THE “LIMITS OF EXPERIENCE” OBJECTION

The general idea here is simple and powerful. By observing nature, we can only make inferences about what is observable in nature, not what is unobservable. Here are claims of two prominent antirealists.

Now these two questions—Does there exist a material reality distinct from sensible appearances? and What is the nature of this reality?—do not have their source in experimental method, which is acquainted only with sen­sible appearances and can discover nothing beyond them.^{[122] [123]}

Second, van Fraassen:

I explicate the general limits [of experience] as follows: experience can give us information only about what is both observable and actual.¹¹⁸

The claim, then, is that any argument from what is observed to the truth of claims about what is unobservable is unjustified. Since Perrin claimed truth for his theory about unobservable molecules, and since his argument is based on observed results of experiments that he and others conducted, his argument is unjustified.

What is the basis for the claim that by observing nature one can make inferences only about what is observable? A simple defense, one indeed suggested by Duhem, is that this is the essence of empiricism; inferring beyond the observable is metaphysics, which is not empirical. Indeed, immediately following the passage quoted here, Duhem claims that the resolution of questions as to what lies beyond the “sensible appearances” “transcends the methods used by physics; it is the object of metaphysics” (p. 10). But this begs the issue, which in this case is that of characterizing empiricism. Perrin, no less than Duhem, regarded himself as engaged in empirical science, not metaphysics. He believed that empirical arguments could be given for the truth of a theory about unobservable molecules. It will not do simply to claim that his argument is not (entirely) empirical, or that he has gone beyond what empiricism allows.

Van Fraassen regards as extreme a policy that permits inferences beyond the “range of possible additional evidence,” that is, inferences to the truth of claims about unobservables.^{[124] [125] If only evidence can justify a belief he goes on to say, then belief in the truth of a theory is “supererogatory,” since we can have evidence for truth only via evidence for empirical adequacy (p. 255). Van Fraassen's first claim presupposes, or at least strongly suggests, that inferences to the truth of claims about unobservables are inferences beyond what the evidence allows. His second claim is that one can have evidence only for the empirical adequacy, not the truth, of a theory postu­lating unobservables, or that evidence for truth only amounts to evidence for empirical adequacy. Both claims, I suggest, require justification. Perrin, for example, clearly believed that he had evidence for the truth of propo­sitions about unobservable molecules, not simply for their empirical ade­quacy; he had evidence that molecules exist, not simply evidence that the molecular theory saves the phenomena. Why was he wrong in this belief?}

Let me suggest an argument an antirealist might offer in the spirit of the two quoted passages. Suppose that having observed a great many As and found them all to be B, one infers that all As are B. For example, from the fact that all observed bodies have mass one infers that all bodies, including any unobservable ones, do too.²⁰ (In Perrin's case, we might consider an inference from “All observed accelerating bodies in contact with other bodies exert forces on them” to “All accelerating bodies, in­cluding molecules (if any exist), in contact with other bodies exert forces on them.”) Is an inference of this sort legitimate? Not necessarily. The sample observed may be unrepresentative. Even if we do not know that it is unrepresentative, we may have no positive empirical reason to think it is not biased. Suppose that all the As chosen for observation satisfy a condition C.

Now, says the antirealist, let C be the condition of being observable. All observed As satisfy this condition. Indeed, one can never observe an A that fails to satisfy it. So, unless one can demonstrate that this does not bias the sample of As observed with respect to B, from the fact that all As observed have been B, all we can legitimately infer is that all observable As are B. The only way of demonstrating that “observability” does not bias the sample of As observed with respect to B is by collecting a suitable sample of unobservable As and showing that in this sample the unob­servable As are all B (or perhaps by showing that in other cases, samples of unobservables match observables with respect to some property). But, of course, one cannot do that! One cannot observe unobservables. So in order to exclude potential bias, we are restricted to making (inductive and causal) inferences to what is observable. This argument can be extended to cover all types of nondemonstrative inferences. That is why Duhem is right in claiming that we can discover nothing beyond appearances, and why van Fraassen is justified in claiming that experience can give us infor­mation only about what is observable.

Reply

An antirealist who argues in the previous manner looks at the situation as one involving so-called stratified sampling. The population of As is divided into two classes or strata: the observables and the unobservables (if any). To make inferences about the entire class of As with respect to a property B, the antirealist is claiming, one needs to select randomly members from both strata for observation. Since one cannot select unobservable mem­bers for observation, one cannot legitimately, without potential bias, make inferences about this stratum, but only about the observable stratum.

If this argument is derived from the general principle that to make an inductive inference about a class, samples must be taken from all strata of that class, then no inference from any observed sample will be possible. Let the stratifying condition C be ‘has been observed' (or ‘has been ob­served prior to 2500).

Assuming that the antirealist does not accept this general principle with respect to all stratifying conditions C, what reason can he offer for supposing that the specific stratifying condition “being observable” is a biasing condition? Does he have any empirical reason for supposing that, in general, when considering a class of As, the unobservable stratum of As, if any exist, is different with respect to B from the observable stratum? No, he does not, since, by his own admission, unobservables cannot be sampled. Nor does he have some a priori argument showing that the two strata are different with respect to B. All he has is the weak a priori claim that the strata may be different—a claim about a mere logical possibility. But this is not sufficient to justify the methodological injunction that to make a legitimate inductive inference about the entire class consisting of observables and unobservables one must sample both strata. Indeed, if it were sufficient, the antirealist would be prevented from making any gen­eralizations from what has been observed to what is observable but never observed. If one stratifies the class of observable As into those which are or will be observed and those which never will be observed, then it is log­ically possible that these strata are different with respect to B.

The realist, who wants to make inferences about unobservables as well as observables, must also reject the general principle of stratification. Can he do so without rejecting the idea that sampling requires variation? Can the realist offer an argument that provides support for his claim that “observability” is not, in general, a biasing condition?

I suggest that here the realist has an advantage over the antirealist. The realist can offer an empirical argument that provides support for his claim.

Suppose we do the latter and find that bodies have mass even when we observe bodies that have different sizes, different distances from us, different durations, and different numbers and kinds of interactions with other bodies. In the absence of any contrary empirical information, we can then infer that size, distance, duration, and numbers and kinds of interactions do not alter the situation: bodies observed with different sizes, distances, du­rations, and interactions all have mass. So we infer that differences in these properties—differences that make some bodies observable and others not— make no difference as far as having mass is concerned.^[126] If we vary the con­ditions in virtue of which bodies are observable and find no differences in whether bodies have mass, and if we have no contrary empirical information, then we have offered an empirical argument to support the claim that the fact that all observed bodies are observable does not bias the observed sample with respect to the property of having mass. In so doing, we have provided empirical grounds for inferring that all bodies have mass, whatever their size, distance, duration, and so on, and hence, whether or not they are observable.^[127]

An antirealist may vehemently repudiate this argument by insisting that the only way to vary a condition C that is satisfied by all observed As is to observe As that satisfy C and As that do not.

For example, suppose an antirealist makes an inference he considers legitimate from the fact that all observed bodies have mass to all observ­able bodies have mass. In doing so, in order to preclude bias, one might vary the size of the bodies examined, their distance from other bodies, and a host of other properties. More generally, some variations involve changing “qualitative” properties by examining As that have such properties and As that do not. But there are also variations that involve changing “quantitative” properties by examining As with varying amounts of such properties. If the antirealist rejects this second type of variation for eliminating bias, he needs an argument for doing so. If he accepts it, but only for infer­ences to observables, again he needs an argument for this restriction. The claim that unobservable bodies may be different from the observed ones with respect to having mass—in the sense of logical possibility—cuts no ice, since exactly the same could be said for the observable (but not yet observed) bodies. Nor, again, does the antirealist have either an empirical or an a priori argument to show that the observed bodies are unrepresen­tative of the unobservable ones with respect to having mass.^[128]

5.