1. Introduction

As is well known, Rudolf Carnap in his “Intellectual Autobiography” notes that in his philosophical discussions with friends he is prone to adopt differing philosophical views depending on who he was talking to, “adapting [himself] to their ways of thinking and speaking” (Reference Carnap and Schilpp1963, 17). For example, with one friend he might be a realist, with another an idealist. As he continues, some friends apparently called him on this, pointing to his inconsistent philosophical attitude, and when pressed to clarify his position he would remark that his view is “closer to that of physicists and those philosophers who are in contact with scientific work” (17–18). On further reflection, Carnap realizes that in fact he is neutral with respect to these differing philosophical forms of language, a neutrality based as he explains on the principle that “everyone is free to use the language most suited to his purpose.” He coins this ‘principle of tolerance’ in his Logical Syntax of Language (LSL; Carnap Reference Carnap and Smeaton1934/1937). This neutral attitude, he avers, “has remained the same throughout my life” (18).

Friedman (Reference Friedman, Floyd and Shieh2001, 228–35) traces the evolution of the principle of tolerance in Carnap's thought, starting from Carnap's doctoral work in the early 1920s, through the Aufbau period (late 1920s and early 1930s), subsequently to LSL, and up to “Empiricism, Semantic and Ontology” (ESO; Carnap Reference Carnap1956). In detail, as Friedman's discussion makes evident, Carnap's neutral attitude has not always stayed the same. In his doctoral work, Carnap expresses neutrality regarding differing notions of space: one can interpret ‘space’ in different ways—the “formal, intuitive and physical” (Friedman Reference Friedman, Floyd and Shieh2001, 228; italics removed)—and dependimg on the interpretation adopted, one can be led to formulate alternative foundations for geometry. Nevertheless, no one interpretation is said to have priority over the others; each is said to contain “part of the truth” (229). The Aufbau in turn expresses neutrality in two different ways. First, it is neutral with respect to different possible bases for constitutional systems. One can legitimately start with an autopsychological, heteropsychological, physical, or even a cultural basis to a constitutional system—which basis one prefers depends on one's methodological perspective. Second, the Aufbau expresses neutrality about the different epistemological schools of realism, idealism, and phenomenalism—Carnap expresses no preference for any of these metaphysical positions. Subsequently, as Friedman describes the matter (229–30), Carnap's adoption of the logicist framework that underpins the Aufbau itself became a target for liberalization in light of the foundational debates in the late 1920s, and as a result in LSL there is a further move to liberalizing even one's choice of logic. As Carnap expresses one version of the principle of tolerance, “in logic, there are no morals” (Reference Carnap and Smeaton1934/1937, sec. 17, 52); that is, there is no a priori restriction on what foundation for logic or mathematics (intuitionist, formalist, or logicist) one can adopt. When we turn to ESO, Friedman asserts, we find a kindred liberality, this time in terms of a choice of linguistic frameworks: in choosing a language, we are faced with an ‘external’ question that is answered in practical terms, just as the choice of syntax or logic in LSL is practically motivated (Friedman Reference Friedman, Floyd and Shieh2001, 235). Now, despite these variations, we can say that in general terms Carnap's attitude of tolerance is consistent—with regard to a variety of philosophical issues, Carnap reserves judgment and refuses to take sides. As he recalls in “Testability and Meaning” (TM; Carnap Reference Carnap1936/1937), as a member of the Vienna Circle his reservation was due to his considering numerous philosophical alternatives to be simply meaningless (5). With TM his view moderates to simply regarding alternative philosophical viewpoints as different linguistic ‘proposals’ that possess differing degrees of practical value (Friedman Reference Friedman, Floyd and Shieh2001, 232–33). This more modest view persists in Carnap's work until at least ESO.

A key issue we will be considering here is where one draws the line between those claims that have no cognitive status other than as ‘proposals’ and about which our choices are only practically motivated and those claims that are legitimate assertions. Carnap's response to this issue, as I will argue, is to draw the line on empirical grounds, an answer that is contrary to the received view on how Carnap applies the principle of tolerance, which regards that principle as binding independently of any empirical considerations. I plan to defend my interpretation of Carnap by examining, in the main, relevant passages in LSL, TM, and ESO that I believe make it clear that it is on empirical grounds that Carnap decides to be neutral or tolerant on certain theoretical issues. In effect, Carnap's position is at once tolerant and conservative: where empirical facts are uninformative, one is free to offer theoretical proposals regarding syntactical usage or the choice of a linguistic framework on practical grounds; however, where empirical facts have some bearing, one is to that extent forced to withdraw one's tolerant attitude.

The launching pad for my investigation will be discussions by Thomas Ricketts and Michael Friedman of an objection by Kurt Gödel to Carnap's conventionalism. Ricketts and Friedman independently argue that Gödel's objection misrepresents Carnap by wrongly attaching to him advocacy of an empirical criterion in the application of the principle of tolerance. I argue in response that Gödel correctly represents Carnap's conventionalism and that the reasons Ricketts and Friedman provide for resisting Gödel's depiction are misleading. Later on, I discuss and reject some objections to my empiricist interpretation of Carnap. In the end, I hope to gain some clarity on how Carnap proposes to apply the principle of tolerance.

2. Gödel's Critique

Ricketts (Reference Ricketts, Clark and Hale1994, Reference Ricketts, Friedman and Creath2007) and Friedman (Reference Friedman, Floyd and Shieh2001) cite and respond to the following objection raised by Kurt Gödel to Carnap's application of the principle of tolerance to the philosophy of mathematics. According to Gödel's interpretation of tolerance, one decides among various candidate foundational positions regarding mathematics, such as between logicism, intuitionism, or formalism, by conventionally choosing one or another of these positions subject to an ‘admissibility requirement’. As Ricketts explains,

To show the admissibility of these transformation rules, we need to show the consistency of the language generated by such rules, specifically, the consistency of its mathematics, for from an inconsistent language all and any sentences can be derived, including those expressing empirical facts. However, Gödel's second incompleteness theorem dictates that showing the consistency of a mathematics requires the assumption of a mathematics in the metalanguage that is itself consistent and of equal or greater strength. It follows, as a result of the admissibility requirement, that the convention of choosing a language system leads one into a vicious circularity—the consistency of this system can only be shown by assuming its consistency. The conclusion is that it is not possible for there to be an adequate, that is, noncircular, conventional choice of a language system that abides by the admissibility requirement.

In response, Ricketts and Friedman independently aspire to defend Carnap from Gödel's circularity objection by questioning Gödel's assumption that Carnap is allied to the admissibility requirement. For example, Ricketts remarks that “the requirement of an admissibility proof … rests on a language-transcendent notion of empirical fact, of empirical truth and falsity, a notion that is not tied to a particular calculus” (Reference Ricketts, Friedman and Creath2007, 210; see also Ricketts Reference Ricketts, Clark and Hale1994, 180; Friedman Reference Friedman, Floyd and Shieh2001, 241). But as Ricketts and Friedman interpret him, Carnap rejects any such notion of an empirical fact and so rejects the correlative admissibility requirement. Here is how Ricketts explains Carnap's rejection. To begin with, adopting “such a language-transcendent notion of empirical fact imposes constraints on the choice of a calculus, making some choices incorrect because those choices stipulate sentences to be true which are in fact false” (Reference Ricketts, Friedman and Creath2007, 210). Yet, Ricketts contends, because of the principle of tolerance, Carnap's position is that there are not any epistemic constraints on the choice of a language system that depend on the truth or falsity of any assertions—there are only pragmatic constraints that deem a language choice as valuable for this or that practical aim. Thus, the principle of tolerance effectively undermines the admissibility requirement (Ricketts Reference Ricketts, Clark and Hale1994, 180; Reference Ricketts, Friedman and Creath2007, 207). Friedman's approach, by comparison, emphasizes that for Carnap one determines which assertions count as analytic and which count as synthetic (i.e., empirical) by setting up a language system. As Friedman remarks, “the conventional statements relative to a given formal-logical system are just the sentences that are analytic relative to this system; the empirical statements relative to a given formal-logical system are just those sentences that are synthetic relative to this system” (Reference Friedman, Floyd and Shieh2001, 228; see also 241). It follows that there is no prelinguistic or language-transcendent notion of an empirical fact on hand to guide one's conventional choice of a language system, such as would be needed in applying the admissibility requirement. Accordingly, for both Ricketts and Friedman, Gödel's requirement of admissibility is asserted to be an unfaithful representation of Carnap's conventionalism, and thus Carnap's conventionalism is insulated from Gödel's critique.

My plan here is to argue that Carnap does (and should) subscribe to the fundamentals of Gödel's admissibility requirement and that in fact he would (and should) embrace Gödel's rejection of a pure conventionalism that seeks to delineate, a priori as it were, the distinction between the empirical and the analytic. To begin with, it is worthwhile commenting in support of Gödel's admissibility constraint. For example, it sounds right to require that logically valid truths do not have any empirical implications, for if they did, they could hardly be called logically valid since they would then be empirically testable. In a sense, then, we have a partial test for the logical validity of a claim, to wit, does it have any empirical implications? If so, there is no logical validity to be had. But still, does an advocacy of Gödel's requirement not land us squarely in the circularity problem cited above? One response here, suggested by Awodey and Carus (Reference Awodey, Carus, Awodey and Klein2004), is that there is no need to show that a language is consistent in order for such a language to be admissible—it need only be the case that a language is consistent and thus is admissible (206–9). An alternate answer might be that, although there is a need to show that a stipulated language is consistent on pain of its entailing all sentences, it is not only noncircularly impossible to satisfy this requirement—it is also practically unachievable for any but the simplest languages. Particularly, the proponents of scientific theories would surely demand that their theories be consistent, but they nevertheless recognize that providing such consistency proofs would be elusive given the wide scope and explanatory depth of most worthwhile theories. Thus, the absence of a demonstration of the consistency of a language (or scientific theory) is not an unexpected result and is a conundrum we are faced with, Gödel's result or not.

But let us step back a moment and consider what LSL has to say about how we determine that we are faced with a logical question (i.e., a syntactical question) as opposed to an object question (a question about objects in the world). Carnap, in section 74 (286), distinguishes (1) object sentences, (2) pseudo-object sentences, and (3) syntactical sentences. Object sentences, strictly speaking, pose no problems for the logic of science since they are “concerned with things which … occur in the empirical sciences” (Carnap Reference Carnap and Smeaton1934/1937, sec. 74, 278). Nor do syntactical sentences pose any problems: they are simply sentences phrased in the ‘formal mode of speech’ referring to the use of linguistic objects such as terms and sentences. More problematic are pseudo-object sentences, sentences phrased in the ‘material mode of speech’ that appear to refer to “extra-linguistic objects” (sec. 78, 298), such as numbers and things, but in fact refer (more formally) to linguistic items such as number words and thing words. Note that the problem with pseudo-object sentences is not so much that they refer to extra-linguistic objects (object sentences do that unproblematically) but that they do so illegitimately. The question is what we mean here by ‘illegitimately’.

Suppose, for example, we have a candidate ‘P-primitive sentence’, where such a sentence stands for a fundamental physical law (Carnap Reference Carnap and Smeaton1934/1937, sec. 82, 316). This sentence will presumably make reference to objects—so is it an object sentence or only a pseudo-object sentence? If the latter, we need to change it from the material mode of speech to the formal mode. If the former, no such translation is required since as Carnap explains, “translatability into the formal mode of speech constitutes the touchstone for all philosophical sentences, or, more generally, for all sentences which do not belong to the language of any one of the empirical sciences” (sec. 81, 313; italics removed). That is, the formal mode of speech is unneeded for object sentences since they stand as legitimate scientific sentences. So how can we tell when a candidate P-primitive sentence is a pseudo-object sentence? Here is Carnap's answer in LSL: “a newly stated P-primitive sentence is shown to be a pseudo-sentence if either no sufficient rules of formation are given by means of which it can be seen to be a sentence or no sufficient rules of transformation [are given] by means of which it can … be submitted to an empirical test” (sec. 82, 322). In other words, a P-primitive sentence is a pseudosentence if it is ill formed or untestable. Conversely, a well-formed sentence counts as an object sentence if it is empirically testable. So suppose we find some P-primitive sentence to be untestable. Then we must suppose that its constituent terms do not refer to any empirical objects, although we might be led to think they do, if we persist in using the material mode of speech. Consequently, we need to formulate this sentence in the formal mode of speech to cure us of the tendency to suppose that the terms in this sentence are genuinely referring.

A similar, empirical approach to determining what questions count as pseudoquestions occurs in ESO. In that paper, Carnap considers the question of whether there really are numbers, and about those philosophers who purport to take this as a serious ontological question, he says, “unfortunately, [they] have so far not given a formulation of their question in terms of the common scientific language. Therefore our judgment must be that they have not succeeded in giving to the external question and to the possible answers any cognitive content. Unless and until they supply a clear cognitive interpretation, we are justified in our suspicion that their question is a pseudo-question, that is, one disguised in the form of a theoretical question while in fact it is non-theoretical” (Reference Carnap1956, 209). What does Carnap mean by a “common scientific language” that can “supply a clear cognitive interpretation” for a question? Consider, for example, what Carnap says about what constitutes ‘understanding’ a symbol in physics: such understanding will be generated by providing an interpretation for this symbol “given directly by semantical rules referring to elementary signs together with the formulas connecting them with [this symbol]” (Reference Carnap1939, 210), where the meanings of “elementary signs” relate directly to the observable properties of things (204–5). This sort of empiricist semantics clearly underpins Carnap's work in ESO in which his stated concern is to reconcile language that makes reference to abstract entities with “empiricism and strictly scientific thinking” (Reference Carnap1956, 206). Thus in ESO, as in LSL, a question becomes a pseudoquestion if answering this question is detached from empirical data and so lacks cognitive content, strictly speaking.

Let us now return to Gödel's admissibility requirement for logically valid sentences, which demands that such sentences be logically independent of “the truth or falsity of any proposition expressing an empirical fact.” We have just seen that pseudosentences for Carnap have this feature—this is what makes them ‘pseudo’ sentences. Qua pseudosentences, Carnap's view is that they are more properly phrased in the formal mode of speech, by means of which they become ‘syntactical sentences’, sentences in the ‘logic of science’, and that such sentences are the concern of philosophical investigation. As it happens, philosophers have substantial liberty concerning which syntactical sentences to adopt if they wish to construct a language (as opposed, alternatively, to describing an extant language), as per the principle of tolerance, and their choices here are solely motivated by practical concerns (see Friedman Reference Friedman, Floyd and Shieh2001, 232–33; where the task is to describe an existing language, one is obviously constrained by the need for descriptive accuracy). So what would happen if a syntactical question, the answer to which is the stipulation to use language in a certain away, turned out to have empirical ramifications? It would turn out that an answer to this question could no longer be construed as a logically valid sentence (with respect to the language constructed on its basis), and this sounds like exactly the sort of result we are looking for.

So why do Ricketts and Friedman resist this result? Let us recall their reasons from above. Ricketts’ concern is that, with admissibility, a “language-transcendent notion of empirical fact imposes constraints on the choice of a calculus,” thus violating tolerance. But here it is worth pointing out that tolerance is only a constraint on logic. As Carnap says in LSL, “in logic, there are no morals” (Reference Carnap and Smeaton1934/1937, sec. 17, 52), but there certainly are ‘morals’ when it comes to empirical science, specifically, our duty to adjust our theories in the face of empirical facts. So let us suppose some candidate logically valid sentence (contained in a calculus) is found to have an empirical implication—that is, it implies the truth or falsity of some proposition expressing a (language-transcendent) empirical fact. It follows by Gödel's admissibility requirement that the candidate sentence is not logically valid. (Indeed, it is not a sentence of logic at all but one that concerns empirical issues.) It is in this fashion that the choice of a calculus is empirically constrained. However, it does not follow that we have found a violation of tolerance here, for tolerance is irrelevant when it comes to empirical matters. The applicability of tolerance is preserved since it only applies to those cases in which the admissibility requirement is satisfied and we have, accordingly, a genuine question of logic.

So what of Friedman's reason for resisting Gödel's admissibility constraint? Can we really not talk about empirical facts outside of a system of language? In what way can we talk about empirical facts, or about the truth or falsity of propositions expressing empirical facts, without presupposing a prior choice of a linguistic framework that sets out what is to count as an empirical fact in the first place?

When Carnap describes in ESO the sorts of questions philosophers are prone to ask (e.g., Are numbers real? Are there propositions?) as external questions, he means that by adopting a language form, say, one stipulating the reality of numbers, we should not be taken to be asserting the existence of numbers. Adopting a language form for him is a decision to speak in a certain way, a proposed language, not the assertion of the reality of this or that entity. However, once this language form is adopted and we are in the practice of discoursing in the way stipulated by our chosen linguistic proposal, there arise legitimate questions of existence, internal questions, that do have true or false answers and that are decidable, if the questions are not simply logical, by empirical inquiry (Carnap Reference Carnap1956, 207). This empirical inquiry, as we indicated above in our discussion of what Carnap means by a “common scientific language” that supplies “a clear cognitive interpretation” for a question, relates technical symbols to elementary signs by means of semantical rules. Clearly, then, the crux to answering an (empirical) internal question is to identify the observation reports that form the testing basis for theoretical questions in science and that contain elementary signs as components. How then does one go about identifying such reports? As Ricketts explains, Carnap's approach in TM is to “[characterize] observation predicates in behavioral terms, not semantic terms,” and in this respect Carnap's account of observation reports, Ricketts asserts, coheres with (what Ricketts takes to be) Carnap's rejection of a “language-transcendent notion of empirical fact or truth” (Reference Ricketts, Friedman and Creath2007, 210; see also 205 and Carnap [Reference Carnap1936/1937, 454], for Carnap's account of observation predicates in TM).

At this stage, it is worthwhile spending some time emphasizing Carnap's behaviorist perspective on observation reports during the time of his writing LSL. For instance, Carnap describes observation reports as involving, simply, human verbalizations prompted under certain physical circumstances (Reference Carnap1932/1987, 459); as such, they are of a kind with a variety of repeatable physical processes. Carnap comments, “whether it concerns a machine which is so built as to given signals, or any other object exhibiting regular observable reactions under specified conditions, makes no fundamental difference. … Even here we can use the reaction as a signal and construct a system of rules for the translation of the signals into sentences of our language” (459). As such, the identification of observation reports is an empirical process and not at all a syntactical or logical process or even a conventional one. Consider further what Carnap says: “Logically speaking, every group of protocol sentences has equal rights with every other group. … The demarcation of the ‘real, true protocol sentences’ is certainly not possible in pure semantics, involving merely logical means, but in descriptive semantics with real science, namely, historical conceptions. ‘True protocols’ are described as spoken utterances or written records (i.e., physical-historical creations) originating from any person, especially those scientists of our culture circle” (Reference Carnap1932, 179–80; my translation; see also Hempel Reference Hempel1935, 57). This ‘historicist’ understanding of observation reports reaches into LSL: “it is not the task of syntax to determine which sentences of the established protocol form are to be actually laid down as protocol-sentences, for ‘true’ and ‘false’ are not syntactical terms; the statement of the protocol-sentences is the affair of the physicist who is observing and making protocols” (Carnap Reference Carnap and Smeaton1934/1937, sec. 17, 317; square brackets removed).

When we arrive at TM, Carnap reiterates the behaviorist character of observation reports and entertains the extension of these considerations to confirmation and testing as well:

The definition of ‘confirmable’ occurs shortly thereafter: “a sentence S is called confirmable (or completely confirmable, or incompletely confirmable) if the confirmation of S is reducible (or completely reducible, or incompletely reducible, respectively) to that of a class of observable predicates” (456; italics removed). Once we reflect on the fact that observation predicates are behavioristically defined, it follows from this definition that the confirmability of a sentence is defined in behavioral terms as well. So if we are concerned to assess a potential pseudo-object sentence and decide whether it is a genuine object sentence or only a pseudo-object sentence that deserves to be formulated in the formal mode of speech, what we can do is ascertain whether this sentence is linked to observation reports in such a way that these reports either confirm or disconfirm this sentence, and we can do this without necessarily interpreting either the sentence or the observation report through a language system. For example, suppose we hypothesize that “numbers are real” and find that neither the confirmation of this stipulation nor its negation is reducible to a class of observation predicates. It follows that it is a pseudo-object sentence since it contains content going beyond observational testimony.

At this point, the following objection arises: in assessing whether an observation report bears on a proposed object sentence, must not this report be expressible in a language? Moreover, must not this language contain transformation rules that would express the relevant logical connection between the observation report and the pseudo-object sentence subject to test? If we answer these questions affirmatively, it follows that we have not found the language transcendence we seek. Just as Friedman suggests, the notion of empirical testability would presuppose a language and with it criteria for when sentences count as either analytic or synthetic. As such, the sort of language transcendence we are ascribing to Carnap fails to hold.

It is of interest to note that a recognition of this problem, and a resolution sympathetic to Friedman's interpretation, can be found in Carnap (Reference Carnap1932/1987). In addressing the issue of how one might link observation reports (e.g., machine signals) standing outside a system language to sentences inside the language, Carnap comments: “the signals of the machine and the statements of the man are treated like sentences of a language in that translation rules are constructed for them. We call them therefore ‘protocol sentences’ of the ‘protocol language’ of the machine or of the foreign man and distinguish this language from the language of our system” (459). In other words, observation reports (protocol sentences) are located in an extrasystematic language containing translation rules that link them to sentences in the system language. Accordingly, such reports could hardly be said to be language transcendent.

Indeed, an analogous problem (needing an analogous solution) arguably afflicts the distinction presented in ESO between external and internal questions. In addressing the controversy between those who believe in the reality of numbers and their nominalist opponents, Carnap provides comments that on the surface resonate well with my proposed interpretation: “I cannot think of any possible evidence that would be regarded as relevant by both philosophers, and therefore, if actually found, would decide the controversy or at least make one of the opposite theses more probable than the other. … Therefore I feel compelled to regard the external question as a pseudo-question, until both parties to the controversy offer a common interpretation of the question as a cognitive question; this would involve an indication of possible evidence regarded as relevant by both sides” (Reference Carnap1956, 219). (Here “possible evidence” could be either empirical or logical, as Carnap notes; see 218). But how can one discuss, in language, an external question regarding the reality of numbers when the result of answering this question will be to set up a language system that guides us on how we are to talk about numbers? This problem can be put in very general terms as follows: what sense can be made of formulating in language an external question that, by definition, purports to address an issue that is language transcendent? The appeal of Friedman's approach that eschews language transcendence should be obvious.

However, there is an error in this line of thinking as an interpretation of what Carnap means when he suggests that a question is a pseudo-object, external question, the answer to which is a proposal to use language in a certain kind of way. Our suggestion is that answers to such questions must meet Gödel's admissibility condition in that they “do not imply the truth or falsity of any proposition expressing an empirical fact.” Yet this condition can be met simply by omission: it might be the case that there are no translation nor transformation rules at hand to negotiate any empirical implications, rather than there being such rules that, as it happens, fail to be instantiated with reference to actual fact. Put another way, with reference to nominalists and their opponents, we seek, Carnap says, evidence that would decide the issue between them, and finding none we call their dispute a pseudoproblem resolved by stipulation to use language in one way or another. What we mean, though, by ‘finding none’ is not that we have looked and have failed to turn up any evidence as yet—this would make the dispute an unresolved empirical matter—but that there are neither translation nor transformation rules in place that could mediate such empirical implications. Similarly, with regard to the general question of how to formulate language-transcendent external questions, they are questions that are answered by a pragmatically motivated stipulation to use language in a certain way, to set up ab initio transformation and translation rules; this is possible so long as there are no extant rules that are asserted to be binding, such as would be the case if there were empirical factors bearing on the matter. If there are not, the matter is turned over to the principle of tolerance, and we are free to use language in this way or that, depending on our orientation.

We are now in a position to answer Friedman's objection to Gödel's admissibility requirement. Gödel requires that the logically valid sentences of a calculus (of one's chosen language system) “not imply the truth or falsity of any proposition expressing an empirical fact.” Again, the worry is that to carry out this requirement one needs a prelinguistic or language-transcendent notion of an empirical fact, which is not to be had since it is by setting up a language that Carnap apparently determines which assertions count as analytic and which count as synthetic (empirical) to begin with. Moreover, because Gödel's admissibility requirement is found to be flawed in this way, Friedman asserts it to be an unfaithful representation of Carnap's conventionalism. In turn, we have found that Carnap's notion of an empirical fact is represented, in its first instance, not as an observation report defined semantically but rather as one defined pragmatically (i.e., in terms of human behavior). As such, observation reports can be used to assess the character of a proposed object sentence as either a pseudo-object sentence, if it lacks empirical implications, or an object sentence, if it does have empirical implications, and can do so without implying the truth or falsity of any empirical facts (as Gödel requires) since such reports, as defined, are uninterpreted. I argued earlier that there are good intuitive reasons to adopt Gödel's requirement; we can now say that there are good reasons to assert that Gödel's requirement is faithful to Carnap's project in the logic of science.

3. Critiques of ‘Carnapian Empiricism’

I have been arguing that the (philosophic) claims about which Carnap says we should be tolerant are those for which there is no empirical evidence, either for or against, and I have defended the reference to empirical evidence here on the grounds that such evidence can be formulated in behavioral terms not requiring semantic interpretation. Let us now consider some further critiques of my approach.

The first critique is derived from the Alan Richardson who advocates an approach very similar to the one advanced by Friedman. Richardson's perspective places heavy emphasis on the following passage from TM:

Reflection on this quote might lead one to question my thesis that, for Carnap, empirical factors bear on the determination that some sentence is a pseudo-object sentence (and so is best formulated as a syntactic proposal). For what is being suggested is that the ‘principle’ of empiricism is (itself) a syntactic proposal, and my contention is that it is by means of an empirical inquiry that one arrives at such a description of empiricism (i.e., the principle of empiricism is a ‘proposal’ if this inquiry reveals that this principle has no empirical consequences). As such, on my approach, one arrives at the decision that empiricism is a proposal only after having assumed empiricism, which seems circular. However, I am not convinced that there is a circularity here. Where empiricism states that all knowledge is gained through sensory experience, one can imagine finding empirical evidence for or against empiricism without having assumed that all knowledge is arrived at in this way. Put another way, the proposal, ‘empiricism’, is a much more profound sort of thing than the suggestion to try and find empirical evidence that bears for or against some claim—in no way am I assuming that for Carnap one needs to assume empiricism in arriving at the conclusion that empiricism is a proposal.

But there is another way one might take the above quote as challenging my ‘Carnapian Empiricism’, a way promoted by Richardson. In the above-quoted formulation of empiricism, it is stated that synthetic sentences must have some connection to observation. But for this formulation to make sense, one needs to presuppose a distinction between analytic and synthetic sentences. Thus, the analytic/synthetic distinction comes before empiricism, and it cannot be the case that one determines empirically the components of this distinction. Richardson expresses this point as follows: “we see here that an analytic/synthetic distinction is relied upon in giving content to the proposal to employ empiricist languages—we seek the suitable relation to experience only for synthetic sentences of a given language. It is clear that in this sense within Carnap's philosophy, a commitment to an analytic/synthetic distinction comes prior to a commitment to empiricism” (Reference Richardson1997, 157). Again, reflecting on the TM quote above, Richardson suggests that “[it] indicates how … analyticity serves to first clarify and make sense of empiricism. … Analyticity is not a dogma but a presupposition of empiricism for Carnap” (Reference Richardson, Awodey and Klein2004, 68).

For Richardson, this feature of Carnap's view sharply distinguishes it from Quine's perspective in which, with the assumption of empiricism, ‘analytic’ sentences are defined as those sentences that are impervious to empirical testimony (which, I might add, would speak to the poverty of analyticity since any claim can be deemed to have empirical import; see Richardson Reference Richardson1997, 153–57). However, my impression is that there is a different way to read the above quote from TM. We should not read it as starting with a notion of a synthetic sentence and then delimiting what counts as a synthetic sentence from the perspective of empiricism; rather, we should read it as defining what counts as a synthetic sentence and then correlatively defining what counts as an analytic sentence, to wit, as a sentence that lacks any connection with possible observations. There is an obvious appeal to such an approach since, alternatively, if one is to arrive at the analytic/synthetic distinction in advance, one is faced with the question of how this distinction is motivated. Is it completely arbitrary? What practical determinants are relevant or acceptable? Under what circumstances is the distinction legitimately revisable? These questions are easy to answer from the empiricist perspective I am attaching to Carnap. In addition, my interpretive approach coheres better with Carnap's clear allegiance to an empiricist's criterion of meaning in TM. As Carnap says, “it is not [my] aim … to defend the principle of empiricism against apriorism or anti-empiricist metaphysics. Taking empiricism for granted, we wish to discuss, the question what is meaningful. The word ‘meaning’ will here be taken in its empiricist sense; an expression of language has meaning in this sense if we know how to use it in speaking about empirical facts, either actual or possible ones” (Reference Carnap1936/1937, 2). Using such a criterion of meaning, a meaningful sentence (i.e., a legitimate object sentence) will be one that has empirical significance, and a meaningless sentence (a pseudo-object sentence) will lack empirical significance and thus best serve as a ‘syntactic proposal’. In this respect, there is found to be a clear synchrony in Carnap's views as expressed in LSL and TM. Finally, my approach provides a straightforward, and I think more traditional, perspective on the differences between Carnap and Quine on the issue of analyticity: they simply disagree on whether there are, in fact, sentences that lack empirical import. It is the key and contentious burden of Quine's “Two Dogmas of Empiricism” to show that there are no such sentences.

As a final point to be made about the above quote from TM, one might express the concern that Carnap, in expressing empiricism as a proposal, seems committed to denying that empiricism itself has any empirical significance. Presumably, such a result should make empiricism highly unattractive to the empiricist. Yet there are other reasons why an empiricist might nevertheless avow a proposal such as empiricism, and one very clear reason Carnap gives is his interest to capture “the actual practice of science” (Reference Carnap1936/1937, 20; see also 26). That is, although empiricism as a proposal is on an epistemic par with every other proposal, there is for Carnap a clear practical reason for preferring empiricism, which is that such a view is more in line with the views of “physicists and those philosophers who are in contact with scientific work” (Reference Carnap and Schilpp1963, 17–18).

The second critique of ‘Carnapian Empiricism’ I will consider derives from comments Carnap himself makes in ESO and LSL, comments that might be taken to run counter to the empiricist interpretation I am suggesting. In ESO, when considering the dispute between realists and subjective idealists, Carnap notes that the controversy has gone on “for centuries without ever being solved” (Reference Carnap1956, 207). In then diagnosing the problem, he comments, “[the dispute] cannot be solved because it is framed in a wrong way. To be real in the scientific sense means to be an element of a system; hence this concept cannot be meaningfully applied to the system itself. Those who raise the question of the reality of the thing world itself have perhaps in mind not a theoretical question as their formulation seems to suggest, but rather a practical question, a matter of a practical decision concerning the structure of our language” (207). In other words, the realist/subjective idealist debate has reached a stalemate, Carnap seems to be suggesting, not because of the empirical undecidability of the issue but because the issue is illegitimately phrased. We are asking, “Is the external world real?” in a sense that cannot be formulated in a language. As it were, we are standing outside language and asking about the reality of our conception of the external world. But how can one formulate a question, any question, outside a language? A language must be posited to begin with in advance of formulating a question, and once a language is posited it will contain in itself the resources for answering whether the external world is real. So, when it comes to the question of the reality of something, we may as well admit that the answer we are providing here is determined by our choice of language. Let us then be explicit about this fact and talk about the debate between realists and idealists as purely a matter of differing linguistic stipulations.

This is a compelling picture of how Carnap's conventionalism works, and it has a clear connection to Friedman's perspective on what is wrong with Gödel's admissibility constraint. Does it provide a useful interpretation of how the principle of tolerance is applied? If it did, tolerance would say something like what follows. Language systems can be set up using any number of linguistic ‘proposals’. To cite some of Carnap's examples, we can stipulate the existence of an external world or an idealistic mind-dependent world, a reality with numbers or without, a deterministic physics or an indeterministic physics, and so on. It is open to us to choose any one of these proposals, as guided by our practical aims, and the proponents of any one of these proposals can vindicate their approach by citing the reasoning in the above quote from ESO. That is, a believer in the reality of numbers can say that “to be real in the scientific sense means to be an element of a system” and can then assert the reality of numbers in the context of her system. To the critic standing outside her system, she can claim that the “concept [real] cannot be meaningfully applied to the system itself,” and thus the extrasystematic skeptic who denies the existence of numbers is offering an illegitimate criticism. Rather, the proponent of the real existence of numbers, as dictated by her own practical goals, is free to set up a language system of her own choosing that includes and sanctions her proposal that numbers are real.

However, my impression of this picture of Carnap's conventionalism is that it leads to an excessive liberality in acceptable language systems. We find ourselves confronted with an enormous plethora of language systems, all equal candidates for consideration, pending this or that practical motivation. Most any language system could be said to constitute the ‘logic of science’, if we are without a restriction on what constitutes a legitimate practical aim, and this seems to be a result that is at odds with typically conservative scientific practice. Nevertheless, it should be pointed out that Carnap in the above quote from ESO is in fact discussing the concept ‘real’, as opposed to the concepts ‘observable’ or ‘confirmable’. ‘Real’ is clearly a theoretical notion that has meaning only within a language system; to say that something is real, one needs an idea of its properties that would be described within this language system. Thus, there cannot be a completely extralinguistic notion of the ‘real’. However, observability for Carnap is a semantically neutral notion describing the behavior of people under certain circumstances and presupposes no particular language system (assuming we eschew the sort of perspective we saw Carnap [Reference Carnap1932/1987] express). To illustrate the difference, consider substituting the word ‘real’ with the word ‘observable’ in the above quote, giving us: “to be observable in the scientific sense means to be an element of a system.” This statement is manifestly not true for Carnap given the behavioral criteria he sets forth for observability. Similar comments apply for the word ‘confirmable’. As I argued above, the confirmability of a claim, for Carnap, can also be construed purely behavioristically and so independently of a language system. Thus, “to be confirmable in the scientific sense” does not mean “to be an element of a system.” In all fairness, we might acknowledge that to determine the truthfulness of an observation report or confirmed claim, one needs an interpretation in the context of a system. But for the purposes of confirmability, all we need is the identification of such reports and claims; as such, the assessment of a potential object sentence in terms of whether it is subject to empirical confirmation or disconfirmation can proceed outside the context of a language system, even if its interpretation and the assessment of its truth value cannot.

As it turns out, the interpretation of conventionalism one can draw from the above quote from ESO need not make reference to realism. Let us reformulate our objection to ‘Carnapian Empiricism’ on the basis of the above quote from ESO in the following way, taking into account remarks Carnap makes in LSL that “as regards those object-questions whose objects do not occur in the exact sciences, critical analysis has revealed that they are pseudo-problems” (Reference Carnap and Smeaton1934/1937, sec. 72, 278). This is a similar approach to the one portrayed in ESO, except that here something is the topic of a legitimate object question by occurring in a science instead of, as before, being real by occurring in a language system. Do we then have a workable, nonempiricist account of pseudo-object sentences? Whether this alternative approach works depends on what it means for an object to ‘occur in the exact sciences’. Suppose this means that scientists are committed to the existence of such things. Scientists, as we know, are committed to the existence of an external world, but of course the existence of the external world is not, for Carnap, the topic of a legitimate object question. So what else can it mean to ‘occur in the exact sciences’? On the criterion I suggest, for an object to occur in an exact science is for the existence of this object to have some degree of empirical support. For to be an ‘exact’ science, it is not enough that a set of beliefs be logically consistent and clear—if that were enough, all sorts of ludicrous sets of beliefs could count as exact sciences. Rather, we thus need some way to sort through the theoretical alternatives and pick out those that contain genuine object sentences, and my suggestion is that an empirical criterion is the obvious choice to do this task. So the renewed interpretation of the above quote from ESO would now go as follows: to be the topic of a legitimate object question means to be confirmable on the basis of observations. Of course, such a claim is precisely what I am defending here as an interpretation of Carnap.

I close the article by emphasizing the need for an empirical criterion in explaining a characteristic feature of Carnap's principle of tolerance, to wit, the fact that a choice of language is only expressible as a ‘proposal’—a recommendation to speak in a certain way—and does not stand as an assertion. For example, if I choose to be a realist and so stipulate, which I am allowed to do by tolerance, that “the external world is real” (and not a figment of my mind), this stipulation for Carnap cannot stand as an assertion that is either true or false. Yet, one would think, if I choose to adopt this stipulation as part of my language system, would the doctrine of realism become not only true for me (and so express an assertion) but necessarily true with respect to this language system since realism then expresses the ‘logic’ of this language? The answer to this query is straightforward, once we acknowledge an empiricist criterion of meaningfulness. For the linguistic proposals that are subject to tolerance are claims that have no distinguishing empirical feature as compared to alternative, competing linguistic proposals. For example, as regards the two alternatives of realism and idealism, these alternatives are empirically on a par—there is not, nor can there be, any empirical fact showing realism to be true and idealism false and vice versa. Thus, since there is no empirical evidence pointing toward either of these stipulated proposals, and given the empiricist criterion of meaning to which Carnap is allied, we are lead to conclude that each alternative lacks cognitive content (strictly speaking); neither claim can stand as an assertion, and each can only stand as a proposed language form. Comparatively, it is difficult to see without such an empiricist criterion on what grounds Carnap could defend the nonassertoric nature of linguistic proposals from within a language system.