February 26, 1970: Sellars to Harman
March 24, 1970: Harman to Sellars
November 20, 1970: Sellars to Harman
December 9, 1970: Harman to Sellars

New York, N.Y. l0021

February 26, 1970

Professor Gilbert Harman
Department of Philosophy
Princeton University
Princeton, New Jersey

Dear Gil:

I have read and reread your insightful and challenging review of Science and Metaphysics, and have been accumulating notes, I won't say for a reply, but for a first attempt at rising to your closing challenge. I shall concentrate on the points you raise concerning meaning and truth, leaving to another occasion topics which, though important, do not touch the very heart of the enterprise. Thus I shall not discuss your comments on the availability and/or adequacy of concepts of thinking-out-loud and its modes as models for the construction of a theoretical framework which introduce classical thought-episodes. I certainly did not mean to imply that all the thinking of Jones' Rylean contemporaries was thinking-out-loud, but only that they did more of it than we do. Most of their thinking would be, in the terminology of my "Language as Thought and as Communication," extremely short term proximate propensities to think such and such out loud. (I thought I had made this clear in SM.) Also, I have stressed from the beginning that inner thought episodes, though construed on the model of overt speech, are not to be thought of as "words going through ones head." No verbal imagery need occur, and even when it does, the imagery is not itself the thinking, but at most a symptom of it.

Moving closer to home, the second paragraph on p. 8 [410] is indeed "truer to (my) intention" than "the view (of inference) deplored above." (I have long been aware that if P implies Q then not-Q implies not-P.) The distinction between 'rules of action' and 'rules of criticism' plays a central role in my interpretation of the rule-governed nature of thought. Logical rules in the primary sense are rules of criticism. Inferring is not an action. It is not the sort of thing that can be done voluntarily -- though it may occur in a process which as a whole is voluntarily undertaken. I shall shortly pick up this general theme in the context of "observation".

I turn now to the central theme of truth. You write (pp. 8-9 [410])

"Sometimes he speaks of semantical rules as if they were principles of the theory of truth. At other times he speaks of them as if they were principles of the theory of evidence. The puzzle is whether he has confused two different sorts of principles, or has rather found a way to exploit connections between the two sorts of principle."

I am not clear why you suppose me to think of semantical rules in my sense as 'principles' of a 'theory' -- whether of truth or of evidence. Of course the concept of a semantical rule belongs to a theory, but the theory is a 'theory of knowledge' and deals with a whole family of related topics which include, but are by no means exhausted by, truth and evidence. Each type of semantical rule is in its own way relevant to the explication of each of these topics, without thereby becoming a 'principle' of the 'theory' of that topic.

I do not have much to say about the theory of evidence in SM, primarily because my views on induction and probability place them in the broader context of a theory of practical reasoning, to which the concluding chapter is a prolegomenon. As a result, I was primarily concerned with the bearing of semantical rules on problems pertaining to meaning and truth. I do, however, touch briefly on the notions of a law-like statement and inductive reasoning at the beginning of chapter V (sections 4-7), and the discussion of the evolution of conceptual frameworks toward the end of the chapter presupposes (rather cryptically, I am afraid) my general account of theoretical explanation.

Since I had avoided problems in the theory of evidence, I was puzzled by the fact that you seem to find them playing a pervasive (if confused) role in the argument. When I read "at other times he speaks of (semantical rules) as if they were principles of the theory of evidence," I asked myself 'where?' I thought you must be referring to my inclusion of L-rules and P-rules of inference among my semantical rules. Imagine, then, my surprise when on the following page you offer as an example which shows that "other things that Sellars says makes it seem that semantical rules are principles of the theory of evidence after all," the view, which you attribute to me, that there are "semantical rules connecting observation and thought." "Why," you expostulate, "bring in observation, if these rules are rules of truth rather than rules of evidence?" (p. 10 [410]).

Now nowhere do I distinguish between 'observation' and 'thought' and discuss their relation to one another. To do so suggests that an observation is a non-thought which is the awareness of a 'state of affairs,' and that the relevant semantical rule says that it is correct to think that-p if one is aware that a certain state-of affairs (perhaps that-p itself, or, perhaps, that it looks to me that-p) obtains. Surely this is exactly the kind of view which (dubbed 'the myth of the given' in "Empiricism and the Philosophy of Mind") I have been vigorously attacking since 1948 ("Realism and the New Way of Words"). To embrace the myth is to treat 'thought entry transitions' as though they were a special class of inferences, those in which the premises are given rather than merely thought to be the case. My view, on the contrary, is that 'observations' are themselves thoughts; they are thought-tokens which are correct responses to the objects which caused them. I deny that there is any such thing as an awareness that-p which isn't a thinking that-p. The semantical rules in terms of which the concept of observation is to be explicated concern, for example, not the correctness of thinking 'here is a red object' when one is aware that there is a red object in front of one, or that it looks to one that there is a red object in front of one, but rather the correctness of the thought 'here is a red object' as a response to a red object in front of one. (See also the opening sections of "Language as Thought and as Communication" and pp, 10-15 of "Actions and Events.")

It is therefore within a framework which I reject that you develop the options between which, as you see it, I must choose. Thus you write "everything depends on the exact nature of the rules connecting observation and thought. Evidential rules would correlate stimulation (or how it looks to one) with specific thoughts. Truth rules might correlate the actual (and not just apparent) observation of something with a thought of that thing." (p. 10 [410-11]) Presumably an example conforming to the first interpretation would be one in which the datum, its looking to one that there is a red object in front of one, would be evidential support for the thought that there is a red object in front of one. Presumably an example conforming to the second interpretation would be one in which the observation that there is a red object in front of one would verify one's thought that there is a red object in front of one. Although it is clear that neither interpretation fits my views, the second, involving as it does the idea that what is correct is the thought that an object is red, on the occasion of stimulation by a red object, is closer than the first. You sense this, and, therefore, write "presumably Sellars has something like the latter sort of rule in mind."[411] Curiously, you give a reference to two passages which, though I admit that they are cryptic, are surely incompatible with the framework in terms of which you are confronting me with the above choice.

This concludes your initial discussion of the question as to whether I confuse truth and evidence. You turn next to my theory of truth. Here, indeed, the chips are down.

The first place at which I began to feel sensitive (like a patient in a dentist's chair) was right at the beginning (p.11 [411]) where you ascribe to me the view that "certain representations are true...if and only if they picture particular facts according to rules of projection in C". Surely I refer to what is pictured as objects rather than facts. Indeed, already in "Truth and Correspondence" I had emphasized the distinction between the pseudo-relation of 'correspondence' between true statements and facts (a pseudo-relation because it turns out that true statements are facts) and the genuine relation of picturing which holds between certain statements as items in the spatio-temporal-causal order and the objects with which they are correlated by the rules of projection of the language. The schema for picturing is not

(statement) pictures (fact)

but rather

(object) pictures (objects).

But, then, since the word 'fact' is used in different ways by different philosophers, there was, as yet, no serious cause for alarm.

The remarks on p. 12 [412] are misleading in that they do not reflect my views on the nature of law-like statements. (They are material rules of inference.) On my view, also, the concept of two worlds which agree in their histories, but differ in their laws, is incoherent. But this is because attributes and laws are not conceptually independent, as traditional empiricism holds them to be. Thus geometrical attributes are not conceptually independent of geometrical principles, nor micro-physical attributes of micro-physical principles. On my view, law-like statements are material rules of inference of the form (roughly):

from 'K is a finite unexamined class of As,' it is correct to infer 'n/m K is B.'

An open quantified counterpart of the result of a finite application of this rule would be a statement of the form

n/m (of all) A is B.

And, where n/m is 1,

(x) Ax --> Bx.

Law-like statements themselves would, as meta-linguistic rules, be "determined" by the evidence, in the sense that the adoption of the rule is vindicated by practical reasoning in which the description of the evidence occurs as a premise. Needless to say the equivalence

'it is correct to infer . . . from ---' is true iff it is correct to infer . . . from ---.

holds as in all cases of truth. But the semantical rules with respect to which the truth (semantic assertibility) of 'it is correct to infer . . . from ---' is to be understood, are in no simple way related to the truth of atomic representations. For the rules in question are those which define inductive reasoning, and fall outside the recursive specification of truth conditions for molecular and quantified matter-of-factual statements.

Thus when you ascribe to me the view that

atomic representations are true if and only if they correctly picture the facts (sic), whereas the truth of other representations is completely determined by the set of true atomic representations (p. 13 [412])

I not only have serious reservations about the first clause, since the term 'fact' tends to short-circuit theories of truth, but would categorically reject the second clause, if it is intended to apply [to] all truths, (ethical principles, law-like statements, theoretical principles) other than true atomic representations. On the other hand, it would be difficult not to agree that the truth of truth-functions (including quantificational derivatives) of atomic sentences is so determined. Yet even here the relation of inferential connection to the truth of basic sentences must not be overlooked in assessing the force of "is determined by".

I come now to the heart of the matter, your critique of my critique of "Carnap-Tarski semantics." Much of what you have to say seems to me to rest on a misunderstanding for which I must share at least part of the responsibility. My attack is directed not against the idea of a recursive specification of truth conditions (how else would one explain the relation of the truth of 'p or q' to the truth or falsity 'p' and 'q'!), but to the idea that a recursive specification of truth conditions provides an explication of the concept of truth. I was looking for an account of the intension of the predicate 'true' and, hence, was concerned to distinguish carefully between the intension of 'true' and the conditions which sentences of various logical form must satisfy for this predicate to be (truly) applicable to them. One is tempted to put this by invoking the traditional contrast between the 'meaning' and the 'criteria' of truth, but 'criteria' is an accordion word, and its use might contaminate the discussion with the problem of evidence discussed above.

The same motivation guided my critique of Carnap's account of such syntactical predicates as 'sentence (of L)' in my Carnap volume essay. The point at issue is, in a familiar sense, philosophical rather than substantive. Thus there is an important sense in which what I reject is not Carnap-Tarski semantics, but rather the Carnap-Tarski philosophical interpretation of Carnap-Tarski semantics.

It is probably my fault that you did not pick up the significance of my claim to give an account of the intension of 'true.' Otherwise it might have occurred to you that I could agree with the Carnap-Tarski recursive account of truth-conditions (and satisfaction-conditions), without agreeing that they had captured the intension of either 'true' or 'satisfies.' You would have then been prepared to find me claiming that an adequate explication of 'true' and 'satisfies' must break out of the family of semantical concepts and relate them to such 'pragmatic' concepts as 'would be a correct perceptual response to,' i.e. to notions pertaining to the observation-inference-action game at all its levels.

The above general remarks should provide the background against which the following more detailed comments might fall into place. Notice, to begin with, that whereas I write in terms of a contrast between the intensions of 'true' and 'true of' (or its converse 'satisfies') and the conditions which govern the applicability of these predicates, you blur this distinction by formulating the issue as one which concerns the "characterization" of truth. Thus you write (p. 14 [413])

. . . because it attempts to offer a recursive characterization of truth, Tarski's theory can also be interpreted as providing a correspondence theory somewhat similar to Sellars' picture theory of truth.

The preceding remarks should make it clear that from my point of view, Tarski does not "provide" a "correspondence theory somewhat similar to Sellars' picture theory of truth," though it is compatible with a picture theory of the truth of matter-of-factual statements. Such a genuine picture theory would give a pragmatic account of the conditions which a basic matter-of-factual statement must satisfy to be "semantically assertible," and a recursive account of the semantical assertibility of molecular compounds of such statements. I can say 'pragmatic', because the analysis of the conditions which a basic sentence (I should now begin to speak in terms of statements) must satisfy in order to be 'semantically assertible' will be in terms of such notions as 'correct perceptual response to object Oi (or objects Oi, Oj)' and 'is an appropriate tensed counterpart of statement S.'

In the passage which follows the sentence quoted above, you write:

Things are complicated by the fact that Tarski envisions a recursive characterization (sic) of satisfaction as well as of truth, where Sellars envisions a non-recursive characterization (sic) of truth apart from satisfaction.

Without running the point into the ground, let me stress again that my aim was to give a non-recursive definition of truth, i.e. analysis or explication of 'true'. I am quite happy with the idea that truth-conditions are to be recursively specified. This does not mean, of course, that all truth-conditions (e.g. truth conditions for ethical statements) belong in the same recursive hierarchy, as would be the case if all sentences were truth functions of atomic matter-of-factual sentences, or quantifications of open matter-of-factual sentences. Thus, I do not envision a non-recursive characterization (as contrasted with explication) of truth. What, then, of satisfaction? Do I think that to characterize truth we must make use of the 'relation of satisfaction?' The answer is in the affirmative. The truth conditions of quantified statements are, indeed, to be specified recursively in terms of satisfaction. This, however, is compatible with the fact, on which I have laid great stress, that the focal concept of semantics is that of truth, the cash value of which is to be found in the 'truth move' illustrated by

'Snow is white' is true
so, snow is white

Reflection on the following two examples should help me make the point I wish to make about satisfaction:

'It is not day or the sun is shining' is true
so, it is not day or the sun is shining

'(x) x is a man --> is mortal' is true
so, (x) x is a man --> x is mortal

The identity of sense of 'true' as it occurs in these two examples is exhibited by the common form of the sequence. This identity of sense, however, goes along with the fact the truth condition of 'either it is not day or the sun is shining' is recursively specified in terms of the formulas

'not-p or q' is true iff 'not-p' is true or 'q' is true
'not-p' is true iff 'p' is false.

Similar considerations hold in the case of the second example. But before I elaborate this point I shall examine the following passage from your review,

. . . this brings us to Sellars' second reason for denying that "Tarski-Carnap semantics" provides any sort of correspondence theory of truth. Sellars claims (in effect) that the relation of satisfaction, which plays an essential role in Tarski's account of truth for quantified sentences, is not really a semantic relation at all. That is, he denies that satisfaction is a relation between language and the world. [413]

Now I certainly do not deny that satisfaction is a semantical concept. And I certainly do not deny that when we get to the level of basic sentences our account of truth conditions must include a relation between language and the world. What I deny is that satisfaction, as characterized in Tarski-Carnap semantics, is that relation. It is close by and near to that relation, but not identical with it. Leaving aside for a moment the considerations which focus attention on a concept of satisfaction which applies to sequences of objects, an object (or an n-tuple of objects) satisfies an open basic sentence if and only if (an appropriately tensed) closure of that sentence which contains designations of the object or objects would be true. Again, a present tensed closure would be true if and only if it would be a correct perceptual response to (i.e. would picture) the object(s) (Compare my concept of a 'verifying token' in "Realism and the New Way of Words.") Other differently, but appropriately, tensed counterparts would picture the object (or objects) by virtue of their relation to such a verifying token.

The above considerations point to a more precise formulation of the point at issue between the theory of truth I am defending, and the standard philosophical interpretation of the Tarski formalism. Granted that the truth of matter-of-factual statements involves at some point a direct relation between language and the world, and that statements which are not true by virtue of a direct relation to the world are true by virtue of being indirectly related to it in a way which can be specified by recursion -- i.e. granted all this, is this language-world relationship captured by the intention of either 'true' or 'true of' or 'satisfies?' My answer is no. The answer of these who hold the standard interpretation of the Carnap-Tarski account -- most recently Donald Davidson -- is yes. Thus, according to the latter "...the property of being true has been explained... in terms of a relation between language and something else . . . the relation, satisfaction . . ."{1}

Very well, then, what is satisfaction if it is not a relation between language and the world? The clue to the answer lies in the fact that the concept of the satisfaction of basic open sentences provides the basis for a recursive account of the truth of quantified sentences just as, abstracting from considerations of quantification, the concept of the truth of basic closed sentences provides the basis for a recursive account of the truth of their molecular compounds. Tarski, indeed, goes on to show us how it is possible to unite these recursive structures by defining the truth of a basic closed sentence in terms of satisfaction. Thus, if satisfaction were, indeed, a relation between open sentences and objects (more accurately, sequences of objects), we would, indeed, have been shown that the concept of truth is at bottom the concept of a relation between language and the world.

Now we can imagine a philosopher who is a 'finitist' -- not in the sense that he abjures the Cantorian paradise, but in the sense that he interprets references to varieties of infinite collections to be non-perspicuous references to varieties of open ended mathematical rules -- arguing that Tarski's recursive formalism can be restructured to take the truth of basic closed sentences as primary, and to treat the truth of quantified sentences as special case of the truth of molecular compounds. He could be expected to argue that only if the truth of basic closed sentences could be shown to be (not just 'involve' -- as of course they do) a relation between language and the world, could something like the traditional correspondence theory of the meaning of 'true' have been established. Of course in such a hypothetical (and heretical) reconstruction, the semantical predicates 'true of' and 'satisfies' would be definable in such a way that, for example,

'fx' is true of a iff 'fa' is true

would be true by definition. But this would provide grounds for saying that the truth of 'fa' is a relation between 'fa' and something non-linguistic only if we already knew informally, as a criterion for the adequacy of the definition, that 'true of' stands for a relation between open sentences and objects. If, to push the matter one step further, it is pointed out that the above example should read

'fx' is true of a iff 'fa' is true and 'a' denotes a

our philosopher would claim that we are committed to the idea that 'true of' stands for a relation between open sentences and objects if and only if either the truth of 'fa' is (not just 'involves') a relation between 'fa' and something non-linguistic, or 'denotes' stands for (not just 'involves') a relation between a linguistic and non-linguistic item.

I have introduced this hypothetical position to focus attention on the manner in which a number of theses which I defend in SM stand or fall together. Thus:

(1) I insisted that the truth of a basic sentence (like all truth) is semantic assertibility. The truth of basic sentences, however, involves, indeed is grounded in, the existence of a relation of picturing between tokens of the sentences and the objects to which they refer.

(2) I insisted that 'denotes' does not stand for a relation between a linguistic expression and an object. On the other hand the denotation statement would not be true unless the expression did stand in complex matter-of factual relations to the object.

(3) I insisted that the 'stands for' in

'f' stands for __________

does not stand for a relation between a linguistic and non-linguistic item, although the indented statement would not be true unless 'f' stood in complex matter-of-factual relations to verbal behavior and non-linguistic objects.

(4) I insisted that the contexts

'fx' is true of a
a satisfies 'fx'
'b' denotes a

are, in Frege's sense oblique. In scholastic terms, the expression 'a' in the above examples is used in second intention.

All of this is by way of comment on an intriguing passage in Davidson's paper in which, commenting on 'correspondence to fact' theories of truth, he writes:

One well-explored consequence is that it becomes difficult to describe the fact that verifies the sentence except by using that sentence itself. The other consequence is that the relation of correspondence (or "picturing") seems to have direct application to only the simplest sentences (Dolores loves Dagmar). This prompts fact-theorists to try to explain the truth of all sentences in terms of the truth of the simplest and hence, in particular, to interpret quantification as mere shorthand for conjunctions or alternations (perhaps infinite in length) of the simplest sentences. The irony is that, in so far as we can see quantification in this light, there is no real need for anything like correspondence.

That I am not a 'correspondence to fact' theorist should, by now, be clear. Yet if 'fact' is replaced by 'objects' I am a correspondence-to-objects theorist. For it is an essential feature of my theory that although 'true' does not mean corresponds to objects, but rather semantically assertible, nevertheless, correspondence to the relevant objects is a necessary and sufficient condition of the semantical assertibility of a basic statement, and this would be so even if quantificational theory could, in principle, be treated as an extension of propositional logic, i.e. even if our recursive account of truth conditions could have as its basis the enumeration,

'fa' is true iff Tom is tall

'gb' is true iff Dick is thin

Where, given PMese syntax, and the dictionary,

'a' des Tom
'b' des Dick

'f' des tall things
'g' des thin things

we could also make such finer-grained statements as that

'fx' is true of Tom iff Tom C tall things
'gx' is true of Dick iff Dick C thin things

and recursively specify the truth conditions of molecular senses (including quantified sentences) belonging to the language thus characterized.

Against this background, let me return to your comments. After pointing out that

Strictly speaking, one should say that an open sentence is satisfied by a sequence of objects [414]

(by which you are, I take it, calling attention to the fact that according to Tarski the immediate concept we need to provide a recursive account of the truth of quantified sentences relates open sentences to sequences of objects, and is to be distinguished from a more basic concept of satisfaction according to which an object or n-tuple of objects satisfies, i.e. belongs to the extension of, an open sentence, where this latter concept is to be taken a semantical primitive.) You continue,

for our purposes we may think of a predicate expression as satisfied by certain objects when it is true of them.[414]

Here you are, I take it, calling attention to the fact that the semantical primitive of Tarski's account is the idea of an object or n-tuple of objects belonging to [the] extension of an open sentence. In other words, the semantical primitive of Tarski's account is the concept of an open sentence being true of an n-tuple of objects. If one begins to feel that the Tarski account is moving in a circle -- usefully, however, and in a circle which is large enough for its purpose -- one might also begin to feel that the connection between Tarski's semantical primitive and the world, which connections involves learned patterns of verbal behavior, remains to be explicated. This is what my account attempts to do. That I find the focal point of the connection to be at the level of 'true' rather than 'true of' does not mean that I minimize the role of 'true of' in explicating the semantics of quantification. I stress 'true' because here is where the matter-of-factual relationship of picturing, definable in the framework of language as a rule-governed phenomenon in rerum natura, joins up with the 'truth move' which is characteristic of all modes of truth and which is exhibited by the schema

That-p is true
so p

I continue with your text.

. . . for our purposes we may think of a predicate expression as satisfied by certain objects when it is true of them. Thus, "is wise" is satisfied by Tarski. Now Sellars says, in effect, that what I have just said in the preceding sentence only appears to assert a relation between language and the world. He argues that my claim really asserts a relation between the expression "is wise" and "Tarski" and he suggests that it would be more accurate to write it in any of the following (for him equivalent) ways:
"is wise" is satisfied by "Tarski"
"is wise" is true of "Tarski"
"Tarski"^"is wise" is true
"Tarski is wise" is true
He concludes that "Tarski-Carnap" semantics does not show how language is related to the world. (p. 15 [414])

Now I do indeed claim that "Tarski-Carnap" semantics does not show how language is related to the world. I criticize its claim to have explicated the tie between language and the world. My present purpose, however, is to reply to your criticism of my treatment of satisfaction.

The first point is a relatively trivial one. I do not hold that

"Tarski is wise" is true

is a 'way of writing'

"is wise" is satisfied by (or true of) "Tarski"

I hold, simply, that satisfaction and truth of are definable in terms of 'true'. You criticize this claim as follows

Notice that Sellars' analysis works only if every object that satisfies any expression has a name.

I do not think that this is true, but to show it I must mobilize features of my account of meaning and truth on which you do not touch. It is an essential feature of my philosophy of language that the nature of quantification and the role of variables is to be explicated in terms of the concept of an ideal language which contains names of all objects, and predicates for all basic attributes and relations. Such a language I called, in "Realism and the New Way of Words," the language of omniscience. In such an ideal language there would be the resources for constructing basic sentences which adequately picture all objects. I argued that we should view our language as a "schema" of such an ideal language. The cash value of this, for my present purposes, is that we should view the variables of our language as taking for their substituends not only the constants we now have, but the constants which are to be introduced in accordance with rational methods. (See SM, V, 66). Looked at behavioristically the matter-of-factual connection between language and the world includes acquirable as well as acquired patterns of verbal behavior. It is the very nature of language to have a reach which exceeds its grasp. This feature belongs not only to object language variables but to meta-linguistic variables as well. It is a feature of natural languages which recurs at all levels.

This feature of quantification enters as follows into my account of 'true of' and 'satisfies.' Abstracting from my concern with inter-linguistic identity of roles in the conceptual game of observation-inference-action, and limiting myself to roles as embodied in English sign-designs, and emphasizing that English is not be taken as English 1970 but as the continuing English of the English speaking scientific cammunity, the following (regimented) account expresses my views. Let 'IND,' 'PRED,' 'SENT,' etc. be meta-linguistic variables which take quoted expressions as substituends. With these resourses we can say not only

'is wise' is true of 'Tarski' iff 'Tarski is wise' is true


(E IND) 'is wise' is true of IND iff (E IND) IND^'is wise' is true.

I need no more be able to replace

(E IND) IND 'is wise' is true

by a disjunction which begins

'Harman is wise' is true or 'Tarski is wise' is true . . .

then I need be able to replace

(E x) x is tall

by a disjunction which begins

Tom is tall or Dick is tall . . .

Nor does

'x' ranges over a, b, c . . .

formulate a 'ranging relation' between the variable 'x' and non-linguistic entities. It is rather the material mode for

'x' has as substituends 'a,' 'b,' 'c,'...

Note that it would be a serious mistake to suppose that the two indented statements are equivalent to

'x' ranges over 'a,' 'b,'. . .

The formal and material modes of speech must not be mixed together. I have put this elsewhere by saying that contexts such as 'ranges over . . .' already play a quoting role, so that to follow them with quoted expressions moves them up one higher level than is intended. Actually the same is true of the context 'true of . . .' That is why it is best to underline the expression following it rather than put the expression in quotes. Yet putting the expression following 'true of' in quotation marks, though technically incorrect, does serve the purpose of emphasizing that the expression which properly follows 'true of' is being mentioned (autonymously) rather than simply used.

If the interpretations of range statements as material mode of speech for substitution statements seems shocking, it is because we tend to overlook the fact that the right hand side involves a rule of construction. In the above model the rule is based on the alphabet. In a natural language it would involve the rules concerning the introduction of referring expressions into the language. These involve something like a coordinate system, and can initially be characterized as syntactical and pragmatical rules governing the use of such a system. Thus, to oversimplify we can envisage the use of numerals as subscripts making referring expressions out of variables.

'x' ranges over x1, x2, x3 . . .
'x' has as substituends 'x1,' 'x2,' 'x3,'. . .

Once again, as in the case of other semantical concepts, the semantical concept of 'ranging over' is connected with a pragmatic notion pertaining to patterns of verbal behavior in relation to extra-linguistic objects, but is not this notion.

Strictly speaking, of course, objects are not locations, but have locations. This is why the substituends for 'x' in 'x is wise' are not as immediately available as are the terms in an arithmetical sequence of which the rule is known. As I pointed out above, the potential substituends for the variables of a continuing language are potential with respect to the scientific enterprise as a whole, and not with respect to anything as simple as the rules governing a coordinate system.

You write

. . . Tarski introduced to the notion of satisfaction in order to give truth conditions for quantified sentences. He would not have had to do so if he had been able to assume that all objects in the range of a quantifier have names in the language, for if they do one can make do with so-called substitutional quantification. [And indeed, Sellars' version of quantification] is the substitutional variety.(15-6)[414].

I have always been uneasy about the phrase 'substitutional variety of quantification,' which implies that there are two 'varieties' of quantification, 'substitutional' and 'objectual' (or, as you put it, 'referential'). As I see it there is just quantification, and the distinction between 'substitutional' and 'objectual' is a not too happy way of contrasting two theories of the sense of quantified statements. My version of quantification is 'substitutional' only in the sense that

(1) I interpret statements about the range of variables along the above lines, in keeping with my general approach to semantical statements.

(2) I regard the following statements as quantificational in exactly the same sense:

(Ec) c(p,q)
thus, or(p,q)

(Ef) Tom is f
thus, Tom is tall

(Ex) x is tall
thus, Tom is tall

This parallels the above account of the range of 'x'; thus

'c' ranges over connectives

has the sense of

'c' has as substituends, 'and' 'or' 'not' . . .


'f' ranges over properties

has the sense of

'f' has as substituends 'red' . . . 'square' . . .

I, therefore, see no reason why I cannot avail myself of Tarski's method of giving truth conditions for quantified sentences. I quite agree with him that it is necessary to use the concept of satisfaction to specify these conditions. I see nothing incompatible with this in my analysis of satisfaction. You write

Sellars is just wrong when he says that the relation of satisfaction in "Tarski-Carnap semantics" does not relate language and the world. On the contrary it only works because it does relate language and the world.[414]

On the contrary, it works because of its formal properties and not because of the Tarski-Carnap conviction that 'true of' stands for a relation between linguistic and non-linguistic entities. You admit, in the paragraph which follows the above, that I can "give truth conditions . . . for substitutionally quantified sentences," though I do not make use of "the notion of satisfaction," although you insist that Tarski can "give truth conditions for referentially quantified sentences" because he makes use of "the notion of satisfaction." I reply that Tarski and I are both making use of the same semantical concept of satisfaction, and giving truth conditions for the same quantified sentences -- but are giving different philosophical accounts of what we are doing. The difference between us is not a semantical one, but a meta-semantical one. It is in this spirit that I accept your olive branch

So perhaps no real objection to Sellars' theory arises here.[415]

Now I come to a comment which really perplexes me:

Tarski requires that a truth characterization satisfy condition T{2} referred to above. Sellars does not require this.[415]

You give no grounds for this claim. It apparently has something to do with my treatment of law-like statements, but since I have little to say about this in SM, and since you give no hint as to what feature of my treatment motivates your remark, I simply do not see what you are driving at. What I do say about law-like statements (V,5) characterizes them as material principles of inference. But why equivalences of the form

'That something is f materially implies that it is g' is true iff that something is f materially implies that it is g.

should be exceptions to condition T, I completely fail to see.

You do, however, go on to make another point about the place of condition T in my account which does touch closer to home. You note that whereas I have quite a bit to say about the sense in which 'it is true that Carnap is wise' implies 'Carnap is wise,' I offer no ". . . argument to show that Carnap is wise only if it is true that Carnap is wise." Thus, you conclude, I "do not offer any account of why Carnap is wise if and only if it is true that Carnap is wise.

This is a fair criticism of the argument of the chapter on truth, even though at no point in that chapter did I "pretend to explain" why Carnap is wise if and only if it is true that Carnap is wise. I should have had something to say about the sense in which 'Carnap is wise' implies 'it is true that Carnap is wise.' My failure to do so was simply another expression of the fact that my disagreement with what I (unfortunately) called "Carnap-Tarski semantics" was philosophical rather than formal.

Let me attempt to remedy this omission, or, more accurate to sketch the general line such a remedy would take. The point is an important one, because it requires an explicit discussion of the connection between the concept of truth and the concept of semantical rule. Thus, if I had filled this gap in my discussion of truth, the question of whether my semantical rules are rules of truth [or rules] of evidence, to which the first part of this letter was devoted, would never have arisen.

[editor's note: The following passages were amended by Sellars in his Nov. 20 letter to Harman.]
The problem is most pressing with respect to basic sentences. Yet, to fix our ideas (as Nagel says), it is worth looking at a less problematic case. In what sense does 'p or q' imply 'it is true that p or q?' Or, to translate this in terms of my analysis, but omitting dot-niceties, in what sense does 'p or q' imply " 'p or q' is semantically assertible (with respect to relevant rules)?" Or, again, what gives kosherity to the sequence

1. p or q
2. 'p or q' is SA (semantically assertible).

Now, to inscribe the design p or q in a language-using frame of mind, implies that it can be classified as a token of some expression or other. And, indeed, the question presupposes that inscription 1 can be classified as a 'p or q.' Furthermore, although we want to say that in an extended sense the implication we are seeking to understand holds in the case of all tokens of 'p or q' whether they occur in play rehearsals, examples, tape-recorder-play-backs, tall stories, etc., we distinguish as primary the candid, cognitively motivated, use of language. Assuming this primary context, we consider the alternatives

1 conforms to the relevant semantical correctness (is SA)
1 does not conform to the relevant semantical correctness (is not SA)

So far we have

1: p or q
2: 1 is a primary token of 'p or q'
3: 1 is SA iff 'p or q' is SA.

Now the statement

p or q, but 'p or q' is not SA

is, in the sense which Moore attempted to analyze incoherent. Thus if we are prepared to stick with 1, the derivation continues

4: 1 is SA (pragmatic implication)
5: 'p or q' is SA.

So far our sequence exhibits the connection between the semantical assertibility (truth) of 'p or q' and the semantic assertibility (truth) of " 'p or q' is SA." To show how the formulation of this connection is justified, we continue the derivation as follows:

6: 5 is a primary token of " 'p or q' is SA"
7: 5 is SA iff " 'p or q' is SA" is SA
8: 5 is SA (pragmatic implication)
9: " 'p or q' is SA" is SA

This derivation, as before, contains a pragmatic step. Nevertheless it is conceptually binding and explains the incoherence of the sequence

p or q
'p or q' is not SA (true)

and hence the sense in which 'p or q' implies 'it is true that p or q.' And, after all, if I am right that the 'if' direction of the equivalence

it is true that p or q iff p or q

involves a pragmatic consideration (the 'truth-move') it should not be surprising that pragmatic considerations enter into the 'only if' direction as well.

But what, you may ask, are the 'semantical correctnesses' to which you refer? 'Surely the above account is empty until they are specified. To this the answer is that in part they are correctnesses which relate to the function of the connective 'or' and are reasonably unproblematic. But in part they concern the propositional constituents of 'p or q' and raise the puzzles of 'truth' all over again. To resolve them we must first take into account the recursive characterization of truth conditions which is so important a feature of Tarski's account. This tells us that

it is true that p or q iff it is true that p or it is true that q

Thus, given that 'p', for example, is a basic sentence, this takes us to the problem of the equivalence

it is true that p iff p

Assuming, as before that 'truth-move' strategy takes care of 'if' direction, the problem remains of explicating the sense in which 'p' implies 'it is true that p.'

[editor's note: This is the end of the amended paragraphs.]
The strategy is the same, but this time we come face to face with the problem: what are the relations between basic sentences and objects (not facts) which define the truth conditions of the sentences, and, hence, indirectly, the truth conditions of molecular and quantified sentences at this level?

In this context, of course, tenses (and demonstratives generally) become important, and attention must be paid to the fact that tokens of differently tensed sentences can "make the same statement" i.e. translate into one and the same tensed sentence which is appropriate to our use now.

I sketched a classification of the semantical rules relevant to our problem in the table in the end of chapter V. Of particular importance are those belonging to categories II, (a), (alpha) and II, (a), (beta). The fundamental principle is that a basic sentence, in the appropriate tense for our use now, is true if and only if a token of the corresponding present tense sentence would be (would have been) a correct language entry (perceptual) response to a certain object (or objects). Let us call such a correct response a 'verifying token' of the basic sentence. Its correctness involves (a) that the singular term or terms it contains be appropriate; (b) that the predicate be appropriate.

That a perceptual response token of 'this is red' be a verifying token for 'a was red' requires that semantical rules connect the singular term 'a' with an object located at the place and time with respect to which the perceptual response was made. This 'connection' of the singular term 'a' with an object is not the pseudo-relation of denotation, but is connected with the latter much as the truth condition of a specific proposition is connected with its character of being true. It is an empirical connection in the causal order between 'a' and other expressions in a schematic and developing world story, a sub-set of the sentences which are here-now perceptual responses. The identification criteria for 'a' must connect it with other sentences which, in present tense form, would have been correct perceptual responses to a and to other objects spatio-temporally related to it. The application of the language essentially involves the application of a (schematic) world story formulated in that language (compare V, 30). A further exploration of this point would tie it in with the discussion of the range of variables for individuals in an earlier paragraph of this letter.

Again, that a perceptual response token of 'this is red' be a verifying token of 'a is red' requires that tokens which contain the predicative expression 'red' be correct responses to red objects. Here is where my interpretation of the subject-predicate distinction, and the status of abstract entities, becomes important. For if they are not taken into account the confusions of fact-ontologies become, as I see it, unavoidable. My analysis is intended to explain, without reference to accessibility to quantification, the sense in which predicates are syncategorimatic expressions, and, together with my account of abstract singular terms, to explain why quantification over predicate variables does not commit one to a platonistic ontology.

A token of 'fa', viewed in terms of what might be called its depth-grammar, is not an 'f' concatenated with an 'a' but rather an f*/'a'/, where an f*/'a'/ is an 'a' which is concatenated with an 'f,' just as a white/dog/ is a dog which is white. The second requirement, then, is to the effect that red* singular terms are appropriate responses to red objects.

The importance of this account is the fact that it keeps us from looking for an object 'red', and by getting us away from the name-object model, enables us to see how the semantical rules for the predicate 'red', i.e. the rules for being a red* singular term can essentially involve its relations to other color predicates and to predicates of the space-time family. It enables us to see how these rules can be, as I put it in "Realism and the New Way of Words," conformation rules. It is but a step from here to argue the involvement of law-like statements in the meaning of predicates.

I know that I have barely scratched the surface of the concept of a verifying token. But the above might be useful by suggesting useful questions. Let me continue, then, by saying that verifying tokens are primary pictures of objects. Other tokens are pictures if they are appropriate transforms of primary tokens.

Notice that the correctness of a verifying token (primary picture) is, in the terminology of "Language as Thought and as Communication," a matter of its conforming to ought-to-bes, rather than ought-to-dos. For perceptual responses are, as such, involuntary; passions rather than actions. It is ultimately in terms of such ought-to-be correctness, that the truth conditions, i.e. conditions of the semantic assertibility, of basic sentences is to be defined. Roughly

'fa' is true (SA) iff 'fa's picture a

The right-hand side of this equivalence formulates the truth condition for 'fa.' This time, however, unlike the other cases in the recursive chain, the condition directly involves the relation of picturing which, unlike concepts belonging to the family explored by classical semantics, (stands for, denotes, extension, intension, reference, truth, truth of, etc,) is definable in naturalistic terms, thus word-word and word-world uniformities. To be sure, the uniformities are controlled, along the lines discussed in "Language as Thought and as Communication," by ought-to-bes (principles of criticism), which are, themselves, related to ought-to-dos (principles of action) relating to verbal behavior. But picturing itself is a natural, not a normative, relationship. I have explored the 'causal' force of normative language in a number of papers, but it would be folly to touch on it here.

This concludes what I have to say (on this occasion) in reply to your critique or my critique of "Tarski-Carnap semantics". I shall close with brief comments on the remaining paragraphs of your review.

My argument to the effect that the world as conceived in the Manifest Image is, in the Kantian sense, phenomenal, hinges not on considerations pertaining to sets, sums, and slices, as you seem to imply [417], but on considerations pertaining to the way in which sense qualities are involved in the constitution of these objects. It has nothing to do with the question whether on choice of primitives for a given framework is better than another. Incidentally, your use of [the] term 'colorless' on p, 19 [417] trades on an ambiguity. It may mean 'looks like clear water', and stand in this sense for a visual quality, or it may mean 'not having a visual quality at all.' Atoms are not colorless in the first sense, but rather in the second.

I have nowhere said that the Scientific Image doesn't include clusters, configurations, sets, wholes, slices, etc. of atoms nor that it doesn't include atomic events, states, processes, interactions, etc., etc. My general views on ontological categories applies to the Scientific as well as to the Manifest Image. I have said it doesn't include chairs, tables, etc., as conceived in the Manifest Image (see above). It does, however, include 'successor concepts' of the relevant Manifest concepts. Thus, as I point out at the end of Chapter V, there is a sense in which the chairs, tables, etc, of the Manifest Image really exist.

As for your concluding reprise of the rules of truth, rules of evidence theme, I think that my letter has directly or implicitly covered this ground. If not, however, I hope that my remarks have not been so baffling that they have failed to advance our (I hope) continuing dialogue.

As ever,

Wilfrid Sellars



Princeton University
Department of Philosophy
1879 Hall
Princeton, New Jersey 08540

March 24, 1970

Professor Wilfrid Sellars
The Rockefeller University
New York, New York 10021

Dear Wilfrid:

Thank you for the comments on my review of Science and Metaphysics. I think I finally see the difference between what you are trying to do and what Tarski and Carnap were trying to do. You want to say what "true" means, whereas they attempted to give necessary end sufficient conditions for truth in one or another particular formal language.

I think the difference can be expressed more clearly if what you are doing is compared with what Quine does in section 4 of "Two Dogmas of Empiricism." Quine argues that Carnap does not explain what "analytic" means. He points out that Carnap defines "analytic-in-L0" for a particular language L0 but never defines "analytic-in-L" for varying L. Thus Carnap never characterizes analyticity in general. Similarly, you might say, the semantic theory of truth never defines "true-in-L" for varying L; at best it defines "true-in-L0" for a particular language L0.

Quine points out that any adequate account or analyticity in general would have to refer to "mental or behavioral or cultural factors." Similarly, you point out that the same holds for any adequate account or truth in general. In particular, the semantical rules of truth must have something to do with the use of language, If I understand you, you go on to suggest that these rules are rules in accordance with which users of the language behave.

This way of looking at what you are doing clears up some of my confusions about what you say about Tarski. For example, on page 21 you say that Tarski's account moves in a large circle. I was originally puzzled by this, since Tarski's account seems no more circular than any other recursive definition. Your point is, not that Tarski's account is circular as an account of truth-in-L0, but that it becomes circular if turned into an account of truth-in-L for variable L.

Similarly, I take it that when you deny that Tarski's relation of satisfaction relates objects with linguistic expressions, what you mean is that it cannot do this if it is treated as a relation introduced into an account of truth in general. For it is obvious that as a relation introduced simply to account for truth-in-L0 satisfaction does relate (sequences of) objects to linguistic expressions. It is simply defined to do that. Suppose a relation R is defined as follows:

"Rxy =df x is wise and y is the expression 'is wise'."

R is defined to be a relation between objects and a linguistic expression. Tarski's relation satisfaction(-in-L0) is simply a more complex version of R.

Quine specifies what he takes to be the relevant "mental or behavioral or cultural factors" in Word and Object. They turn out to be a meager lot, definable in terms of a speaker's disposition to assent to or dissent from sentences as the result of varying sensory stimulation.

What you count as relevant "mental or behavioral or cultural factors" differ in two important respects from what Quine counts as such factors. For one thing, you think there are more factors: in addition to perceptual responses, you mention inference and the effect of practical reasoning on action. Furthermore, where for Quine a perceptual response would be a response to certain stimulation, for you it is a response to a perceived object that causes the response. And this second point is a special case of a more general difference between Quine's approach and yours. An obscure way to put this (rightly condemned in your letter) is that Quine thinks that speakers of a language act in accordance with rules of evidence, which give meaning to their words, whereas you think that speakers act in accordance with rules of truth, which give meaning to their words.

Thus for Quine, what speakers learn when they learn "Rabbit" is to assent to "Rabbit" given certain conditions of stimulation. For you, speakers learn to assent to "Rabbit" given the perception of an actual rabbit. Or, and better, for you, speakers learn to assent to "That is a rabbit" given the perception of an actual rabbit. Such a response pictures a perceived rabbit as a rabbit.

(I went wrong in my review when I spoke of responses to observation where I should have spoken of perceptual responses. I went wrong again when I spoke of picturing the facts where I should have spoken of picturing objects.)

Since you wish to give a general characterization of truth and of semantical rules, you are committed to the claim that in a situation of radical translation it is empirically discoverable what the semantical rules concerning picturing are for the native language in question. That is, you are committed to the claim that it is empirically discoverable whether or not a given native utterance is a correct perceptual response to rabbits, i.e. whether it correctly pictures rabbits. (Quine would of course deny this claim. I do not see that you have any adequate reply to Quine -- but let us by-pass that issue; there are enough other issues.)

I must confess that I am still puzzled by something I discussed on pages 20-21 [418?] of my review. Perhaps I can now express my puzzlement in a clearer fashion than I did there. Given the way I have formulated your objection to Tarski (on page one of this letter), you are committed to characterizing "true-in-L" for varying L. Since you characterize truth in part in terms of picturing, you are committed to characterizing "pictures-in-L" for varying L. Now, in much of what you say, picturing seems to be something that occurs by way of perceptual responses; indeed one might attempt to formulate a general characterization of picturing in terms of perceptual responses and how such responses can be criticized. But what about Science and Metaphysics V, 92? There you assert that "singular statements in the language of microphysics . . . constitute pictures of microphysical objects and events . . ." But you also agree that "no singular statement about microphysical particles can occur in a language entry transition or observation" by which I take it you mean a perceptual response. This seems to break the connection between pictures and perceptual responses and seemingly permits me to return your complaint against Tarski: you do not characterize "picturing-in-L" for varying L, since you do not say what picturing is if it is not tied to perceptual responses. You have not explained what "picturing-in-L" comes to if L is taken to the language of microphysics. That means you have not explained what "truth-in-L" comes to if L is the language of microphysics.

Finally I would like to remark on your discussion of the schema: "p" is true if and only if p. On page 32 you allow a derivation to use a principle of "pragmatic implication" and you refer me to Moore's discussion of such implications. As I recall, one case Moore discussed involved belief. There is a pragmatic implication in either direction between "I believe that p" and "p". Notice, however, that it is not true that I believe that p if and only if p. Pragmatic implications do not in general support conditional statements.

On pages 34-38 of your letter you discuss the important case where "p" takes the form of a basic statement "fa," true if and only if "f"s picture a. But I do not see that you have shown that "f"s picture a if and only if fa. That ought to follow from your account of picturing. I do not see that it does. You haven't said enough about what picturing is to get that conclusion.

I hope that you will think I have learned from your letter. I look forward to learning more.


Gilbert H. Harman




November 20, 1970

Professor Gilbert Harman
Department of Philosophy
Princeton University
Princeton, New Jersey

Dear Gil:

It is about time I put down on paper some of the things I said in response to your letter of March 24 during lunch at that unfindable Italian restaurant. I made notes shortly afterward and have been writing in fits and starts a formal 'reply' a la Sellars-Chisholm. But your letter raises so many important questions which I must think through again, that if I wait until I have them all in hand, I will have written Son of Science and Metaphysics. What seems to be called for, therefore, is a series of less ambitious letters which take up separate questions as my ideas about them take shape.

First a note about the way in which you relate my distinction between analyzing the sense of 'true' and giving a recursive specification of truth conditions, to Quine's distinction between giving a definition of 'analytic-in-L' for variable L, and giving a definition of 'analytic-in-L0.' You suggest that "the semantic theory of truth never defines 'true-in-L' for variable L; at best it defines 'true-in-L0' for a particular language L0." The analogy, as you see it, yields a possible interpretation of my claim that "Tarski's account moves in a large circle," which puzzled you "since Tarski's account seems to be no more circular than any other recursive definition." You find it less puzzling if my claim is interpreted as the "point, not that Tarski's account is circular as an account of truth-in-L0, but that it becomes circular if turned into an account of truth-in-L for variable L."

I think that this is a very helpful move, particularly since I independently made closely related distinctions in my essay in the Carnap volume. Yet I would want to press the point a bit further, and distinguish, as I did in that essay, between, on the one hand, 'defining' (to take a simple case)

Sentence (in L0)

by a recursive specification of sentential expressions of L0, -- which 'definition' amounts to a recursive listing of sentences of L0, and does not mention (though it makes a tacit use of) the function by performing which these expressions are sentences of L0 -- and, on the other hand, giving a definition of 'sentence (in L0)' which both

specifies the items in L0, which do the sentence job


explicates what the sentence job is.

Only the latter would I count as a definition proper of 'sentence (in L0).' On the other hand, as I pointed out in the essay, we can introduce 'sentence-in-L0' as a contrived predicate which stands to 'sentence (in L0)' as the contrived predicate

x and y are the-twins-in-the-Jones-family

(where this is introduced to mean

x = Tom and y = Dick or x = Jack and y = Jill or x = Bob and y = Carol)

stands to

x and y are the twins in the Jones family

(where this means something like

x = Tom and y = Dick and x is connate dual sibling of y or
x = Jack and y = Jill and x is connate dual sibling of y or
x = Bob and y = Carol and x is connate dual sibling of y)

Notice that the former predicate is introduced, and as introduced does not contain as an element in its sense the concept of twin. It simply abbreviates the listing of twins in the Jones family -- even though the concept of twin (and, for that matter, of the Jones family) is used in picking out the items to be listed. Nevertheless if the stipulation 'catches on', the definition of the contrived predicate by means of the disjunction of conjunctions which follows it would be a definition proper, since the predicate, thus introduced, would have this sense.

If now, we were to use the definiens of the contrived predicate as the 'definition' of open sentence 'x and y are the twins in the Jones family, which does contain as an element of its sense the concept of twin, then, in my terminology, we would be giving an extensional definition' of this predicate, but not a definition proper.

I have no objection to 'extensional definitions' nor, in particular, to 'extensional definitions' which list recursively. I also have no objection to such contrived predicates as 'sentence-in-L0' (as contrasted with 'sentence (in L0).' Correctly used, both are legitimate and can be clarifying. I do however, think that when their role is misunderstood, they can be the source of philosophical perplexity, for they may make it seem to philosophers that certain concepts have been explicated, when what has been explicated is their contrived extensional counterparts.

Thus even though the contrived predicate 'sentence-in-L0' can be explicated without any explication or the concept of sentence, the use of this contrived predicate presupposes the concept of sentence. Again, the explicans of the contrived predicate does not explicate even the concept of 'sentence (in L0)' -- let alone 'sentence (in L)' for variable L -- the corresponding non-contrived predicate, but simply gives its extensional exposition.

Thus, from my point or view, the interesting question about Tarski's definition of 'true-in-L,' is the following: How are we to understand such statements (at the basic recursive level) as

a satisfies 'fx' (in L0)
'fx' (in L0) is true of a
a belongs to the denotation of 'fx' (in L0).

There are two alternatives: (1) 'satisfies' or 'true of' is used in its ordinary sense, and this ordinary sense involves something like my concept of semantic assertibility or correctness of dequoting, in which case the account of truth (in L0) brings in concepts of meaning and truth pertaining to our background language, which concepts it does not attempt to explicate, though it uses them in the process of giving a recursive account of necessary and sufficient conditions for being a true sentence (of L0).

The second alternative, (2), is to construe

x satisfies y

as used in the 'definition' of 'true sentence (of L0)' as simply a more complicated case of the type of recursive specification illustrated by

x denotes y =
x = 'New York' and y = New York or
x = 'Chicago' and y = Chicago or
x = 'Pittsburgh' and y = Pittsburgh
. . .
Here, however, there are two sub-alternatives: (2a) Our account of 'true (in L,)' doesn't make direct use of the concept truth-of, the latter coming into the account only as the ordinary sense of 'sentence' came into the recursive specification of sentence (in L0). Or, (2b), the ordinary sense or 'true of' simply is a recursive specification (in our background language) of pairs of expressions and objects. The latter, (2b) would amount to the claim that our background semantical concepts can be explicated in terms of the pattern (where 'Sigma' is a semantical predicate),
Sigma (xy) =
x = expression-1 and y = item-1 or
x = expression-1 and y = item-2 or
x = expression-2 and y = item-1 or
x = expression-2 and y = item-2 or
x = expression-n and y = item-n or
Only on alternatives (1) and (2a) would Tarski's account be subject to my charge of 'circularity' as an explication of the concept of truth. On alternative (2b), the charge of circularity would fail, because the account of 'true-in-L0' would not make use of unexplicated semantical concepts in the background language, because these semantical concepts themselves would have been successfully explicated in terms of the resources of the functional calculus.

Now Tarski seems to claim as philosopher, that he has explicated the concept of truth (in L0) in terms of purely logical resources. From my point or view, this would have to mean that his account is circular (i.e. presupposes and does not explicate the concept 'true (in L)' for variable L), unless he can substantiate the thesis that the background concept of satisfaction or truth of, or belonging to the denotation of which are used at the ground floor of the recursive account are themselves explicable in terms of the resources of the functional calculus.

On the other hand, if without advancing any such thesis about the background concepts, he is simply using these background concepts to give a (recursive) extensional exposition of 'true sentence (of L0)' or, alternatively, a (recursive) explication of the contrived extensional predicate 'true-sentence-of-L0', then he is indeed not guilty of circularity, for he has simply abandoned the philosophical claim to be explicating the intention of 'truth (in L0)', let alone 'truth (in L)' for variable L.

Now I am not clear where you stand with respect to all this (given that you find what I am saying intelligible). On the one hand you seem to agree with me that (and I apply your remarks in the second paragraph of your letter to the case of truth) that "any adequate account of [truth] in general would have to refer to 'mental or behavioral or cultural factors' (Quine)" and applaud my attempts to do so in terms of rule governed linguistic behavior (page 2, paragraph 2 ff.). On the other hand, in the paragraph which begins at the bottom or page 1, you seen to go along with the idea that our background concept of satisfaction can be explicated in terms of the pattern

Sigma (x,y) =
x = expression-1 and y = item-1 or
If you can make sense of this, perhaps you can, as usual, help me say it better. In any event, the above remarks set the stage for the main purpose of this letter, which is to agree with you that the account which I give in my original letter of a 'pragmatic' sense in which

(1) Carnap is wise


(2) It is true that Carnap is wise

does not explicate the proper sense in which (1) implies (2) and, for that matter (2) implies (1).

In other words, I agree that the account in my letter is incomplete and, which is worse, directs attention away from the correct account of the connection between the meaning of truth and the criteria of truth -- in the sense of truth conditions, not evidence. Since I think that I am clear in my own mind about this, I am chagrined that my letter messed this up.

If letters could have second editions, I would replace the material from p. 31, second paragraph, to p. 34, paragraph ending at the top of the page, by the following text.[editor's note: These passages have been marked.] The deleted material with an appropriate introduction, could then be added as a relatively uninteresting appendix.

The problem is most pressing with respect to basic sentences. Yet to fix our ideas (as Nagel says) it is worthwhile to look at a less problematic case. In what sense does 'p or q' imply 'it is true that p or q'? Or, to translate this in terms of my analysis, but omitting dot-quote niceties, in what sense does 'p or q' imply ' 'p or q' is semantically assertible (with respect to relevant rules)'?

The structure of the answer is given by the recursive specification of truth conditions. In this respect my account does not differ from that of "Carnap-Tarski semantics." With respect to the sense of the answer, however, there is a difference which, though it may seem minor at this stage in the recursive structure, is of a piece with my account of the sense in which basic statements correspond to the world, and my critique of Tarski's explication of this correspondence in his classic paper on "The Semantic Conception of Truth."

As far as structure, then, is concerned, consider the following schema (where 'p' and 'q' represent basic sentences in the language we use, and ''p'' and ''q'' represent designations of these sentences in the metalinguistic level of this language {3}):

1. p --> 'p' is true Truth Cond.: basic sentences.
2. q --> 'q' is true Truth Cond.: basic sentences.
3. p or q --> 'p' is true or 'q' is true (1), (2): Prop. Calc.
4. 'p' is true or 'q' is true --> 'p or q' is true Truth Cond. 'or'
5. p or q --> 'p or q' is true (3), (4),: Tans. -->.

This derivation has two distinctive features. The first is that premises (1) and (2) have the same form as the conclusion to be derived, i.e. the form

. . . --> '. . .' is true

This feature is not surprising, for it simply illustrates the fact that we are dealing with a recursive specification of truth conditions (conditions of semantic assertibility). The second feature, however, is of greater interest, for it involves the distinction between semantic assertibility (truth) and a criterion or a ground of such assertibility. Thus step (4) is an appeal to the principle that a disjunctive statement is semantically assertible iff at least one of its disjunct is semantically assertible. And it is of the utmost importance to realize that in the equivalence

'p or q' is SA <--> 'p' is SA or 'q' is SA

the left hand and right hand sides do not have the same sense. For the LHS authorizes the inscription of 'p or q's, whereas the RHS is a disjunction of an authorization to inscribe 'p's and authorization to inscribe 'q's.

A possible source of confusion arises most clearly in the case of conjunction. Thus consider

'p and q' is SA <--> 'p' is SA and 'q' is SA

Here, again, the equivalence is not an identity of sense, for the LHS authorizes the inscription of 'p and q's whereas the RHS is a conjunction of an authorization to inscribe 'p's and is an authorization to inscribe 'q's. But is there not also the rule of Conjunction Introduction

From 'p' and 'q' to infer 'p and q'

according to which, if we are entitled to inscribe a 'p' and entitled to inscribe a 'q' we are entitled to inscribe a 'p and q'? The RHS of the above equivalence authorizes us to inscribe as follows

. . . .
. . . .
. . . .
. . . .

Conjunction Introduction authorizes us to add, in each case, a 'p and q,' thus

p,q so, p and q
p,q so, p and q
. . . .
. . . .
. . . .
. . . .

Consequently, it is a fact of the meaning of logical words that the authorizations to inscribe 'p's and 'q's carry with them "as a matter of sense" an authorization to inscribe 'p and q's. This is quite true. It does not follow, however, that the authorization to inscribe 'p and q's consists in a conjunction of an authorization to inscribe 'p's and an authorization to inscribe 'q's. The situation is rather that Conj. Int. (and its converse Conj. Elim.) give the sense of 'and,' and exhibit conjunction as a purely 'linguistic' or syntactical concept. Parallel points can be made about disjunction and negation. (Of course, a careful discussion of the semantics and syntactics of connectives would have to take into account the possibility of non-classical connectives.)

Before turning our attention to the distinction between the sense of 'true' as applied to basic sentences and the conditions of its applicability, we should note that the above derivation of

p or q --> 'p or q' is true

can be matched with a derivation of the converse hypothetical,

'p or q' is true --> p or q

it goes as follows:

1. 'p or q' is true --> 'p' is true or 'q' is true Truth Cond.: 'or'
2. 'p' is true --> p Truth Cond.: basic sent.
3. 'q' is true --> q Truth Cond.: basic sent.
4. 'p' or 'q' is true --> p or q (2), (3), Prop. Calc.
5. 'p or q' is true --> p or q (1), (4), Trans. -->.

Steps (2) and (3) remain, of course, to be discussed, as do steps (1) and (2) in the converse derivation. My present purpose, however, is to emphasize that the fact that this derivation establishes that

'p or q' is true --> p or q

should not tempt us to overlook the special connection between ''p or q' is true' and 'p or q' which is not a matter of implication, and which pertains to the sense of 'true.' Thus, although the sequence

'p or q' is true
so, p or q

can be viewed as an enthemyme with the suppressed premise

'p or q' is true --> p or q

or as an inference licensed by the principle

From 'p or q' is true' infer 'p or q'

which the above derivation would justify, it can also be viewed not as an inference, but as a sequence in which the statement that a certain kind of action is correct is followed by a performance of that kind of action. The latter way of viewing the sequence takes into account the sense of the predicate 'true,' i.e. (on my analysis) 'semantically assertible.'

The distinction between the truth move and the related inference is bound up with the distinction, stressed above, between the meaning of 'true' ('semantically assertible') and the recursive specification of truth conditions. It is because, as I see it, the truth conditions for basic sentences consist in matter-of-factual relations between linguistic and non-linguistic items, by virtue of which the former picture the latter, that my account of truth breaks out of the family of semantical concepts (true, true of, satisfies, belongs to the extension of, etc.) and brings in pragmatic considerations pertaining to the causal involvement of language users in the natural order.

Two points should be carefully noted: (1) By 'relations' in 'matter-of-factual relations' I mean relations. That is to say, I regard the distinction between relations and connectives as of the utmost importance to sound philosophy. Thus, when I deny that 'means' stands for a relation, it is because I analyze

'dreieckig' (in G) means triangular


The 'dreieckig' (in G) is a .triangular.

i.e. as

'dreieckig's (in G) are .triangular.s

Obviously it would be obtuse to treat 'are' as a relation word or predicate. A similar point is to be made with respect to 'stands for' as in

'dreieckig' (in G) stands for triangularity.

This I 'regiment' by the following steps

The 'dreieckig' (in G) stands for triangularity
'dreieckig's (in G) stand for triangularity
'dreieckig's (in G) stand for the .triangular.
'dreieckig's (in G) are .triangular.s

With 'denotes' the same strategy yields the sequence

The 'Vernunftiges Tier' (In G) denotes Featherless Biped
(E s) 'VT' (in G) stands for s, and s is materially equivalent to .FB.

Where 's' is a predicate variable taking dot-quoted expressions as its primary substituends. Regimented as above, this becomes

(E s) 'VT' (in G) C s and s ME .FB.

which is verified by

'VT' (in G) C .RA. and .RA. ME .FB.

of which the right hand conjunct tells us (in accordance with my convention of using '.IND [epsilon] RA.' as a predicate which applies to tokens consisting of an IND concatenated with an .[epsilon] RA.)

(IND) .IND [epsilon] RA. is true iff .IND [epsilon] FB. is true

which, in turn, is equivalent{4} to

(x) x [epsilon] RA --> x [epsilon] FB.

The second point to be noted, (2), is that if by relations I mean relations, it should be clear that by

sentence affirming a matter-of-factual relation

I don't mean simply

Contingent multi-term sentence.


Man C Featherless Biped

is a contingent sentence involving the two terms 'Man' and 'Featherless Biped,' but it does not, in my terminology, affirm a matter-of-factual relation between men and featherless bipeds. By parity of reasoning, the facts that

'dreieckig' (in G) means triangular


'Vernunftiges Tier' (in G) denotes Featherless Biped

are contingent truths should not lead us to say that they assert the obtaining of matter-of-factual relations.

This letter is designed to lay the foundation for a discussion of picturing. Thus I hope to take up in my next [letter] your question concerning 'pictures-in-L' for variable L. I must also satisfy you that it follows from my account of picturing that (leaving dot-quote niceties and tenses aside)

"a"s which are concatenated with an "f" [f*"a"s] picture a (correctly) iff fa

(not, as you put it,

"f"s picture a iff fa).

If I can pull this off by our satisfaction, I will take a vacation.

As ever,

Wilfrid Sellars


1879 HALL

December 9, 1970

Professor Wilfrid Sellars
Department of Philosophy
University of Pittsburgh
Pittsburgh, Pennsylvania 15213

Dear Wilfrid:

Three comments on your letter of November 20.

(1) According to Tarski, a general notion of truth applicable to any language leads to paradox. So he defined only truth in L0 for a particular language L0. One might extend his account to any language L' translatable into L0 by counting a sentence in L' true just in case its translation is true in L0. It is this extension that I supposed you would find circular. It makes use of a notion of translation; and I take it that you think the notion of truth figures in the explanation of translation.

(2) What justifies your assumption that "p or q" is SA if and only if "p" is SA or "q" is SA? Tarski can make use of an analogous assumption because in his definition of satisfaction he has a clause saying that a sequence satisfies a disjunction just in case it satisfies at least one disjunct. But your truth definition has no such clause.

This is no mere technical problem. For mathematical statements you equate semantic assertability with provability. But it is not in general true that a disjunction is provable if and only if one of its disjuncts is provable.

(3) A similar question can be raised about the justification of the truth condition cited for "and": "p and q" is SA if and only if "p" is SA and "q" is SA. In your letter you find a connection between this truth condition and the rules of conjunction introduction and conjunction elimination. Elsewhere, in your discussion of induction, you have rejected the rule of conjunction introduction. Does that mean you also reject the truth condition "and"?


Gilbert H. Harman



{1}Donald Davidson, "True to the Facts," The Journal of Philosophy, 66, 1969, pp. 748-64.[Back]

{2} T is the condition that "ones theory of truth should entail, for every sentence of the language in question an instance of the . . . formula

x is true iff p

with a name of the sentence replacing x and with a translation of that sentence replacing p." (p. 13)[412-13][Back]

{3} Notice that by not taking into account the internal structure of 'p' and 'q,' I am abstracting from essential features of the specification of truth conditions for basic sentences. These features concern the subject-predicate structure of basic sentences, to which reference must be made in characterizing the picturing relationship which defines the truth conditions for such sentences. [Back]

{4} I realize, of course, that this bald assertion of equivalence raises all the issues involved in the controversy over 'substitutional' versus 'objectual' quantification. [Back]

{*} edited in hypertext by Andrew Chrucky
Editor's note: The correspondence, which I probably obtained from either Sellars or Harman -- though I don't remember from which one of them, consists of five items:
(1) a 20 page (double-spaced) typescript of Harman's essay, "Sellars' Semantics," (subsequently published in Philosophical Review 79 (1970) 404-419), which is a critical review of Sellars' Science and Metaphysics: Variations on Kantian Themes (New York: Humanities Press, 1968);
(2) Sellars' letter to Harman, which is a 40 page (double-spaced) critique of Harman's essay;
(3) a 3 page (single-spaced) reply by Harman;
(4) an 11 page (double-spaced) rejoinder by Sellars; and
(5) a final 1 page (single spaced) retort by Harman.

Within my replication of the correspondence, next to the references to Harman's 20 page typescript of his essay (and some other places), I have inserted in square brackets the corresponding pages of the published version.

Because of the limitations of hypertext, substitutions of characters and designs have to be used. For Greek characters I simply use their names. The summation sign expressed by an upper case sigma is rendered by the word "Sigma", while the epsilon of set membership is expressed by "[epsilon]". The existential quantifier (inverted E) is rendered by a bold upper case "E"; material implication is expressed by a bold "-->"; set inclusion, by a bold upper case "C"; and where "L" has the subscript "0", this is rendered as "L0"; the triple bar of material equivalence is rendered by a bold "<-->"; and dot quotation is expressed by bold periods. All underlining and names of books and journals have been italicized. Minor grammatical and spelling mistakes have been corrected.

I wish to thank Professor Harman for his permission to let me publish his correspondence on the Internet.[Back]

[Contents][Return to Main Page]