system - and this
involves the semantic and pragmatic issues to
which we will turn shortly.
V Digital and Analog
Automatic formal systems are, by definition, digi-
tal computers. There is another kind of device,
sometimes called an analog computer, which is
really quite different. Digital systems, remember,
are self-contained, perfectly definite, and finitely
checkable - all with regard to which moves are
legal in which positions, according to the rules.
An analog device, on the other hand, doesn't
even have clearly defined moves, rules, and posi-
tions - though it may have states (which may
change), and there is usually some way that it is
supposed to work. The crucial difference is that in
analog systems the relevant factors have not been
defined and segregated to the point where it is
always perfectly definite what the current state is,
and whether it is doing what it is supposed to do.
That is, there will often be slight inaccuracies, and
marginal judgment calls, even when the device is
working normally. To take the very simplest
example, it is like the difference between a multi-
position selector switch and a continuous tuning
dial on a stereo. A switch has a number of click-
stops, and it is always set definitelv at one or
another of them - you cannot set it between AM
and F:'v1. Of course, when you rotate it, it passes
through intermediate angles; but these are irrelev-
ant to its function, and (if it is a good switch) can
be ignored. A tuning dial, on the other hand,
moves smoothly without click-stops, and each
angle tunes a different frequency. It is perfectly
possible to set the dial between two stations, and,
in fact, getting it "right on" a station can require a
sensitive judgment call.
F or some purposes, analog devices are very con-
venient. Scale models are a case in point. Suppose
an architect wants to find out how the light and
shadows will fall on a complicated structure at
different times of day. A good analog approach is
to build a model, illuminate it with a spotlight from
various directions, and then look at it from various
directions. A trickier but similar case is the use of
laboratory animals in medical research. To esti-
mate the physiological effects of weird drug com-
binations, experimenters can give corresponding
(or exaggerated) doses to rats and then just wait
to see what happens. Other common examples of
analog devices are slide rules, electronic harmonic
systems (modeling mechanical harmonic systems),
and string-net pathfinders for railroad or highway
networks. Though "analog" is itself a rather ill-
defined notion, it clearly encompasses quite valu-
able tools.
Digital systems, however, have several inherent
advantages. First, of course, universal machines
are, by their very nature, extremely versatile; and
that makes them more economical for each
application. Second, analog systems can them-
selves often be digitally simulated - which makes
the digital system even more versatile. For exam-
ple, the architect's model could be simulated by
writing equations for all the opaque surfaces in the
building, and then calculating the paths of indi-
vidual light rays at, say, one-inch intervals (for
various positions of the sun). The whole thing
might then drive a TV display, set for various
viewing angles; and if all the intervals are small
enough, the result can be pretty good. The amount
of calculation involved can be prodigious, and the
general technique has all the theoretical elegance of
sandblasting; but computers are cheap, fast, and
tireless, so it often works nicely.
It is sometimes said (with a grand air) that any
,10 device can be digitally simulated to any
ana g h" I' I d'
desired precision; t IS. IS gross y mls ea mg.
Digital simulation I.S only when all the
tl
'\'e relationships m the analog svstem can
opera . '.
be described in a relatively compact and precise
. e g not too many equations with not too
wa\ - .. , .
ma'n
v
\ariables. The architectural simulation, for
depends on assuming that all the light
comes from the same direction and travels in a
straight line until it's blocked by one of the few
specified opaque surfaces. Without this simple
structure to work from, the simulation could not
get off the ground. Thus there is no comparable
general description of the physiology of laboratory
rats pumped full of odd chemicals - there are
billions of potentially relevant subtle interactions
and responses in a complex organic system, and
the scope for combinatorial explosion is essentially
indescribable. Hence, digital simulation is out of
the question, except in very special cases where it
is known in advance that only a certain few vari-
ables and relationships are relevant. It might seem
that, in principle, simulation must be possible any-
way, because rats are made of atoms, and each
atom obeys known equations. But such a principle
is utterly out to lunch. A single large molecule may
have so many interdependent degrees of freedom,
that no computer yet built could simulate it reli-
ably in reasonable time; and one rat contains more
molecules than there would be minicomputers if
the entire volume of the earth were packed solid
with them.
The real theoretical advantage of digital systems
lies in quite another direction, and depends
specifically on their digital nature. Consider two
conventions for keeping track of money in a poker
game. Each uses different colors for different
denominations: blue, red, and white for a hundred,
ten, and one, respectively. But in one system, the
unit of each denomination is a colored plastic disk
(i.e., a poker chip), whereas in the other system it
is a tablespoon of colored sand. The latter arrange-
ment does have some merits - particularly the fact
that fractional bets are possible - using less than a
full tablespoon of white sand. But the chip system
has one overwhelming advantage: all the bets are
exact. By contrast, in measuring volumes of sand,
there is always some small error, no matter how
careful you are. Using the chips, and a very modest
degree of care, it is possible to wager exactly 13
units (plus or minus zero); but with the sand this is
impossible, even given the finest instruments in
Semantic Engines
the world. The difference is that the chip system is
digital; each token is perfectly definite, and there is
no need for judgment calls.
The more complex and interdependent a system
becomes, the more vulnerable it is to errors that
propagate and get out of contra!' A small error in
one component, affecting a more sensitive com-
ponent, can result in a larger error going to the
next component, and so on. \Ne can see a rudi-
mentary form of this, even in the poker example.
Suppose the measurement error on sand volumes
is 2 percent; then it would be perverse to try to
bet, say, 613 units, because the measurement error
on the six blue tablespoons is worth four times as
much (on the average) as all three white table-
spoons - the latter simply drop out as irrelevant.
There are ways to control such errors, of course,
but only to a certain extent, and they tend to get
expensive. The perfect definiteness of digital
tokens, though confining in some cases, pays otT
in the elimination of this sort of error; thus,
though you cannot bet a fraction of a unit with
poker chips, there is no problem in betting 613
units - since six blue chips can be counted exactly,
the white chips remain perfectly significant. This
advantage is progressively more important for lar-
ger and more complicated systems - hence the
contemporary predominance of digital methods.
VI Semantics
Formal systems (and computers) can be more than
mere games, because their tokens can have inter-
pretations that relate them to the outside world.
This is the domain of semantics and pragmatics.
Sometimes we say that the tokens in a certain
formal system mean something - that is, they are
"signs," or "symbols," or "expressions" which
"stand for," or "represent," or "say" something.
Such relations connect the tokens to the outside
world (what they are "about"), making it possible
to use them for purposes like record-keeping, com-
munication, calculation, and so on. A regular, sys-
tematic specification of what all the tokens of a
system mean is called an interpretation; and the
general theory of interpretations and meanings is
called semantics. Accordingly, what any token
means or says, and hence also whether it is true
or false, and so on, are all semantic properties of that
token.
Semantic properties are not formal properties. A
formal system as such is completely self-contained
J. Haugeland
and, viewed in that way, is just a meaningless
game. In strictly formal terms, interpretation and
meaning are entirely beside the point - they are
extraneous "add ons" that are formally irrelevant.
(When discussing a system that is to be inter-
preted, we call its purely formal characteristics
and structure its syntax; "syntactical" is just
another word for "formal," but it is generally
used only when a contrast to "semantic" is rele-
vant.)
So, formal tokens can lead two lives: syntactical
(formal) lives, in which they are meaningless mar-
kers, moved according to the rules of some self-
contained game; and (if the system is interpreted)
semantic lives, in which they have meanings and
significant relations to the outside world. The
story of how these two lives get together is the
foundation of modern mathematics and logic; and
it is also the philosophical inspiration of cognitive
science. We review the mathematical!logical case
first, however, because it is better understood and
there are fewer complications.
Consider a formal system (like the algebra game,
discussed above in section 2) in which the posi-
tions are just sets of tokens, and the legal moves are
just to add more tokens to the current position
(depending, of course, on what is already in it).
And suppose this system is interpreted in such a
way that each of its (well-formed, complete) tokens
"asserts" something - that is, each token (accord-
ing to the meaning assigned to it by the interpreta-
tion) expresses some claim about the world. Then,
depending on what the claim is and what the world
is like, each of these tokens will be either true or
false (relative to that interpretation). Now, the
rules of such a system will have the semantic prop-
erty of being truth-preserving if and only if they
meet the following condition: for any position
which already contains only true tokens, any
other token which can be added to that position
(according to the rules) will also be true. In other
words, if you start with tokens which are all
true (under the interpretation), and if you obey
the (purely formal) rules, then you can be sure
there are no false tokens in any position you
ever reach.
The rules of standard logical and mathematical
systems, relative to their standard interpretations,
are all truth-preserving; and, of course, the tokens
in their starting positions (i.e., their axioms) are all
true. Therefore, any token in any legal position of
one of these systems is guaranteed also to be true!
That is why we know in advance that their theo-
rems (which are defined in purely syntactical! for-
mal terms) are all true (which is a semantic
property). Or, what comes to the same thing, in
order to establish the semantic truth of a token in
such a system, it suffices merely to prove it form-
ally (play the game). This is how the "two lives" of
the tokens get together; and it is the basic idea
behind the formalization of modern logic and
mathematics. In effect, given an interpreted formal
system with true axioms and truth-preserving
rules, if you take care of the syntax, the semantics
will take care of itself7
Most mathematical and logical systems are only
partially interpreted, in the sense that some of
their atomic tokens are left as variables, whose
exact interpretation is to be specified by the user
(in certain allowable ways). For example, in ordin-
ary algebra, you can specify what the letters (vari-
ables) stand for in any way you want - so long as
they stand for numbers. The important thing is
that the rules are still truth-preserving, no matter
what specific interpretation is given to the vari-
ables. For example, if a user knows, relative to
some specific interpretation of the variables (as
values in a physics problem, say) that the token
a - b d - b e( I + e) - e
--+c=--+----
c c c
is true, then he or she can apply the purely formal
rules for multiplying through, deleting parenth-
eses, collecting terms, etc., and be assured that
the token
is also true (relative to the same interpretation).
This is a vivid example of how useful it can be to
have the semantics "take care of itself," if only one
plays the syntactical game correctly.
An even more vivid example is computers,
because precisely what an automatic formal system
can do is "take care of the syntax" - i.e., play by
the rules. The machine does not have to pay any
attention to the interpretation or meaning of any of
its tokens. It just chugs along purely formally; and
if the starting position, rules, and interpretation
happen to be of the sort we have been discussing,
then it will automatically produce only truths.
Given an appropriate formal system and
interpretation, the semantics takes care of itself.
This, fundamentally, is why computers can be so
useful for calculations and the like - why they can
be computers and not just electronic toys.
An automatic formal system with an interpreta-
tion such that the semantics wiJ1 take care of itself
. 'hat Daniel Dennett (1981) caJ1s a semantic
15 "
"
ne The discovery that semantic engines are
t!IIg .
possible - that with the kind of formal sys.tem
and interpretation, a machme can handle meanmgs
_ is the basic inspiration of cognitive science and
artificial inteJ1igence. Needless to say, however,
mathematics and logic constitute a very narrow
and specialized sample of general inteJ1igence.
People are both less and much more than auto-
matic truth-preservers. Consequently, our discus-
sion of interpretation needs to be expanded and
generalized considerably; unfortunately, the issues
will get messier and murkier as we go.
VII Interpretation and Truth
Interpretation is especiaJ1y straightforward in the
special case of logic and mathematics because, in
the final analysis, the only semantic property that
matters is truth. Hence, it suffices if the theorems
are guaranteed true, given the truth of the axioms
(and special assumptions, if any) - that is, it suf-
fices if the rules are truth-preserving. We shall see
in the next section that there are many other cases
where truth is far from all that matters in judging
interpretations. But it is worth staying with the
special case a little longer, to ask why truth has the
importance it does; then it wiJ1 be clearer why
other considerations are important in other cases.
Imagine finding an automatic formal system
that uses ordinary numerical tokens, and generates
"theorems" (i.e., outputs) like the following:
= 5 =
71 92x
= - 1 - 8
= 61 = 040
+ 1 - 5940
84 - 1 x 5 =
Formally, of course, there is nothing wrong with
these tokens; we can imagine any number of
(strange and boring) games in which they would
be perfectly legal moves. Semantically, on the
other hand, they look like nonsensical, random
gibberish - "arithmetic salad." That is, it seems
impossible to construe them as expressing claims
about the relationships among numbers (e.g.,
equations). But hold on: this reaction depends on
a tacit adoption of the familiar Arabic interpreta-
tion of what the numerals and signs mean (the
digit "I" stands for the number one, the "+"
sign stands for addition, and so on). Formally,
Semantic Engines
however, these numerals and signs are just neutral
marks (tokens), and many other (unfamiliar) inter-
pretations are possible (as if the outputs were in a
code). Suppose, for instance, we construed the
atomic tokens according to the following non-
customary scheme (using ",*" to abbreviate
"stands for"):
" 1" '* equals "6" '* zero
"+" '* fiye
"2" '* plus "7" '* one "-" '* six
"3" '* minus "8" '* two " x" '* seven
"4" '* times "9" '* three '* eight
"5" '* div. by "0" '* four "=" '* nine
Then, with this table, we could translate the sys-
tem's outputs back into the familiar notation as
98 - 6 = 62
1 = 83 + 7 90 = 94 x 40 2 x 6 = 7 9
Superficially, these look more like equations, and
hence not nearly as random or crazy as the raw
(untranslated) outputs. Unfortunately, they're all
false - wildly false. In fact, on closer inspection, the
digits look just as random as before; we stilJ have
arithmetic salad, only disguised in regular equation
format.
But there are over a trillion possible ways to
interpret these fifteen atomic tokens (even sticking
to permutations of the ordinary one). Here is just
one more possibility, together with the translations
it would yield for the original outputs:
"1" '* equals
"2" '* div. by
" 3" '* times
"4" '* minus
"5" '* plus
1+1=2
9 = 27 3
"6" '* zero
"+" '* five
"7" '* nine "-" '* four
"8" '* eight " x" '* three
"9" '* seven '* two
"0" '* six "=" '* one
12 x 4 = 48 5 = 4 + 7 - 6
10 = 16 - 6 8 - 4 = 3 + 1
What a difference! These not only look like equa-
tions, they are equations - they are true. And,
intuitively, that strongly inclines us to prefer this
interpretation - to think it "better" or "righter"
than the ones which yield random nonsense as
readings. This intuition has several theoretical
considerations behind it.
In the first place, formal tokens in themselves
never intrinsically favor one interpretation scheme
over any other - from their point of view, aJ1
interpretations are equalJy extraneous and arbi-
trary. So if some particular interpretation is to be
J. Haugeland
adopted, over all the other possibilities, there must
be something distinctive about it, which makes it
stand out in the crowd. But, as our example sug-
gests, an interpretation that renders a system's
theorems as truths is a rare and special discovery
(the alternatives yield mostly gibberish). Hence,
such an interpretation is distinctive.
Second, if we regard an interpretation as relating
or (onnecting a formal system to the outside world
(i.e., to whatever its tokens are "about"), then the
distinctiveness of the preferred interpretation
should lie in that relation. But if a system's theo-
rems are all true under an interpretation, then
there is, in fact, a relation strong enough to sup-
port inferences about the world from the theo-
rems, and about the theorems from (knowledge
of) the world. Thus we can discover new facts
about numbers, using an interpreted formalization
of arithmetic; likewise, the above-noted {practical}
utility of the semantics "taking care of itself'
depends on the truth-preservingness of the rules.
In the other direction, we {sometimes} can predict
theorems, given what we know about numbers.
For instance, if our example system produced the
theorem fragment "7401 ... ", we would be hard
pressed to guess its completion, from its form
alone. The first interpretation scheme {which
gives" 1 x 4 = ... "} is no help either, because its
"translations" come out randomly false. But
knowing that the second scheme gives "9 - 6 =
... ", and that the theorem will be true, makes the
prediction easy: the completion will be whichever
token means "three" {namely, "x "}. The reliabil-
ity of such inferences (both ways) indicates that
this relation between the tokens on the one hand
and the numbers and operations on the other is not
arbitrary - i.e., this interpretation somehow genu-
inely connects them.
Finally, and most important, interpreting is tan-
tamount to "making sense of'; hence, if the system
doesn't end by making sense {but, rather, makes
nonsense}, then the interpretation attempt has
failed. Arithmetic salad does not make sense
{whether clothed in the outer form of equations,
or not}; and that, primarily, is why random inter-
pretation schemes don't really give interpretations
at all. In the context of arithmetic, true equations
make eminently good sense - hence the preferabil-
ity of our second scheme. (An occasional lapse
could be understood as meaningful though false;
but constant, random falsehood is, in effect,
impossible - because the tokens cease to make
any sense at all.) So truth matters to interpreta-
tions not only because it provides a nonarbitrary
choice among candidate schemes and because this
choice reflects some relation between the system
and what it is (interpreted to be) "about," but also,
and most of all, because wild falsehood amounts to
nonsense - the antithesis of meaning.
VIII Interpretation and Pragmatics
In most activities involving meaningful tokens
(other than logic and mathematics) truth is far
from the only semantic property that matters.
Take ordinary conversation, for example. In the
first place, many speech acts - such as questions,
commands, expletives, and even most banter and
quips - are neither true nor false; so some other
kind of appropriateness must be what matters.
And even in the case of statements {which typically
are true or false}, much else is important besides
just whether they are true. For instance, in con-
versation it is also important {usually} to stick to
the topic, be polite, say only what is worth saying
(not obvious, inane, redundant, etc.) - and, in
general, to avoid making an ass of oneself. It is
not clear to what extent such conditions should be
called "semantic" {some authors prefer to call
them "pragmatic"}, but they are all relevant to
the acceptability of an interpretation as soon as
you get away from special cases like logic and
mathematics.
In these cases we are driven back to the more
general but also somewhat fuzzier notion of "mak-
ing sense" as our criterion for the adequacy of
interpretations. There is, to my knowledge, no
satisfactory philosophical account of what it is to
make sense; indeed, it is questionable whether a
precise, explicit definition is even possible. Still,
there is much to be said, and philosophers have
worked out a variety of rough characterizations
and prerequisites (sometimes referred to by the
peculiar misnomer "principles of charity"). The
truth of simple, uncontroversial declarative out-
puts {i.e., tokens interpreted as such} is, of course,
one rough prerequisite on making sense in general;
and a few philosophers (notably, Davidson, 1973)
are inclined to stick with that. But most would
agree there are other considerations.
The most widely discussed, rationality, is
loosely analogous to truth-preservation in mathe-
matical systems. The idea is that "obvious con-
sequences" of tokens in the current position
should be relatively easy to evoke from the system
t
s) or to get it to add to the position. In
(as outpU '. ..
I
the point applies not only to lOgIcally valid
genera, ..
s but to common sense mferences ot all
IOference , .. .
. d ncIuding the formation of subSidiary goals
klO s, 1 .
. d attitudes (given the facts, goals, and attitudes
lIn d' at hand) and also the ability to solve suit-
a rea \
ably "problems." Further, the system
sh;uld have a tendency to root .and
'ncompatibilities among the tokens m Its positions
c.g., by altering or eliminating one of the offend-
ing parties. It goes without this notion
is not very precisely defined; but It IS equally clear
that "rationality" in some such sense is an import-
ant factor in "making sense."
."- second important factor is reliable interaction
with the world via transducers. A transducer is a
kind of automatic encoder or decoder which either
reacts to the physical environment and adds tokens
to the current position (an input transducer), or
else reacts to certain tokens in the current position
and produces physical behavior in the environ-
ment (an output transducer). So transducers are
analogous to sense organs and motor systems in
animals and people; and they become, in effect,
part of the whole system, when it comes to deter-
mining whether (according to some interpretation)
the system is making sense. Roughly, when an
input transducer adds a token that is interpreted
as a report, then (other things being equal) that
report ought to be correct; and when an output
transducer reacts to a token interpreted as a deci-
sion to do something, then (other things being
equal) the behavior produced ought to be that
action. For instance, if the token that regularly
gets added in response to a rabbit running by
were interpreted to mean "Lo, a flaming giraffe,"
then the system isn't making much sense - which
is to say, the interpretation is probably defective.
This is essentially Q!.tine's (1960: ch. 2) criterion of
preserving "stimulus meaning." Though he
doesn't mention it, there is a corresponding con-
dition on the reliability of output transducers.
A third condition on making sense is conversa-
tional cooperativeness, in more or less the sense
introduced by Grice (1975) - assuming, of course,
that the system is interpreted as "conversing" or
communicating in some way. Suppose, for
example, that you ask it what the capital of Illinois
is (i.e., give it a token so interpreted); then "It is
Springfield" or "I don't know" would be perfectly
sensible replies, but "Easy on the mustard,
please," or "Wow, 400 squirrels" would not.
Even a false but relevant answer, like "It is Chi-
Semantic Engines
cago," makes much more sense in context than
some utterly irrelevant truth, such as
"2 + 7 = 9." Answering the question one is
asked is a very basic example of being cooperative;
there are many subtler points as well, concerning
how much information is given in the answer, the
manner in which it is provided, how it pertains to
the larger topic, whether it is rude, obvious, or
funny, and so on. Cooperativeness is clearly a
matter of degree, involving judgment and compro-
mise; but if it is ever lacking altogether, a "con-
versation" quickly reduces to impossible nonsense.
A related consideration is what Austin (1970)
called "felicity conditions" (compare also Searle's
1969 "constitutive rules"). Roughly, the idea is
that one cannot or cannot "properly" make an
unkeepable promise, threaten someone with some-
thing he or she wants anyway, give commands for
which one lacks the authority, propose marriage to
one's spouse, offer to sell (or buy) the planet Mars,
and so on. That is, these "speech acts" or "per-
formances" have various prerequisites or presup-
positions; and if the prerequisites are not satisfied,
there is something wrong with doing them. As
with all of our conditions, a few such violations
look merely foolish, or perhaps dishonest; but
wholesale and flagrant disregard will yield an over-
all pattern of outputs which fails to make sense -
and hence the interpretation will be unsatisfactory.
These last two points, more than the earlier
ones, bring in the relevant "context," and indicate
the importance of considering the outputs in rela-
tion to one another and in relation to the situation
in which they are produced. Further work in this
direction must confront issues like what is
involved in the coherence of extended conversa-
tions, dramatic plots, and scholarly essays. Very
little in the way of explicit theories or conditions
has been proposed in this area - which is sympto-
matic of an impoverished philosophical literature
on the problem of understanding and intelligibility
in general. (For a brief overview of some of the
difficulties, see Haugeland, 1979, and the works
cited there.)
Interpreting an automatic formal system is find-
ing a way of construing its outputs (assigning them
meanings) such that they consistently make rea-
sonable sense in the light of the system's prior
inputs and other outputs. In the special case of
logical and mathematical systems, it suffices if the
outputs are consistently true; and this can be guar-
anteed by having only true axioms and truth-
preserving rules. In more ambitious systems,
J. Haugeland
however, including any with aspirations to artifi-
cial intelligence, truth is not a sufficient condition
on the output tokens "making sense" - many other
considerations are important as well. Hence, there
is no reason to believe that truth-preservingness is
the only, or even the most important, requirement
to be imposed on the system's rules. Unfortun-
ately, the alternative - nonasininity-preservation,
perhaps? - is not at all clear. And the foregoing
motley list of amorphous "further conditions"
holds out little promise of any quick or clean
solution. I think that this problem constitutes one
of the deepest challenges that cognitive science
must face.
IX Cognitive Science (Again)
The basic idea of cognitive science is that intelli-
gent beings are semantic engines - in other words,
automatic formal systems with interpretations
under which they consistently make sense. We
can now see why this includes psychology and
artificial intelligence on a more or less equal foot-
ing: people and intelligent computers (if and when
there are any) turn out to be merely different
manifestations of the same underlying pheno-
menon. Moreover, with universal hardware, any
semantic engine can in principle be formally imit-
ated by a computer if only the right program can
be found. And that will guarantee semantic imita-
tion as well, since (given the appropriate formal
behavior) the semantics is "taking care of itself"
anyway. Thus we also see why, from this perspect-
ive, artificial intelligence can be regarded as psy-
chology in a particularly pure and abstract form.
The same fundamental structures are under inves-
tigation, but in AI, all the relevant parameters are
under direct experimental control (in the program-
ming), without any messy physiology or ethics to
get in the way.
Of course, it is possible that this is all wrong. It
might be that people just aren't semantic engines,
or even that no semantic engine (in a robot, say)
can be genuinely intelligent. There are two quite
different strategies for arguing that cognitive
science is basically misconceived. The first, or
hollow shell strategy has the following form: no
matter how well a (mere) semantic engine acts as
if it understands, etc., it can't really understand
anything, because it isn't (or hasn't got) "X" (for
some "X"). In other words, a robot based on a
semantic engine would still be a sham and a fake,
no matter how "good" it got. The other, or poor
substitute, strategy draws the line sooner: it denies
that (mere) semantic engines are capable even of
acting as if they understood - semantic engine
robots are not going to get that good in the first
place. The first strategy tends to be more concep-
tual and a priori, while the second depends more
on experimental results. (Compare: No beverage
made from coal tar would be wine, no matter what
it tasted like; with: There's no w:w to turn coal tar
into a beverage that tastes like wine.)
The most obvious candidate for "X" in the
hollow shell strategy is consciousness; thus "No
computer really understands anything, no matter
how smart it seems, because it isn't conscious."
Now it is true that cognitive science sheds virtually
no light on the issue of what consciousness is
(though see Dennett, 1978, for a valiant effort);
indeed, the term itself is almost a dirty word in the
technical literature. So it's natural to suspect that
something difficult and important is being left out.
Unfortunately, nobody else has anything very spe-
cific or explanatory to say about consciousness
either - it is just mysterious, regardless of your
point of view. But that means that a cognitivist
can say, "Look, none of us has much of an idea
of what consciousness is; so how can we be so
sure either that genuine understanding is
impossible without it, or that semantic engines
won't ever have it (e.g., when they are big and
sophisticated enough)?" Those questions may
seem intuitively perverse, but they are very
difficult to answer.
A different candidate for "X" is what we might
call original intentionality.
8
The idea is that a
semantic engine's tokens only have meaning
because we give it to them; their intentionality,
like that of smoke signals and writing, is essentially
borrowed, hence derivative. To put it bluntly:
computers themselves don't mean anything by
their tokens (any more than books do) - they
only mean what we say they do. Genuine under-
standing, on the other hand, is intentional "in its
own right" and not derivatively from something
else. But this raises a question similar to the last
one: What does it take to have original intention-
ality (and how do we know computers can't have
it)? If we set aside divine inspiration (and other
magical answers), it seems that original intention-
ality must depend on whether the object has a
suitable structure and/or dispositions, relative to
the environment. But it is hard to see how these
could fail to be suitable (whatever exactly that is) if
b'ect (semantic engine cum robot) always acts
the 0 J . d . .
II' nt self-motivate , responSIve to questions
tOte Ige , .
d challenges, and so on. A book, for mstance,
an . h "'" d h '
aYS no attentIOn to w at It says an. at s
p. easo
n
) whv we reallv do not thmk It IS the
(one r . .
k
"'hich is saving anything (but rather the
boo '"
h
r)
A perfect robot, however, would seem to
aut 0 .'
l
'tS opinions defend them from attack, and
act on '
modify them when confronted with counter-evid-
a
ll of which would suggest that they really
cncc -
rc the robot's own opinions.
a d'd C "X'" h h 11 h 11
A third can 1 ate lor m teo ow s e
strategy is caring. Here the intuition is that a
System could not really mean anything unless it
had a stake in what it was saying - unless its beliefs
and goals mattered to it. Otherwise, it is just
mouthing noises, or generating tokens mechanic-
allv. The popular picture of computers as cold
unfeeling) calculators motivates the
view that they could never really care about any-
thing, hence that they could never genuinely mean
(or understand) anything on their own. But, of
course, the legitimacy of this picture is just what
we're inquiring about. If cognitive science is on the
right track, then some semantic engines - starting
with people - can care about things, be involved
with them, have personalities, and so on. And,
indeed, if we had a robot which seemed (in appro-
priate circumstances) to be sympathetic, disap-
pointed, embarrassed, ambitious, offended,
affectionate, and so on, it would be very difficult
to claim that it was merely an unfeeling hunk of
metal - especially if this very remark "hurt its
Notes
There remain, of course, quantum indeterminacies,
even in physics; but these are no consolation to inter-
actionism. It is generally best, in fact, to forget about
quantum mechanics in any discussion of mind/body
metaphysics.
2 Which is not to say, by any means, that those accounts
themselves were stupid, or scientifically worthless.
3 The number of different types of tokens is not a
"deep" property of a formal system, as can be seen
from the fact that the peg-jumping game has only one
type, whereas the domino-switching game has 33
types - yet, in a deep sense, they are the same
game. In effect, the domino version trades more
sophisticated token discriminations for less sophistic-
ated position discriminations.
4 This is a slight generalization of Turing's original
definition (Turing, 1937). In his version, the storage
Semantic Engines
feelings." Again, we need some further criterion
or intuitive test for what it is to have the appro-
priate inner quality, if we are to justify saying that
a semantic engine which merely acts as if it IS
intelligent (etc.) is a hollow shell.
The "poor substitute" strategy - which I,
myself, think is much more likely to succeed -
argues instead that semantic engines will never
even seem to have the full range of common sense
and values of people. The basic suggestion is that
those areas in which computers excel (or can be
expected eventually to excel) are all of a special
sort, where the relevant considerations are com-
paratively few and well defined. This includes
formal games (by definition) and also a number
of other routine technical or micro-world tasks.
But it is an open question whether the intelligence
manifested in everyday life, not to mention art,
invention, and discovery, is of essentially this
same sort (though presumably more complicated).
Cognitive science, in effect, is betting that it is; but
the results are just not in yet. The above issues of
consciousness, original intentionality, and caring
can all be raised again in the poor substitute strat-
egy, in a more empirical form: Does the system in
fact act as if it had the relevant "X"? And if, so far,
the answer is always "No," is there any pattern to
the failures which might give us a clue to the
deeper nature of the problems, or the ultimate
outcome? These, it seems to me, are the most
important questions, but they are beyond the
scope of this introduction.
bins are all connected in a row, called a "tape"; each
bin can hold only a simple token; the in-bin and the
out-bin are always the same bin; and this in/out-bin
is always either the same as or right next to the in/
out-bin for the previous step. So Turing's machine
chugs back and forth along the tape, one bin at a time,
dealing with one simple token at a time. The surpris-
ing and important point, as we shall see shortly, is
that (apart from convenience and efficiency) these
differences don't make any difference.
S This is a way of expressing Church's thesis (named
after the logician, Alonzo Church); it has been proven
true for all known, well-defined kinds of automatic
formal system; but it cannot be proven in general, in
part because the general idea of "automatic formal
system" is itself somewhat intuitive, and not precisely
definable.
J. Haugeland
6 "Ve ignore the possibility of "randomizers" - they
don't affect the point.
i Unfortunately, even in mathematics, formalization is
not all that one might have hoped. Ideally, one would
like a system such that not only were all its theorems
true, but also all its true tokens were theorems (i.e.,
theorems were true); such a system is semantic-
al(v complete. But it has been shown (originally by
References
Austin, John L. (l9iO) How to Do Things with Words.
New York: Oxford University Press.
Davidson, Donald (l9i3) Radical interpretation. Dialec-
tica, 2i, 3l3-28.
Dennett, Daniel (19i8) Toward a cognitive theory of
consciousness. In C. Wade Savage (ed.), Perception
and Cognition: Issues in the Foundations of Psychology.
Minnesota Studies in the Philosophy of Science, vol.
9. Minneapolis: University of Minnesota Press. Re-
printed (19i8) in Brainstorms. Montgomery, VT:
Bradford Books.
Dennett, Daniel (1981) Three kinds of intentional psy-
chology. In R. A. Healey (ed.), Reduction, Time and
Studies in the Philosophy of the Natural
Sciences. Cambridge: Cambridge University Press.
Godel, Kurt (1931) Ober formal unentscheidbare Siitze
der Principia Mathematica und verwandter Systeme 1.
Monatshefte fur Mathematik und Physik, 38, li3-98.
Kurt Godel, in 1931) that no consistent formalization
of arithmetic can be complete; and the same applies
to many other important axiomatic systems. Most
people are agreed, however, that this result doesn't
make any difference to cognitive science. (For a poss-
ibly dissenting view, see Lucas, 1961.)
8 "Intentionality" is a philosopher's term for being
about something, or having a meaning.
Grice, H. Paul (19iS) Logic and conversation. In G.
Harman and D. Davidson (eds), The Logic of Gram-
mar. Encino, CA: Dickenson.
Haugeland, John (19i9) Understanding natural language.
Journal of Philosophy, i6, 619-32.
Hobbes, Thomas (1651) The Leviathan.
Kuhn, Thomas (1970) The Structure of Scientific Revolu-
tions, 2nd edn. Chicago: University of Chicago Press.
Lucas, J. R. (1961) Minds, machines and Giidel. PhIlo-
sophy, 36, 120-4.
Quine, W. V. O. (1960) Word and Object. Cambridge,
MA: MIT Press.
Searle, John R. (1969) Speech Acts: An Essay in the
Philosophy of Language. Cambridge: Cambridge Uni-
versity Press.
Turing, A. M. (1937) On computable numbers, with an
application to the Entscheidungsproblem. Proceedings of
the London Mathematical SOCiery, 42, 230-65.