THERE is no doubt that all our knowledge begins with experience, says Kant. He goes on to say that we should not conclude that all our knowledge is derived from experience. He argues that, in addition to the experimental source, there exists a rational source of knowledge, which supplies us with synthetic a priori principles. All his philosophy can be regarded as a commentary on this basic proposition.
Modem empiricism has shown Kant’s thesis to be fallacious. There is no synthetic a priori; what reason contributes to knowledge are analytic principles only. I will not elaborate on this counterthesis within this discussion group, within which the rejection of the synthetic a priori can be regarded as accepted. In particular, Professor Lewis has repeatedly made it clear that he shares this empiricist criticism of Kant’s views, and has made many valuable contributions to the purification of philosophy from the remnants of a philosophic rationalism. It is true that certain contemporary philosophers have attempted to reintroduce a synthetic a priori, reviving arguments that some forty years ago were quite fashionable; but their attempts are too obviously fallacious, and I cannot but regard their attacks on empiricism as a Putschversuch which need not be taken seriously. I will therefore proceed on the assumption that our discussion is to be grounded on a common empiricist basis; and the names of my partners in this dis cussion encourage me to take this common presupposition for granted.
If all knowledge not only begins with experience, but is also validated through experience, there arises the problem of whether there is a certain experiential basis of the elaborate structure of knowledge, a substratum which is composed of the experiential data alone and which carries the total edifice that includes so many results of highest abstraction. Speaking in terms of sentences, we speak here of report propositions. It is well known, and has often been discussed, that the derivational processes, leading from report propositions to laws of nature and predictions of future occurrences, include inductive inferences and therefore can only lead to probable conclusions; certainty is excluded when induction is at work. There remains the question whether such uncertainty also creeps into the experiential basis itself, or whether the basis, at least, is exempt from doubt. Can report propositions be absolutely certain? That is the question I should like to examine as my contribution to the present discussion.
This is the point at which the paths of empiricists diverge. Professor Lewis thinks that there are absolutely certain report propositions; and this view is shared by some other outstanding empiricists, such as Bertrand Russell and C. D. Broad. However, I do not think that this conception is tenable. Within an historical inquiry, I would argue that it is but another remnant of rationalism, taken over even by those who are willing to renounce the synthetic a priori, but who cannot dispense with a synthetic a posteriori which is as apodictic as the corresponding a priori. But historical classification is no logical argument; my thesis requires an analysis on logical grounds, and this is the subject to which I now shall turn.
For this purpose, I should like first to make the thesis of my opponents as clear as possible to me. The absolutely certain report propositions, of which Lewis thinks, are of course not conceived as referring to physical objects. In this respect, Lewis’ view differs fortunately from G. E. Moore’s opinions, according to which the simple statements concerning our daily environment are absolutely certain. Lewis shows convincingly that this cannot be maintained because later observations can always invalidate such statements. He speaks here of nonterminating judgments; they include in their meanings an infinite number of implications for the future and can therefore never be completely verified. In contrast, Lewis calls report propositions of the kind considered terminating judgments; their meaning is completely given in the observation to which they refer. They therefore do not include any implications for the future and can be completely verified in the act of observation. The language of these statements, Lewis insists, is not physical language. When I say, “There is a flight of stairs,” later observation can disprove my statement; this would be the case, for instance, if I find out that I cannot walk up these stairs because they are merely a perspective drawing by a clever architect. However, there is another statement, says Lewis, which remains true even if this further consequence is observed. It seems difficult to formulate this statement. Some of us would give it the wording: “I see a flight of stairs”; others would say that there is a flight of stairs, “optically speaking,” or “subjectively speaking”; others again would speak of a perceptual existence of the stairs. It seems we all know what we mean, and differ only as to the formulation.
The language of these statements Lewis calls expressive language. Using a more familiar name, I would prefer to call it phenomenal language. Its terms strangely duplicate those of physical language; but they do not have the same meanings. They do not, because they do not have the same implications; for instance, the phenomenal term “flight of stairs” does not include any implications concerning my later walking up stairs, whereas the physical term does. If we are aware of the peculiar nature of this language, Lewis maintains, we can understand why report propositions are absolutely certain; they merely express a momentary observation and are not subject to later revision.
Yet these statements are synthetic. They inform us about something that is there, and do not merely give labels to the individual observation. They tell us, for instance, that what there is now resembles something I saw yesterday. They draw lines of comparison into the picture offered by experience.
Here we have to be careful. We cannot directly compare what we see now with what we saw yesterday. We have a recollection image of what we saw yesterday, which we compare with what we see now. We cannot observationally compare the recollection image with what we saw yesterday. This comparison can only be made after certain co-ordinative definitions have been laid down, comparable to those known from space-time measurement, which make it a rule of language (an extension rule) to call the original object of sight similar to its recollection image. Yet there are, in addition, certain empirical statements involved. For instance, when we look at the same blue paper twice, we find that the visual aspect of the second observation is similar to the recollection image of the first. This constancy of the perceptual function1 does not always hold; we know that the subjective color of an object may change when the color of its surroundings changes, while the object retains its color in the physical sense (contrast colors). But, I think, Lewis would not include an hypothesis about the constancy of the perceptual function in his expressive, or phenomenal, language. He would regard it as already belonging in physical language.
Let us assume that this phenomenal language as proposed by Lewis can be carried through. Can we then maintain that there are absolutely certain report propositions, formulated in phenomenal language? I will try to collect the arguments for and against Lewis’ thesis of the existence of such propositions and, examining the pros and cons, attempt to arrive at a conclusion.
The first argument in favor of Lewis’ thesis stems from a consideration of the procedure by means of which we construct phenomenal propositions. When we discover that a report formulated in physical language is incompatible with later observations, we change its interpretation by reducing it to its phenomenal content; we say, for instance, “Though it was false to believe that there was a flight of stairs, it remains true that there was the phenomenon of a flight of stairs.” In other words, the method by means of which we introduce the phenomenal reports seems to guarantee their absolute certainty: we eliminate that component of the original sentence which is falsified by later observation, and which may be called its physical component; thus we construct its phenomenal component as a residue exempt from doubt. Phenomenal language is constructed by elimination of all physical implications; it therefore appears plausible that the sentences of this language cannot be disproved.
Second, if we reduce the sentence to its phenomenal content, it seems quite impossible that we are mistaken if only certain precautions are taken. Observation cannot lie, we would like to argue; a mistake can only spring from our interpretation of the observation. It is true that observers can lie; so we must make sure that lying is excluded. Although this condition may involve some difficulties if reports by other persons are concerned, it offers no difficulties regarding our own reports: we know whether or not we lie. Now it is also true that there are certain sources of error even for our personal language: we sometimes make mistakes in speaking, say “red” when we mean “green,” or “John” when we mean “Peter,” and therefore have to be sure that no mistakes of this kind occur. But we may argue that while we use the words incorrectly they have, at the moment we use them, the correct meaning for us; the word “John,” at this moment, has the meaning of “Peter” for us, and, though a slip of the tongue may misinform other persons, it cannot deceive us who know what we mean. In other words, errors can only spring from the linguistic formulation, but cannot concern its meaning as it is intended by the speaking person. So, if we take care that no error creeps into the linguistic formulation, report sentences are absolutely reliable.
The third argument is of a logical nature. It refers to the empiricist conception that statements about physical objects are merely probable, and proceeds by a sort of deductive reasoning. If something is claimed to be probable, so goes the argument, something else must be certain; for how can we arrive at probabilities unless we have some firm basis from which we can derive those probabilities?
Sometimes this argument is given a more mathematical form. We say, for instance, that the probability of the event is p. Assume that this statement about a probability is not absolutely certain, but is itself merely probable to the degree q. This weakens the probability of the event, which can now be asserted merely with the probability p-q. What if even the sentence concerning the probability q is not certain, but merely probable to the degree r? This would once more weaken the probability of the event, which is now =p.q.r. And so on: If there is no certainty on any level, the ultimate probability of the event is a product of infinitely many factors each of which is smaller than I and which do not converge to I, and thus the product is equal to o. So, if the empiricist claims that he can maintain empirical knowledge with some probability, he must have certainty on some level of language. This argument was advanced by Bertrand Russell2.
These are the most important arguments in favor of the thesis that there are absolutely certain report sentences. I will now turn to analyzing these arguments and thereby develop the arguments which I raise against the thesis. Let me begin with an analysis of the third argument. I will then proceed to advance arguments against the thesis and shall thereby be led to a criticism of the second and first argument set forth above.
The argument that probability presupposes certainty on some other level appears very persuasive, in particular if it is given the mathematical form in which a product of infinitely many factors smaller than I goes to o. Unfortunately – or should I say, fortunately? – there is a flaw in it. If q is the probability of the sentence s, “The probability of the event is p,” then the probability of the event is not given by the product p.q, but by a more complicated formula, the rule of elimination. Its application requires a knowledge about the probability of the event in case the sentence s is false. Let this probability be =p’; then the probability of the event is equal to q.p + (I-q).p’.3 This expression need not be smaller than p, and can even be larger. So the convergence to o for infinitely many levels of language cannot be derived. In fact, the infinite product p.q.r… does not give the probability of the event, but supplies the probability that all these infinitely many sentences are true simultaneously; and such a probability is, of course, equal to o. Thus the argument is invalid for mathematical reasons.
But why is it so persuasive? It seems so clear that something must be certain in order that some other thing can be probable. I think this is just one of those fallacies in which probability theory is so rich. We are accustomed to thinking in terms of truth and falsehood and wish to extend this habit to probability statements. We argue: if events are merely probable, the statement about their probability must be certain, because… Because of what? I think there is tacitly a conception involved according to which knowledge is to be identified with certainty, and probable knowledge appears tolerable only if it is embedded in a framework of certainty. This is a remnant of rationalism. An empiricist theory of probability can be constructed only if we are willing to regard knowledge as a system of posits. And posits do not require certainty on any level. But I will not go into these problems here, as I have treated them in other publications.
I will rather turn to another argument against the thesis of absolutely certain report sentences, an argument also referring to probabilities, but which uses the theory of probability in the opposite direction, so to speak, namely, for the proof that such sentences cannot be certain. The logic of this argument is very simple. Since report sentences, even when formulated in phenomenal language, are the only basis of knowledge, they must enable us to make predictions of further report sentences. Now it is true that, according to Lewis’ definition of phenomenal language, report sentences cannot strictly imply predictions; i.e., any implications to the future cannot be analytic implications, or, what is perhaps the same thing, strict implications in the sense of Lewis’ famous term, or even synthetic nomological implications, when I use this term to denote the form of implication represented by the laws of nature.4 So the implications from report sentences to predictions of further report sentences must be probability implications. This seems to me to be the only solution for Lewis’ problem of terminating judgments which still allow for predictions: if these predictions are not strictly implied, in any of the above senses, they do not belong to the meaning of the terminating judgment, and thus Lewis’ condition is satisfied; whereas it still makes sense to use terminating judgments for predictions if such predictions are implied with probability.
Now if there exist probability implications between phenomenal sentences, they establish a concatenation between such sentences; and this concatenation works in both directions, from the past to the future as well as from the future to the past. If the phenomenal sentence a, in a certain context, makes the phenomenal sentence b highly probable, whereas non-a would make non-b highly probable, then conversely, the verification of b will make a highly probable, whereas the verification of non-b would make a highly improbable. I refer here to the theorem of inverse probability, known as Bayes’ rule. It is true that the use of this rule requires a knowledge of quite a few probabilities, among which the antecedent probabilities play an important part; but the system of knowledge is elaborate enough to supply all these probabilities, though perhaps sometimes only in the form of rough estimates.5
It follows, then, that any phenomenal sentence is capable of being tested in terms of other such sentences; and though the other sentences cannot completely verify, or falsify, the given sentence, they can verify it, or falsify it, to a high degree.
I will call this the argument from concatenation. It shows that phenomenal sentences are not exempt from probability tests. Let us examine in what sense it can be used for a proof that phenomenal sentences are not absolutely certain.
In defense of Lewis’ thesis, the objection might be raised that, although the probability of the phenomenal sentence a may be very low if it is based on the totality of other phenomenal sentences, we need not abandon sentence a. Once this sentence has been verified by direct observation, we shall prefer this kind of evidence to all the evidence collected in the other sentences, for the reason that direct observation confers the probability I upon the sentence a, whereas the probability of non-a derived from the totality of other sentences is smaller than I. In fact, similar situations do admittedly occur. It is very improbable that when I step on the brake my car starts speeding up. But assume such experience did actually occur; would I be willing to doubt my observation? Would I argue that I was mistaken in believing that my car speeded up, because the probability against the production of such an effect by a stepped-on brake pedal is very high? We would usually refuse to make such an inference; the direct observational evidence for the individual occurrence is stronger than the indirect evidence against it derived from other observations. Would we not be willing to defend in a similar way any individual phenomenal sentence against the totality of other sentences?
The logic of this defense must be carefully analyzed. When we are willing to retain the individual sentence in spite of indirect evidence against it, we do so because we believe that further observational evidence will support it. In the example of the car, we might ask other people whether they, too, saw our car speed up; and we might ask them to check, in a repeated performance of the unusual event, whether our foot is actually on the brake pedal. If both observations are endorsed by other observers, we might proceed to further investigation and may perhaps eventually discover that the sole of our shoe, while pressing down the brake pedal, touched the gas pedal slightly and moved it downward. We then have found an explanation of the unusual occurrence. We have found further observation sentences of such a kind that the totality of observation sentences now confers a high probability on the individual sentence considered. This is, in fact, the maxim of any defense of individual observation sentences: retaining an observation sentence against indirect evidence is equivalent to predicting that the class of observation sentences, on suitable experimentation, will be so extended as to turn the negative indirect evidence into positive evidence.
In the example considered I used not phenomenal sentences but physical sentences, since I referred to such physical occurrences as the motion of a brake pedal and the speeding of a car. But I do not see any reason why the procedure should be different in principle for phenomenal sentences. Retaining a phenomenal sentence against the indirect evidence derived from the total system of phenomenal sentences in which it is embedded is equivalent to assuming that the class of phenomenal sentences, on suitable experimentation, will be so extended as to supply positive evidence for the sentence. I will call this maxim the principle of inductive consistency. It goes beyond deductive consistency. A sentence system relative to which, on further extension, the probability of some individual sentence a goes to zero, is deductively consistent if the sentence a is retained in it; but it is not inductively consistent. It appears obvious that for our system of knowledge we require inductive consistency, because this system cannot be constructed by deductive inferences alone but is based on inductive inferences.
I arrive at the result that phenomenal sentences cannot be absolutely certain, because retaining an individual sentence against any possible indirect evidence may lead to the abandonment of inductive consistency. This argument I regard as a conclusive proof against the thesis.
It may perhaps be objected that we would rather sacrifice inductive consistency than give up a phenomenal sentence. However, it can be shown that such a practice contradicts our actual procedure. This consideration leads back to the second argument advanced above in favor of Lewis’ thesis.
It was said there that the only sources of error for phenomenal sentences are lying and mistakes in the verbal formulation. Now assume that such errors actually have occurred; for instance, that I said “John” when I meant “Peter.” How would I find it out later? Obviously, by inferences from the totality of other report sentences. Now, in these inferences I do not use the fact that the error stemmed from a slip of the tongue; I merely refer to the high probability derived from other phenomenal sentences against the correctness of the word “John,” i.e., I use inductive consistency. When I then proceed to maintain that there must have been a mistake in speaking, I merely advance a psychological explanation of the error, and thus attempt to make my inductive system even more consistent. This is, in fact, the meaning of such qualifications as that the observer must not use the wrong words, or must not lie: they represent possible explanations for false phenomenal sentences.
But it is by no means understood that they are the only possible explanations for such sentences. It is true that there are situations in which we say “John” and mean “Peter,” in which, therefore, at least subjectively the correct meaning can be said to exist. There are other situations, however, in which our attention is not fully focused on our words, and for which we cannot maintain that any distinct meaning was attached to our words. These transitional situations play a greater part than is usually recognized. That is true, likewise, for the case of lying. Although we intend to make a truthful statement, wishful thinking may bias our report; this fact is known to astronomers and physicists, who are afraid that their pointer readings may be slightly falsified in the direction of the result which they wish to find for theoretical reasons, and prefer to have their readings checked by an observer who does not know the theoretical background. The psychological sources of error in phenomenal sentences cannot be simply classified into lying and verbal mistake.
It may perhaps be argued that error presupposes truth, that speaking of possible errors in phenomenal sentences indicates our willingness to assume that there exist true phenomenal sentences. But this objection misses its point. What matters is the sentences we possess, or are willing to accept; it does not help us to know that there are other sentences which we should accept. That there are true sentences describing phenomenal occurrences is no greater miracle than that there is a truly descriptive sentence for any physical occurrence. Our knowledge is built up, not from facts, but from what we know about facts; and knowing refers to language. Phenomenal knowledge must somehow be given in linguistic form, even if it is not the articulate form of conversational language; and for this very reason it can be false. The gap between facts and language cannot be abridged by the belief that the phenomenal sentences known to us cannot be false – this belief is incompatible with the principle of inductive consistency.
Although I think that the argument from concatenation settles the problem of absolutely certain phenomenal sentences in a negative way, I should like to add some remarks concerning the use of a phenomenal language. This consideration will lead us back to the first argument which I advanced in favor of Lewis’ thesis.
It was said there that our actual procedure of checking and correcting physical report sentences exhibits a method of denying physical truth to those sentences and reducing them to mere phenomenal reports. In the example used, the physical object “flight of stairs” was replaced by a phenomenal object carrying the same name. Although at first glance this method seems to support the view that there exists a specific phenomenal language, I wish to show now that the contrary is true, that this method speaks against the psychological priority of a so-called phenomenal language.
I wish to argue that, in a psychological sense, the primary meaning of all terms is given by their reference to physical objects. The term “flight of stairs” means the physical object so called; and when I speak of a tree I saw in a dream, the term “tree” refers to the kind of physical object usually denoted by it. It is very well possible to base all empirical knowledge on a list of observation sentences having physical reference, in this sense. This list would include reports of dreams and hallucinations. Subjecting the list to inductive concatenation, i.e., constructing physical knowledge in the usual way, we then find that, if all report sentences are assumed to be true, some must be false, i.e., we eliminate dream reports, etc., by an inductive form of reductio ad absurdum. We then can ask: How must the eliminated sentences be modified in order to make them compatible with the total system? It is at this point that their phenomenal interpretation is introduced. We say, for instance, that there was no tree, but that we had a dream in which we believed we saw a tree. This sentence can be completely translated into physical language.6 It is merely a convenient abbreviation if we speak of a subjective tree, or the phenomenon of a tree, and thus construct a phenomenal language.
The phenomenal language appears, therefore, as the product of a logical construction starting with terms of the physical language. This is the reason that the phenomenal language cannot be freed from terms of the physical language, that we cannot help but use such terms as “something that looks like a tree,” “something that resembles a flight of stairs.” If we wish to define the meaning of phenomenal terms, we have to refer to physical objects.
The usual objection to this view is that there are direct ostensive definitions of phenomenal terms. But I doubt whether such definitions can be completely separated from physical terms. When we define shades of color by means of a color scale, we refer to a physical object, the scale, and say, for instance, “sand beige is the color of this object.” If we try also to replace the scale by a phenomenal object, we come into difficulties when we wish to reproduce the standards of color, and arrive at sentences like “this here equals that here.” This sentence contains terms like “equals,” “this,” the meaning of which must already be known. If it was introduced in a previous ostensive definition, we must somehow be able to reproduce the defining standards, which procedure involves physical objects and physical language. It seems that phenomenal language is not an actual language, but merely a program which some of us hope to be able to carry through.
This leads me to another interpretation of phenomenal language. When a sentence like “There is a tree” is given, in the ordinary meaning of physical language, and we discover that it is false, we ask: How can we reinterpret the sentence in order to make it true? We then come to such interpretations as, “I dreamed that there was a tree.” This sentence contains the term “I,” which belongs to physical language, since the ego is by no means phenomenally given; and the same holds for the term “dreamed.” The program of a phenomenal language seems therefore to be the directive: replace every observation sentence that is false, physically speaking, by a true sentence which explains why we uttered the original sentence. In order words, we look for a psychological explanation of false observation sentences. When we have found a true explanation, i.e., a true reinterpretation of the sentence, the resulting sentence is true-this is, of course, trivially analytic. If we regard the property of being a true explanation as the criterion of phenomenal language, it thus follows trivially that all sentences of this language are true. But we cannot say beforehand what these sentences will be; whether they refer to a dream, or a slip of the tongue, or a falsification of observation by an unconscious motive. The construction of phenomenal language then merely appears as a procedure supplying psychological knowledge; we can achieve it step by step, but will never be able to say that we have found the truly phenomenal language, because we can never say when we have found the ultimate truth.
These are a few sketchy ideas to show how we may incorporate a phenomenal language in an empiricist philosophy. My main argument is that such a philosophy does not need a basis of absolutely certain sentences. The system of knowledge can be based on observation sentences of the physical language of everyday life. Those among these sentences which are tenable are found by inductive concatenation; and those which are not tenable are given a reinterpretation, by the same method, which makes them tenable as psychological reports. Empiricism does not need absolute certainty, either for so-called principles of knowledge, or for observation sentences. There is no synthetic a priori; and there is no synthetic a posteriori that is absolutely certain.
Philosophical Review, vol61, no.2, pp.147-59, 1952
1 As to this term and the use of co-ordinative definitions for the comparison of perceptions, see my book. Experience and Prediction (Chicago, 1938), pp. 183,
2 Bertrand Russell, Human Knowledge (New York, 1948), p. 416.
3 For an exact treatment of this probability see my Theory of Probability (Berkeley, 1949), p. 321 (German edition [Leiden, 1935], p.315).
4 See my Elements of Symbolic Logic (New York, 1947), ch. viii.
5 From the existence of a probability P(a,b) alone we cannot infer that a probability P(b,a) exists; see my Theory of Probability (Berkeley, 1949), pp. 109-no. But this inference can be made if, in addition, the probabilities P(a) and P(a,b) exist; this follows from formula (9) (ibid. , p. 92), when the general reference class A is omitted. I may add the remark that all these considerations, of course, belong in what I have called advanced knowledge.
6 See my Elements of Symbolic Logic (New York, 1947), p.275, and a note in Philosophical Studies, II (1951), 92.