=Paper=
{{Paper
|id=Vol-444/paper-13
|storemode=property
|title=How Analytic Philosophy Has Failed Cognitive Science
|pdfUrl=https://ceur-ws.org/Vol-444/paper13.pdf
|volume=Vol-444
}}
==How Analytic Philosophy Has Failed Cognitive Science==
https://ceur-ws.org/Vol-444/paper13.pdf
How Analytic Philosophy Has Failed Cognitive Science
Robert Brandom
I. Introduction
We analytic philosophers have signally failed our colleagues in cognitive science. We have done that by
not sharing central lessons about the nature of concepts, concept-use, and conceptual content that have been
entrusted to our care and feeding for more than a century.
I take it that analytic philosophy began with the birth of the new logic that Gottlob Frege introduced in
his seminal 1879 Begriffsschrift. The idea, taken up and championed to begin with by Bertrand Russell,
was that the fundamental insights and tools Frege made available there, and developed and deployed
through the 1890s, could be applied throughout philosophy to advance our understanding of understanding
and of thought in general, by advancing our understanding of concepts—including the particular concepts
with which the philosophical tradition had wrestled since its inception. For Frege brought about a
revolution not just in logic, but in semantics. He made possible for the first time a mathematical
characterization of meaning and conceptual content, and so of the structure of sapience itself. Henceforth it
was to be the business of the new movement of analytic philosophy to explore and amplify those ideas, to
exploit and apply them wherever they could do the most good. Those ideas are the cultural birthright,
heritage, and responsibility of analytic philosophers. But we have not done right by them. For we have
failed to communicate some of the most basic of those ideas, failed to explain their significance, failed to
make them available in forms usable by those working in allied disciplines who are also professionally
concerned to understand the nature of thought, minds, and reason.
Contemporary cognitive science is a house with many mansions. The provinces I mean particularly to
be addressing are cognitive psychology, developmental psychology, animal psychology (especially
primatology), and artificial intelligence. (To be sure, this is not all of cognitive science. But the points I will
be making in this paper are not of similarly immediate significance for such other subfields as
neurophysiology, linguistics, perceptual psychology, learning theory, and the study of the mechanisms of
memory.) Cognitive psychology aims at reverse-engineering the human mind: figuring out how we do what
we do, what more basic abilities are recruited and deployed (and how) so as to result in the higher cognitive
abilities we actually display. Developmental psychology investigates the sequence of stages by which those
abilities emerge from more primitive versions as individual humans mature. Animal psychology, as I am
construing it, is a sort of combination of cognitive psychology of non-human intelligences and a
phylogenetic version of ontogenetic human developmental psychology. By contrast to all these empirical
inquiries into actual cognition, artificial intelligence swings free of questions about how any actual
organisms do what they do, and asks instead what constellation of abilities of the sort we know how to
implement in artifacts might in principle yield sapience.
Each of these disciplines is in its own way concerned with the empirical question of how the trick of
cognition is or might be done. Philosophers are concerned with the normative question of what counts as
doing it—with what understanding, particularly discursive, conceptual understanding consists in, rather
than how creatures with a particular contingent constitution, history, and armamentarium of basic abilities
come to exhibit it. I think Frege taught us three fundamental lessons about the structure of concepts, and
hence about all possible abilities that deserve to count as concept-using abilities.1 The conclusion we
should draw from his discoveries is that concept-use is intrinsically stratified. It exhibits at least four basic
layers, with each capacity to deploy concepts in a more sophisticated sense of ‘concept’ structurally
presupposing the capacities to use concepts in all of the more primitive senses. The three lessons that
generate the structural hierarchy oblige us to distinguish between:
• concepts that only label and concepts that describe,
1
It ought to be uncontroversial that the last two of the three lessons are due to Frege. Whether he is responsible also for
the first is more contentious. Further, I think both it and a version of the second can be found already in Kant. (As I
argue in my 2006 Woodbridge Lectures, “Animating Ideas of Idealism: A Semantic Sonata in Kant and Hegel,”
forthcoming in the Journal of Philosophy.) But my aims here are not principally hermeneutical or exegetical—those
issues don’t affect the question of what we philosophers ought to be teaching cognitive scientists—so I will not be
concerned to justify these attributions.
• the content of concepts and the force of applying them, and
• concepts expressible already by simple predicates and concepts expressible only by complex
predicates.
AI researchers and cognitive, developmental, and animal psychologists need to take account of the
different grades of conceptual content made visible by these distinctions, both in order to be clear about the
topic they are investigating (if they are to tell us how the trick is done, they must be clear about exactly
which trick it is) and because the empirical and in-principle possibilities are constrained by the way the
abilities to deploy concepts in these various senses structurally presuppose the others that appear earlier in
the sequence. This is a point they have long appreciated on the side of basic syntactic complexity. But the
at least equally important—and I would argue more conceptually fundamental—hierarchy of semantic
complexity has been largely ignored.
II. First Distinction: From Labeling to Describing
The Early Modern philosophical tradition was built around a classificatory theory of consciousness and
(hence) of concepts, in part the result of what its scholastic predecessors had made of their central notion of
Aristotelian forms. The paradigmatic cognitive act is understood as classifying: taking something particular
as being of some general kind. Concepts are identified with those general kinds.
This conception was enshrined in the order of logical explanation (originating in Aristotle’s Prior
Analytics) that was common to everyone thinking about concepts and consciousness in the period leading
up to Kant. At its base is a doctrine of terms or concepts, particular and general. The next layer, erected on
that base, is a doctrine of judgments, describing the kinds of classificatory relations that are possible among
such terms. For instance, besides classifying Socrates as human, humans can be classified as mortal.
Finally, in terms of those metaclassifications grouping judgments into kinds according to the sorts of terms
they relate, a doctrine of consequences or syllogisms is propounded, classifying valid inferences into kinds,
depending on which classes of classificatory judgments their premises and conclusions fall under.
It is the master-idea of classification that gives this traditional order of explanation its distinctive shape.
That idea defines its base, the relation between its layers, and the theoretical aspiration that animates the
whole line of thought: finding suitable ways of classifying terms and judgments (classifiers and
classifications) so as to be able to classify inferences as good or bad solely in virtue of the kinds of
classifications they involve. The fundamental metaconceptual role it plays in structuring philosophical
thought about thought evidently made understanding the concept of classifying itself a particularly urgent
philosophical task. Besides asking what differentiates various kinds of classifying, we can ask what they
have in common. What is it one must do in order thereby to count as classifying something as being of
some kind?
In the most general sense, one classifies something simply by responding to it differentially. Stimuli are
grouped into kinds by the response-kinds they tend to elicit. In this sense, a chunk of iron classifies its
environments into kinds by rusting in some of them and not others, increasing or decreasing its
temperature, shattering or remaining intact. As is evident from this example, if classifying is just exercising
a reliable differential responsive disposition, it is a ubiquitous feature of the inanimate world. For that very
reason, classifying in this generic sense is not an attractive candidate for identification with conceptual,
cognitive, or conscious activity. It doesn’t draw the right line between thinking and all sorts of thoughtless
activities. Pan-psychism is too high a price to pay for cognitive naturalism.
That need not mean that taking differential responsiveness as the genus of which conceptual
classification is a species is a bad idea, however. A favorite idea of the classical British empiricists was to
require that the classifying response be entering a sentient state. The intrinsic characters of these sentient
states are supposed to sort them immediately into repeatable kinds. These are called on to function as the
particular terms in the base level of the neo-Aristotelian logical hierarchy. General terms or concepts are
then thought of as sentient state-kinds derived from the particular sentient state-kinds by a process of
abstraction: grouping the base-level sentient state-repeatables into higher-level sentient state-repeatables by
some sort of perceived similarity. This abstractive grouping by similarity is itself a kind of classification.
The result is a path from one sort of consciousness, sentience, to a conception of another sort of
consciousness, sapience, or conceptual consciousness.
A standing felt difficulty with this empiricist strategy is the problem of giving a suitably naturalistic
account of the notion of sentient awareness on which it relies. Recent information-theoretic accounts of
representation (under which heading I include not just Fred Dretske’s theory, which actually goes by that
name, but others such as Jerry Fodor’s asymmetric counterfactual dependence and nomological locking
models2) develop the same basic differential responsiveness version of the classic classificatory idea in
wholly naturalistic modal terms. They focus on the information conveyed about stimuli—the way they are
grouped into repeatables—by their reliably eliciting a response of one rather than another repeatable
response-kind from some system. In this setting, unpalatable pan-psychism can be avoided not, as with
traditional empiricism, by insisting that the responses be sentient states, but for instance by restricting
attention to flexible systems, capable in principle of coming to encode many different groupings of stimuli,
with a process of learning determining what classificatory dispositions each one actually acquires. (The
classical American pragmatists’ program for a naturalistic empiricism had at its core the idea that the
structure common to evolutionary development and individual learning is a Test-Operate-Test-Exit
negative feedback process of acquiring practical habits, including discriminative ones.3)
Classification as the exercise of reliable differential responsive dispositions (however acquired) is not
by itself yet a good candidate for conceptual classification, in the basic sense in which applying a concept
to something is describing it. Why not? Suppose one were given a wand, and told that the light on the
handle would go on if and only if what the wand was pointed at had the property of being grivey. One
might then determine empirically that speakers are grivey, but microphones not, doorknobs are but
windowshades are not, cats are and dogs are not, and so on. One is then in a position reliably, perhaps even
infallibly, to apply the label ‘grivey’. Is one also in a position to describe things as grivey? Ought what one
is doing to qualify as applying the concept grivey to things? Intuitively, the trouble is that one does not
know what one has found out when one has found out that something is grivey, does not know what one is
taking it to be when one takes it to be grivey, does not know what one is describing it as. The label is, we
want to say, uninformative.
What more is required? Wilfrid Sellars gives this succinct, and I believe correct, answer:
It is only because the expressions in terms of which we describe objects, even such basic ex-
pressions as words for the perceptible characteristics of molar objects, locate these objects in a
space of implications, that they describe at all, rather than merely label.4
The reason ‘grivey’ is merely a label, that it classifies without informing, is that nothing follows from so
classifying an object. If I discover that all the boxes in the attic I am charged with cleaning out have been
labeled with red, yellow, or green stickers, all I learn is that those labeled with the same color share some
property. To learn what they mean is to learn, for instance, that the owner put a red label on boxes to be
discarded, green on those to be retained, and yellow on those that needed further sorting and decision. Once
I know what follows from affixing one rather than another label, I can understand them not as mere labels,
but as descriptions of the boxes to which they are applied. Description is classification with consequences,
either immediately practical (“to be discarded/examined/kept”) or for further classifications.
Michael Dummett argues generally that to be understood as conceptually contentful, expressions must
have not only circumstances of appropriate application, but also appropriate consequences of application.5
That is, one must look not only upstream, to the circumstances (inferential and non-inferential) in which it
is appropriate to apply the expression, but also downstream to the consequences (inferential and non-
inferential) of doing so, in order to grasp the content it expresses. One-sided theories of meaning, which
seize on one aspect to the exclusion of the other, are bound to be defective, for they omit aspects of the use
that are essential to meaning. For instance, expressions can have the same circumstances of application, and
different consequences of application. When they do, they will have different descriptive content.
2
Dretske, Fred: Knowledge and the Flow of Information (MIT Press—Bradford, 1981), Fodor, Jerry: A Theory of
Content (MIT Press—Bradford, 1990).
3
I sketch this program in the opening section of "The Pragmatist Enlightenment (and its Problematic Semantics)"
European Journal of Philosophy, Vol 12 No 1, April 2004, pp. 1-16.
4
Pp. 306-307 (§107) in: Wilfrid Sellars: “Counterfactuals, Dispositions, and Causal Modalities” In Minnesota Studies
in the Philosophy of Science, Volume II: Concepts, Theories, and the Mind-Body Problem, ed. Herbert Feigl, Michael
Scriven, and Grover Maxwell (Minneapolis: University of Minnesota Press, 1958), p.225-308.
5
I discuss this view of Dummett’s (from his Frege: Philosophy of Language second edition [Harvard University Press
1993], originally published in 1974), at greater length in Chapter Two of Making It Explicit [Harvard University Press,
1994], and Chapter One of Articulating Reasons [Harvard University Press, 2000].
1] I will write a book about Hegel,
and
2] I foresee that I will write a book about Hegel,
say different things about the world, describe it as being different ways. The first describes my future
activity and accomplishment, the second my present aspiration. Yet the circumstances under which it is
appropriate or warranted to assert them—the situations to which I ought reliably to respond by endorsing
them—are the same (or at least, can be made so by light regimentation of a prediction-expressing use of
‘foresee’). Here, to say that they have different descriptive content can be put by saying that they have
different truth conditions. (That they have the same assertibility conditions just shows how assertibility
theories of meaning, as one-sided in Dummett’s sense, go wrong.) But that same fact shows up in the
different positions they occupy in the “space of implications.” For from the former it follows that I will not
be immediately struck by lightning, that I will write some book, and, indeed, that I will write a book about
Hegel. None of these is in the same sense a consequence of the second claim.
We might train a parrot reliably to respond differentially to the visible presence of red things by
squawking “That’s red.” It would not yet be describing things as red, would not be applying the concept
red to them, because the noise it makes has no significance for it. It does not know that it follows from
something’s being red that it is colored, that it cannot be wholly green, and so on. Ignorant as it is of those
inferential consequences, the parrot does not grasp the concept (any more than we express a concept by
‘grivey’). The lesson is that even observational concepts, whose principal circumstances of appropriate
application are non-inferential (a matter of reliable dispositions to respond differentially to non-linguistic
stimuli) must have inferential consequences in order to make possible description, as opposed to the sort of
classification effected by non-conceptual labels.
The rationalist idea that the inferential significance of a state or expression is essential to its conceptual
contentfulness is one of the central insights of Frege’s 1879 Begriffsschrift (“concept writing”)—the
founding document of modern logic and semantics—and is appealed to by him in the opening paragraphs
to define his topic:
...there are two ways in which the content of two judgments may differ; it may, or it may not, be the case
that all inferences that can be drawn from the first judgment when combined with certain other ones can
always also be drawn from the second when combined with the same other judgments…I call that part of
the content that is the same in both the conceptual content [begriffliche Inhalt]. 6
Here, then, is the first lesson that analytic philosophy ought to have taught cognitive science: there is a
fundamental conceptual distinction between classification in the sense of labeling and classification in the
sense of describing, and it consists in the inferential consequences of the classification: its capacity to serve
as a premise in inferences (practical or theoretical) to further conclusions. (Indeed, there are descriptive
concepts that are purely theoretical—such as gene and quark—in the sense that in addition to their
inferential consequences of application, they have only inferential circumstances of application.) There is
probably no point in fighting over the minimal circumstances of application of the concepts concept and
conceptual. Those who wish to lower the bar sufficiently are welcome to consider purely classificatory
labels as a kind of concept (perhaps so as not to be beastly to the beasts, or disqualify human infants, bits of
our brains, or even some relatively complex computer programs wholly from engaging in conceptually
articulated activities). But if they do so, they must not combine those circumstances of application with the
consequences of application appropriate to genuinely descriptive concepts—those that do come with
inferential significances downstream from their application.
Notice that this distinction between labeling and describing is untouched by two sorts of elaborations of
the notion of labeling that have often been taken to be of great significance in thinking about concepts from
the classical classificatory point of view. One does not cross the boundary from labeling to describing just
because the reliable capacity to respond differentially is learned, and in that sense flexible, rather than
6
Frege, Begriffsschrift (hereafter BGS), section 3. The passage continues: “In my formalized language
[Begriffsschrift]...only that part of judgments which affects the possible inferences is taken into consideration.
Whatever is needed for a correct inference is fully expressed; what is not needed is...not.”
innate, and in that sense rigid. And one is likewise developing the classical model in an orthogonal
direction insofar as one focuses on the metacapacity to learn to distinguish arbitrary Boolean combinations
of microfeatures one can already reliably discriminate. From the point of view of the distinction between
labeling and describing, that is not yet the capacity to form concepts, but only the mastery of compound
labels. That sort of structural articulation upstream has no semantic import at the level of description until
and unless it is accorded a corresponding inferential significance downstream.
III. Ingredient vs. Free-Standing Content: Semantically Separating Content from
Force
Once our attention has been directed at the significance of applying a classifying concept—downstream,
at the consequences of applying it, rather than just upstream, at the repeatable it discriminates, the grouping
it institutes—so that mere classification is properly distinguished from descriptive classification, the
necessity of distinguishing different kinds of consequence becomes apparent. One distinction in the
vicinity, which has already been mentioned in passing, is that between practical and theoretical (or, better,
cognitive) consequences of application of a concept. The significance of classifying an object by
responding to it one way rather than another may be to make it appropriate to do something else with or to
it—to keep it, examine it, or throw it away, to flee or pursue or consume it, for example. This is still a
matter of inference; in this case, it is practical inferences that are at issue. But an initial classification may
also contribute to further classifications: that what is in my hand falls under both the classifications
raspberry and red makes it appropriate to classify it also as ripe—which in turn has practical consequences
of application (such as, under the right circumstances “falling to without further ado and eating it up,” as
Hegel says in another connection) that neither of the other classifications has individually. Important as the
distinction between practical and cognitive inferential consequences is, in the present context there is
reason to emphasize a different one.
Discursive intentional phenomena (and their associated concepts), such as assertion, inference,
judgment, experience, representation, perception, action, endorsement, and imagination typically involve
what Sellars calls “the notorious ‘ing’/‘ed’ ambiguity.” For under these headings we may be talking about
the act of asserting, inferring, judging, experiencing, representing, perceiving, doing, endorsing, and
imagining, or we may be talking about the content that is asserted, inferred, judged, experienced,
represented, perceived, done, endorsed, or imagined. ‘Description’ is one of these ambiguous terms (as is
‘classification’). We ought to be aware of the distinction between the act of describing (or classifying),
applying a concept, on the one hand, and the content of the description (classification, concept)—how
things are described (classified, conceived)—on the other. And the distinction is not merely of theoretical
importance for those of us thinking systematically about concept use. A distinctive level of conceptual
sophistication is achieved by concept users that themselves distinguish between the contents of their
concepts and their activity of applying them. So one thing we might want to know about a system being
studied, a non-human animal, a prelinguistic human, an artifact we are building, is whether it distinguishes
between the concept it applies and what it does by applying it.
We can see a basic version of the distinction between semantic content and pragmatic force as in play
wherever different kinds of practical significance can be invested in the same descriptive content (different
sorts of speech act or mental act performed using that content). Thus if a creature can not only say or think
that the door is shut, but also ask or wonder whether the door is shut, or order or request that it be shut, we
can see it as distinguishing in practice between the content being expressed and the pragmatic force being
attached to it. In effect, it can use descriptive contents to do more than merely describe. But this sort of
practical distinguishing of pragmatic from semantic components matters for the semantic hierarchy I am
describing only when it is incorporated or reflected in the concepts (that is, the contents) a creature can
deploy. The capacity to attach different sorts of pragmatic force to the same semantic content is not
sufficient for this advance in structural semantic complexity. (Whether it is a necessary condition is a
question I will not address—though I am inclined to think that in principle the answer is ‘No’.)
For the inferential consequences of applying a classificatory concept, when doing that is describing and
not merely labeling, can be either semantic consequences, which turn on the content of the concept being
applied, or pragmatic consequences, which turn on the act one is performing in applying it. Suppose John
issues an observation report: “The traffic light is red.” You may infer that it is operating and illuminated,
and that traffic ought to stop in the direction it governs. You may also infer that John has a visually
unobstructed line of sight to the light, notices what color it is, and believes that it is red. Unlike the former
inferences, these are not inferences from what John said, from the content of his utterance, from the
concepts he has applied. They are inferences from his saying it, from the pragmatic force or significance of
his uttering it, from the fact of his applying those concepts. For what he has said, that the traffic light is red,
could be true even if John had not been in a position to notice it or form any beliefs about it. Nothing about
John follows just from the color of the traffic light.7
It can be controversial whether a particular consequence follows from how something is described or
from describing it that way, that is, whether that consequence is part of the descriptive content of an
expression, the concept applied, or stems rather from the force of using the expression, from applying the
concept. A famous example is expressivist theories of evaluative terms such as ‘good’. In their most
extreme form, they claim that these terms have no descriptive content. All their consequences stem from
what one is doing in using them: commending, endorsing, or approving. In his lapidary article
“Ascriptivism,”8 Peter Geach asks what the rules governing this move are. He offers the archaic term
‘macarize’, meaning to characterize someone as happy. Should we say that in apparently describing
someone as happy we are not really describing anyone, but rather performing the distinctive speech act of
macarizing? But why not then discern distinctive speech acts for any apparently descriptive term?
What is wanted is a criterion for distinguishing semantic from pragmatic consequences, those that stem
from the content of the concept being applied from those that stem from what we are doing in applying that
concept (using an expression to perform a speech act). Geach finds one in Frege, who in turn was
developing a point made already by Kant.9 The logical tradition Kant inherited was built around the
classificatory theory of consciousness we began by considering. Judgment was understood as classification
or predication: paradigmatically, of something particular as something general. But we have put ourselves
in a position to ask: is this intended as a model of judgeable contents are constructed, or of what one is
doing in judging? Kant saw, as Frege would see after him, that the phenomenon of compound judgments
shows that it cannot play both roles. For consider the hypothetical or conditional judgment
3] If Frege is correct, then conceptual content depends on inferential consequences.
In asserting this sentence (endorsing its content), have I predicated correctness of Frege (classified him
as correct)? Have I described him as correct? Have I applied the concept of correctness? If so, then
predicating or classifying (or describing) is not judging. For in asserting the conditional I have not judged
or asserted that Frege is correct. I have at most built up a judgeable content, the antecedent of the
conditional, by predication. For embedding a declarative descriptive sentence as an unasserted component
in a compound asserted sentence strips off the pragmatic force its free-standing, unembedded occurrence
would otherwise have had. It now contributes only its content to the content of the compound sentence, to
which alone the pragmatic force of a speech act is attached.
This means that embedding simpler sentences as components of compound sentences—
paradigmatically, embedding them as antecedents of conditionals—is the way to discriminate consequences
that derive from the content of a sentence from consequences that derive from the act of asserting or
endorsing it. We can tell that ‘happy’ does express descriptive content, and is not simply an indicator that
some utterance has the pragmatic force or significance of macarizing, because we can say things like:
4] If she is happy, then John should be glad.
7
One might think that a similar distinction could be made concerning a parrot that merely reliably responsively
discriminated red things by squawking “That’s red.” For when he does that, one might infer that there was something
red there (since he is reliable), and one might also infer that the light was good and his line of sight unobstructed. So
both sorts of inference seem possible in this case. But it would be a mistake to describe the situation in these terms. The
squawk is a label, not a description. We infer from the parrot’s producing it that there is something red, because the two
sorts of events are reliably correlated, just as we would from the activation of a photocell tuned to detect the right
electromagnetic frequencies. By contrast, John offers testimony. What he says is usable as a premise in our own
inferences, not just the fact that his saying it is reliably correlated with the situation he (but not the parrot) reports
(though they both respond to it).
8
The Philosophical Review, Vol. 69, No. 2, 221-225. Apr., 1960.
9
I discuss this point further in the first lecture of “Animating Ideas of Idealism” [op.cit.].
For in asserting that, one does not macarize anyone. So the consequence, that John should be glad, must
be due to the descriptive content of the antecedent, not to its force.
Similarly, Geach argues that the fact that we can say things like:
5] If being trustworthy is good, then you have reason to be trustworthy,
shows that ‘good’ does have descriptive content.10 Notice that this same test appropriately discriminates the
different descriptive contents of the claims:
6] Labeling is not describing,
and
7] I believe that labeling is not describing.
For the two do not behave the same way as antecedents of conditionals. The stuttering inference
8] If labeling is not describing, then labeling is not describing,
is as solid an inference as one could ask for. The corresponding conditional
9] If I believe that labeling is not describing, then labeling is not describing,
requires a good deal more faith to endorse. And in the same way, the embedding test distinguishes [1] and
[2] above. In each case it tells us, properly, that different descriptive contents are involved.
What all this means is that any user of descriptive concepts who can also form compound sentences,
paradigmatically conditionals, is in a position to distinguish what pertains to the semantic content of those
descriptive concepts from what pertains to the act or pragmatic force of describing by applying those
concepts. This capacity is a new, higher, more sophisticated level of concept use. It can be achieved only by
looking at compound sentences in which other descriptive sentences can occur as unasserted components.
For instance, it is only in such a context that one can distinguish denial (a kind of speech act or attitude)
from negation (a kind of content). One who asserts [6] has both denied that labeling is describing, and
negated a description. But one who asserts conditionals such as [8] and [9] has negated descriptions, but
has not denied anything.
The modern philosophical tradition up to Frege took it for granted that there was an special attitude on
could adopt towards a descriptive conceptual content, a kind of minimal force one could invest it with, that
must be possible independently of and antecedent to being able to endorse that content in a judgment. This
is the attitude of merely entertaining the description. The picture (for instance, in Descartes) was that first
one entertained descriptive thoughts (judgeables), and then, by an in-principle subsequent act of will,
accepted or rejected it. Frege rejects this picture. The principal—and in principle fundamental—pragmatic
attitude (and hence speech act) is judging or endorsing.11 The capacity merely to entertain a proposition
(judgeable content, description) is a late-coming capacity—one that is parasitic on the capacity to endorse
such contents. In fact, for Frege, the capacity to entertain (without endorsement) the proposition that p is
just the capacity to endorse conditionals in which that proposition occurs as antecedent or consequent. For
that is to explore its descriptive content, its inferential circumstances and consequences of application, what
it follows from and what follows from it, what would make it true and what would be true if it were true,
without endorsing it. This is a new kind of distanced attitude toward one’s concepts and their contents—
one that becomes possible only in virtue of the capacity to form compound sentences of the kind of which
10
Of course, contemporary expressivists such as Gibbard and Blackburn (who are distinguished from emotivist
predecessors such as C.L. Stevenson precisely by their appreciation of the force of the Frege-Geach argument) argue
that it need not follow that the right way to understand that descriptive content is not by tracing it back to the attitudes
of endorsement or approval that are expressed by the use of the expression in free-standing, unembedded assertions.
11
In the first essay of “Animating Ideas of Idealism” [op.cit.] I discuss the line of thought that led Kant to give pride of
place to judgment and judging.
conditionals are the paradigm. It is a new level of cognitive achievement—not in the sense of a new kind of
empirical knowledge (though conditionals can indeed codify new empirical discoveries), but of a new kind
of semantic self-consciousness.
Conditionals make possible a new sort of hypothetical thought. (Supposing that postulating a distinct
attitude of supposing would enable one to do this work, the work of conditionals, would be making the
same mistake as thinking that denial can do the work of negation.) Descriptive concepts bring empirical
properties into view. Embedding those concepts in conditionals brings the contents of those concepts into
view. Creatures that can do that are functioning at a higher cognitive and conceptual level than those who
can only apply descriptive concepts, just as those who can do that are functioning at a higher cognitive and
conceptual level than those who can only classify things by reliable responsive discrimination (that is,
labeling). That fact sets a question for the different branches of cognitive science I mentioned in my
introduction. Can chimps, or African grey parrots, or other non-human animals not just use concepts to
describe things, but also semantically discriminate the contents of those concepts from the force of
applying them, by using them not just in describing, but in conditionals, in which their contents are merely
entertained and explored? At what age, and along with what other capacities, do human children learn to do
so? What is required for a computer to demonstrate this level of cognitive functioning?
Conditionals are special, because they make inferences explicit—that is, put them into endorsable,
judgeable, assertible, which is to say propositional form. And it is their role in inferences, we saw, that
distinguishes descriptive concepts from mere classifying labels. But conditionals are an instance of a more
general phenomenon. For we can think of them as operators, which apply to sentences to yield further
sentences. As such, they bring into view a new notion of conceptual content: a new principle of
assimilation, hence classification, of such contents. For we begin with the idea of sameness of content that
derives from sameness of pragmatic force, attitude, or speech act. But the Frege-Geach argument shows
that we can also individuate conceptual contents more finely, not just in terms of their role in free-standing
utterances, but also accordingly as substituting one for another as arguments of operators (paradigmatically
the conditional) does or does not yield compound sentences with the same free-standing pragmatic
significance or force. Dummett calls these notions “free-standing” and “ingredient” content (or sense),
respectively. Thus we might think that
10] It is nice here,
and
11] It is nice where I am,
express the same attitude, perform the same speech act, have the same pragmatic force or significance.
They not only have the same circumstances of application, but the same consequences of application (and
hence role as antecedents of conditionals). But we can see that they have different ingredient contents by
seeing that they behave differently as arguments when we apply another operator to them. To use an
example of Dummett’s,
12] It is always nice here,
and
13] It is always nice where I am,
have very different circumstances and consequences of application, different pragmatic significances, and
do behave differently as the antecedents of conditionals. But this difference in content, this sense of
“different content” in which they patently do have different contents, is one that shows up only in the
context of compounding operators, which apply to sentences and yield further sentences. The capacity to
deploy such operators to form new conceptual (descriptive) contents from old ones accordingly ushers in a
new level of cognitive and conceptual functioning.
Creatures that can not merely label, but describe are rational, in the minimal sense that they are able to
treat one classification as providing a reason for or against another. If they can use conditionals, they can
distinguish inferences that depend on the content of the concept they are applying from those that depend
on what they are doing in classifying something as falling under that concept. But the capacity to use
conditionals gives them more than just that ability. For conditionals let them say what is a reason for what,
say that an inference is a good one. And for anyone who can do that, the capacity not just to deny that a
classification is appropriate, but to use a negation operator to form new classificatory contents means
brings with it the capacity to say that two classifications (classifiers, concepts) are incompatible: that one
provides a reason to withhold the other. Creatures that can use this sort of sentential compounding operator
are not just rational, but logical creatures. They are capable of a distinctive kind of conceptual self-
consciousness. For they can describe the rational relations that make their classifications into descriptions
in the first place, hence be conscious or aware of them in the sense in which descriptive concepts allow
them to be aware of empirical features of their world.
IV. Simple versus Complex Predicates
There is still a higher level of structural complexity of concepts and concept use. I have claimed that
Frege should be credited with appreciating both of the points I have made so far: that descriptive
conceptual classification beyond mere discriminative labeling depends on the inferential significance of the
concepts, and that semantically distinguishing the inferential significance of the contents of concepts from
that of the force of applying them depends on forming sentential compounds (paradigmatically
conditionals) in which other sentences appear as components. In each of these insights Frege had
predecessors. Leibniz (in his New Essay on the Human Understanding) had already argued the first point,
against Locke. (The move from thinking of concepts exclusively as reliably differentially elicited labels to
thinking of them as having to stand in the sort of inferential relations to one another necessary for them to
have genuine descriptive content is characteristic of the advance from empiricism to rationalism.) And
Kant, we have seen, appreciated how attention to compound sentences (including “hypotheticals”) requires
substantially amending the traditional classificatory theory of conceptual consciousness. The final
distinction I will discuss, that between simple and complex predicates, and the corresponding kinds of
concepts they express, is Frege’s alone. No-one before him (and embarrassingly few even of his admirers
after him) grasped this idea.
Frege’s most famous achievement is transforming traditional logic by giving us a systematic way to
express and control the inferential roles of quantificationally complex sentences. Frege could, as the whole
logical tradition from Aristotle down to his time (fixated as it was on syllogisms) could not, handle iterated
quantifiers. So he could, for instance, explain why
14] If someone is loved by everyone, then everyone loves someone,
is true (a conditional that codifies a correct inference), but
15] If everyone loves someone, then someone is loved by everyone,
is not. What is less appreciated is that in order to specify the inferences involving arbitrarily nested
quantifiers (‘some’ and ‘every’), he needed to introduce a new kind of predicate, and hence discern a
structurally new kind of concept.
Our first grip on the notion of a predicate is as a component of sentences. In artificial languages we
combine, for instance, a two-place predicate ‘P’ with two individual constants ‘a’ and ‘b’ to form the
sentence ‘Pab’. Logically minded philosophers of language use this model to think about the corresponding
sentences of natural languages, understanding
16] Kant admired Rousseau,
as formed by applying the two-place predicate ‘admired’ to the singular terms ‘Kant’ and ‘Rousseau’. The
kind of inferences that are made explicit by quantified conditionals—inferences that essentially depend on
the contents of the predicates involved—though, require us also to distinguish a one-place predicate, related
to but distinct from this two-place one, that is exhibited by
17] Rousseau admired Rousseau,
and
18] Kant admired Kant,
but not by [16].
19] Someone admired himself,
that is, something of the form ∃x[Pxx], follows from [17] and [18], but not from [16]. The property of
being a self-admirer differs from that of being an admirer and from that of being admired (even though it
entails both).
But there is no part of the sentences [17] and [18] that they share with each other that they don’t share
also with [16]. Looking just at the sub-sentential expressions out of which the sentences are built does not
reveal the respect of similarity that distinguishes self-admiration from admiration in general—a respect of
similarity that is crucial to understanding why the conditional
20] If someone admires himself then someone admires someone,
(∃x[Pxx]∃x∃y[Pxy]) expresses a good inference, while
21] If someone admires someone then someone admires himself,
(∃x∃y[Pxy] ∃x[Pxx]) does not. For what [17] and [18] share that distinguishes them from [16] is not a
component, but a pattern. More specifically, it is a pattern of cross-identification of the singular terms that
two-place predicate applies to.
The repeatable expression-kind ‘admires’ is a simple predicate. It occurs as a component in sentences
built up by concatenating it appropriately with a pair of singular terms. ‘x admires x’ is a complex
predicate.12 A number of different complex predicates are associated with any multi-place simple predicate.
So the three-place simple predicate used to form the sentence
22] John enjoys music recorded by Mark and books recommended by Bob,
generates not only a three-place complex predicate of the form Rxyz, but also two-place complex
predicates of the form Rxxy, Rxyy, and Rxyx, as well as the one-place complex predicate Rxxx. The
complex predicates can be thought of as patterns that can be exhibited by sentences formed using the
simple predicate, or as equivalence classes of such sentences. Thus the complex self-admiration predicate
can be thought of either as the pattern, rather than the part, that is common to all the sentences {“Rousseau
admired Rousseau,” “Kant admired Kant,” “Caesar admired Caesar,” “Brutus admired Brutus,” “Napoleon
admired Napoleon,”…}, or just as that set itself. Any member of such an equivalence class of sentences
sharing a complex predicate can be turned into any other by a sequence of substitutions of all occurrences
of one singular term by occurrences of another.
Substitution is a kind of decomposition of sentences (including compound ones formed using sentential
operators such as conditionals). After sentences have been built up using simple components (singular
terms, simple predicates, sentential operators), they can be assembled into equivalence classes (patterns can
be discerned among them) by regarding some of the elements as systematically replaceable by others. This
is the same procedure of noting invariance under substitution that we saw applies to the notion of free-
standing content to give rise to that of ingredient content, when the operators apply only to whole
sentences. Frege called what is invariant under substitution of some sentential components for others a
‘function’. A function can be applied to some arguments to yield a value, but it is not a part of the value it
yields. (One can apply the function capital of to Sweden to yield the value Stockholm, but neither Sweden
nor capital of is part of Stockholm.) He tied himself in some metaphysical knots trying to find a clear way
12
This point, and the terminology of ‘simple’ and ‘complex’ predicates, is due to Dummett, in the second chapter of his
monumental Frege’s Philosophy of Language [op.cit.].
of contrasting functions with things (objects). But two points emerge clearly. First, discerning the
substitutional relations among different sentences sharing the same simple predicate is crucial for
characterizing a wide range of inferential patterns. Second, those inferential patterns articulate the contents
of a whole new class of concepts.
Sentential compounding already provided the means to build new concepts out of old ones. The
Boolean connectives—conjunction, disjunction, negation, and the conditional definable in terms of them
(AB if and only if ~(A&~B))—permit the combination of predicates in all the ways representable by
Venn diagrams, corresponding to the intersection, union, complementation, and inclusion of sets (concept
extensions, represented by regions), and so the expression of new concepts formed from old ones by these
operations. But there is a crucial class of new concepts formable from the old ones that are not generable by
such procedures. One cannot, for instance, form the concept of a C such that for every A there is a B that
stands to that C in the relation R. This is the complex one-place predicate logicians would represent as
having the form {x: Cx & ∀y∈A∃z∈B[Rxz]}. As Frege says, such a concept cannot, as the Boolean ones
can, be formed simply by putting together pieces of the boundaries of the concepts A,B, and C. The
correlations of elements of these sets that concepts like these, those expressed by complex predicates,
depend on, and so the inferences they are involved in, cannot be represented in Venn diagrams.
Frege showed further that it is just concepts like these that even the simplest mathematics works with.
The concept of a natural number is the concept of a set every element of which has a successor. That is, for
every number, there is another related to it as a successor (∀x∃y[Successor(x,y)). The decisive advance that
Frege’s new quantificational logic made over traditional logic is a semantic, expressive advance. His
logical notation can, as the traditional logic could not, form complex predicates, and so both express a
vitally important kind of concept, and logically codify the inferences that articulate its descriptive content.
Complex concepts can be thought of as formed by a four-stage process.
• First, put together simple predicates and singular terms, to form a set of sentences, say
{Rab,Sbc,Tacd}.
• Then apply sentential compounding operators to form more complex sentences, say
{RabSbc, Sbc&Tacd}.
• Then substitute variables for some of the singular terms (individual constants), to form
complex predicates, say {RaxSxy, Sxy&Tayz}.
• Finally, apply quantifiers to bind some of these variables, to form new complex predicates, for
instance the one-place predicates (in y and z) {∃x[RaxSxy], ∀x∃y[Sxy&Tayz]}.
If one likes, this process can now be repeated, with the complex predicates just formed playing the role
that simple predicates originally played at the first stage, yielding the new sentences {∃x[RaxSxd],
∀x∃y[Sxy&Taya]}. They can then be conjoined, and the individual constant a substituted for to yield the
further one-place complex predicate (in z) ∃x[RzxSxd]&∀x∃y[Sxy&Tzyz]. We can use these procedures
to build to the sky, repeating these stages of concept construction as often as we like. Frege’s rules tell us
how to compute the inferential roles of the concepts formed at each stage, on the basis of the inferential
roles of the raw materials, and the operations applied at that stage. This is the heaven of concept formation
he opened up for us.
V. Conclusion
The result of all these considerations, which have been in play since the dawn of analytic philosophy,
well over a century ago, is a four-stage semantic hierarchy of ever more demanding senses of “concept”
and “concept use.” At the bottom are concepts as reliably differentially applied, possibly learned, labels or
classifications. Crudely behaviorist psychological theories (such as B. F. Skinner’s) attempted to do all
their explanatory work with responsive discriminations of this sort. At the next level, concepts as
descriptions emerge when merely classifying concepts come to stand in inferential, evidential, justificatory
relations to one another—when the propriety of one sort of classification has the practical significance of
making others appropriate or inappropriate, in the sense of serving as reasons for them. Concepts of this
sort may still all have observational uses, even though they are distinguished from labels by also having
inferential ones.13 Already at this level, the possibility exists of empirical descriptive concepts that can only
be properly applied as the result of inferences from the applicability of others. These are theoretical
concepts: a particularly sophisticated species of the genus of descriptive concepts.
At this second level, conceptual content first takes a distinctive propositional form; applications of this
sort of concept are accordingly appropriately expressed using declarative sentences. For the propositional
contents such sentences express just are whatever can play the role of premise and conclusion in inferences.
And it is precisely being able to play those roles that distinguishes applications of descriptive concepts
from applications of merely classificatory ones. Building on the capacity to use inferentially articulated
descriptive concepts to make propositionally contentful judgments or claims, the capacity to form sentential
compounds—paradigmatically conditionals, which make endorsements of material inferences relating
descriptive concept applications propositionally explicit, and negations, which make endorsements of
material incompatibilities relating descriptive concept applications propositionally explicit—brings with it
the capacity to deploy a further, more sophisticated, kind of conceptual content: ingredient (as opposed to
free-standing) content. Conceptual content of this sort is to be understood in terms of the contribution it
makes to the content of compound judgments in which it occurs, and only thereby, indirectly, to the force
or pragmatic significance of endorsing that content.
Ingredient conceptual content, then, is what can be negated, or conditionalized. The distinctive sort of
definiteness and determinateness characteristic of this sort of conceptual content becomes vivid when it is
contrasted with contents that cannot appear in such sentential compounds. My young son once complained
about a park sign consisting of the silhouette of what looked like a Scottish terrier, surrounded by a red
circle, with a slash through it. Familiar with the force of prohibition associated with signs of this general
form, he wanted to know: “Does this mean ‘No Scotties allowed’? Or ‘No dogs allowed’? Or ‘No animals
allowed’? Or ‘No pets allowed’”? Indeed. A creature that can understand a claim like “If the red light is on,
then there is a biscuit in the drawer,” without disagreeing when the light is not on, or immediately looking
for the biscuit regardless of how it is with the light, has learned to distinguish between the content of
descriptive concepts and the force of applying them, and as a result can entertain and explore those
concepts and their connections with each other without necessarily applying them in the sense of endorsing
their applicability to anything present. The capacity in this way to free oneself from the bonds of the here-
and-now is a distinctive kind of conceptual achievement
The first step was from merely discriminating classification to rational classification (‘rational’ because
inferentially articulated, according to which classifications provide reasons for others). The second step is
to synthetic logical concept formation, in which concepts are formed by logical compounding operators,
paradigmatically conditionals and negation. The final step is to analytical concept formation, in which the
sentential compounds formed at the third stage are decomposed by noting invariants under substitution.
This is actually the same method that gave us the notion of ingredient content at the third stage of concept
formation. For that metaconcept arises when we realize that two sentences that have the same pragmatic
potential as free-standing, force-bearing rational classifications can nonetheless make different
contributions to the content (and hence the force) of compound sentences in which they occur as
unendorsed components—that is, when we notice that substituting one for the other may change the free-
standing significance of asserting the compound sentence containing them. To form complex concepts, we
must apply the same methodology to sub-sentential expressions, paradigmatically singular terms, that have
multiple occurrences in those same logically compound sentences. Systematically assimilating sentences
into various equivalence classes accordingly as they can be regarded as substitutional variants of one
another is a distinctive kind of analysis of those compound sentences, as involving the application of
concepts that were not components out of which they were originally constructed. Concepts formed by this
sort of analysis are substantially and in principle more expressively powerful than those available at earlier
stages in the hierarchy of conceptual complexity. (They are, for instance, indispensible for even the
simplest mathematics.)
This hierarchy is not a psychological one, but a logical and semantic one. Concepts at the higher levels
of complexity presuppose those at lower levels not because creatures of a certain kind cannot in practice, as
a matter of fact, deploy the more complex kinds unless they can deploy the simpler ones, but because in
principle it is impossible to do so. Nothing could count as grasping or deploying the kinds of concepts that
13
A key part of the higher inferential grade of conceptuality (which includes the former, but transforms it) is that it is
multipremise material inferences that one learns to draw as conclusions (=responses) now to Boolean combinations of
the relatively enduring states that result from one’s own responses.
populate the upper reaches of the hierarchy without also grasping or deploying those drawn from its lower
levels. The dependencies involved are not empirical, but (meta)conceptual and normative. The Fregean
considerations that enforce the distinctions between and sequential arrangement of concept-kinds do not
arise from studying how concept-users actually work, but from investigation of what concept use
fundamentally is. They concern not how the trick (of concept use) is done, but what counts as doing it—a
normative, rather than an empirical issue. That is why it is philosophers who first came across this semantic
hierarchical metaconceptual structure of concept-kinds.
But cognitive scientists need to know about it. For it is part of the job of the disciplines that cognitive
science comprises to examine—each from its own distinctive point of view—all four grades of conceptual
activity: the use of more complex and sophisticated kinds of concepts, no less than that of the simpler and
less articulated sorts. The move from merely classificatory to genuinely descriptive concepts, for instance,
marks a giant step forward in the phylogenetic development of sapience. I do not think we yet know what
non-human creatures are capable of taking that step. Human children clearly do cross that boundary, but
when, and by what means? Can non-human primates learn to use conditionals? Has anyone ever tried to
teach them? The only reason to focus on that capacity, out of all the many linguistic constructions one
might investigate empirically in this regard, is an appreciation of the kind of semantic self-consciousness
about the rational relations among classifications (which marks the move from classification to rational
description) that they make possible. Computer scientists have, to be sure, expended some significant effort
in thinking about varieties of possible implementation of sentential compounding—for instance in
exploring what connectionist or parallel distributed processing systems can do. But they have not in the
same way appreciated the significance of the question of whether, to what extent, and how such
“vehicleless” representational architectures can capture the full range of concepts expressed by complex
predicates. (Their lack of syntactically compositional explicit symbolic representations prohibits the
standard way of expressing these concepts, for that way proceeds precisely by substitutional decomposition
of such explicit symbolic representations.) These are merely examples of potentially important questions
raised by the hierarchy of conceptual complexity that cognitive scientists have by and large not been moved
so much as to ask.
Why not? I think it is pretty clear that the answer is ignorance. Specifically, it is ignorance of the
considerations, put forward already by Frege, that draw the bright metaconceptual lines between different
grades of concepts, and arrange them in a strict presuppositional semantic hierarchy. Any adequately
trained cognitive scientist—even those working in disciplines far removed from computational
linguistics—can be presumed to have at least passing familiarity with the similarly four-membered
Chomsky hierarchy that lines up kinds of grammar, automaton, and syntactic complexity of languages in an
array from most basic (finite state automata computing regular languages specifiable by the simplest sort of
grammatical rules) to most sophisticated (two-stack pushdown automata computing recursively enumerable
language specifiable by unrestricted grammatical rules). But the at least equally significant semantic
distinctions I have been retailing have not similarly become a part of the common wisdom and theoretical
toolbox of cognitive science—even though they have been available for a half-century longer.
The cost of that ignorance, in questions not asked, theoretical constraints not appreciated, promising
avenues of empirical research not pursued, is great. Failure to appreciate the distinctions and relations
among fundamentally different kinds of concepts has led, I think, to a standing tendency systematically to
overestimate the extent to which one has constructed (in AI) or discerned in development (whether by
human children or non-human primates) or reverse-engineered (in psychology) what we users of the
fanciest sorts of concepts do. That underlying ignorance is culpable. But it is not the cognitive scientists
themselves who are culpable for their ignorance. The ideas in question are those that originally launched
the whole enterprise of analytic philosophy. I think it is fair to say that as we philosophers have explored
these ideas, we have gotten clearer about them in many respects. For one reason or another, though, we
have not shared the insights we have achieved. We are culpable for having kept this treasure trove to
ourselves. It is high time to be more generous in sharing these ideas.