1. The Humean Challenge
David Hume articulated a powerful skeptical argument which casts doubt on the justification of one our basic forms of reasoning. In general, the process which underlies such reasoning goes something like this: all other things being equal, given that m/n observed instances or events of type A are or have been B, we can correctly infer that m/n As are (or have) B. Let’s call this the Inductive Principle (IP). Very roughly put, unobserved cases or instances of a certain type will resemble observed cases or instances of the same or similar type. The basic inferential structure of inductive reasoning can be captured by IP or a principle akin to it. Systems of inductive logic formulate rules which assign inductive probabilities to arguments based on the strength of the evidence or degree of support that the premise(s) provide for the conclusion. Arguments which are assigned a high inductive probability will yield true conclusions from true premises most of the time.
Hume claimed that the inductive principle cannot be supported by any rational argument. Recently, however, Laurence BonJour has proposed a novel justification and defense of IP in an attempt to circumvent the Humean argument. In this paper, I elaborate and criticize BonJour’s strategy. After briefly reviewing Hume’s skeptical argument against induction in this section, I proceed to outline BonJour’s apriorist response to the argument, which defends inductive reasoning by means of a counter-argument consisting of a priori premises. Along the way, I’ll draw attention to Anthony Brueckner’s main criticisms of BonJour’s a priori approach, detailing why they fall short of providing an effective rebuttal to BonJour’s argument. I then go on to show why BonJour’s strategy fails to provide a cogent defense of IP, drawing out the implications for a priori attempts to argue for the principle.
Whether self-consciously or unreflectively, rational beings engage in the process of forming inductive inferences in acquiring, maintaining, revising, and discarding beliefs about what is either unobserved or as yet in the future. But what justifies us in reasoning this way? Is there any rationale for thinking that drawing inferences on the basis of IP, or assigning high inductive probabilities to arguments from some system of inductive logic S having rules R, will give us conclusions that are likely to be true given that the premise(s) are true (or that the probability of a certain conclusion, given that the premise(s) are true, is at least greater than it would be otherwise)? This, in a nutshell, is the traditional problem of induction. Hume (Enquiry, section IV) lays out an argument which purports to show that there is no rational justification for inductive reasoning. On the one hand, no piece of demonstrative reasoning can establish either IP or any inductive conclusion generated by it, because relations among matters of fact are not necessary. It is always possible that the future will be unlike the past, or that unobserved cases of a certain sort do not resemble observed cases of a similar sort. In a valid deductive argument, the conclusion can make no factual claim that is not already implicitly contained in the premises. But the premises of an inductive argument contain only information about past and present events or states of affairs which have been observed. Thus a deductively derived conclusion cannot go beyond this to make claims about future or unobserved happenings. But it is precisely these sorts of claims that characterize IP and inductive reasoning; so it looks like there is no way IP or any inductive system S can be justified deductively.
On the other hand, Hume argues that induction cannot be justified by means of an inductive argument either. Any inductive argument will have to assume in advance that IP or S is reliable in order to prove that IP or induction is justified, and this amounts to circular reasoning. One might argue that IP or S is justified because it has a good track record, and since it has worked in the past, it will continue to work in the future. But this argument is an inductive argument which itself either incorporates some form of IP, or has been assigned a high inductive probability by certain rules of S, the system whose justification is the very point at issue. It might be suggested that this type of argument could appeal instead to some version of the principle of the uniformity of nature. However, even assuming that an adequate version of this principle could be formulated, it is difficult to see how it could be justified other than by implicit appeal to IP or S. We’ll revisit the issue of the uniformity of nature when we examine BonJour’s argument.
The Humean challenge can be viewed as a kind of dilemma which runs roughly as follows:
P1: Any rational justification of inductive reasoning must take the form of either an a priori deductive demonstration or else a strong inductive argument.
P2: Inductive reasoning cannot be justified by means of an a priori deductive argument.
P3: Inductive reasoning cannot be justified by means of a strong inductive argument.
C: Therefore, inductive reasoning cannot be not rationally justified.Footnote 1
What should we make of the skeptical argument? Attempts to block Hume’s argument have given rise to several different strategies for defending induction. I have discussed some of these strategies elsewhere.Footnote 2 In what follows, I focus on BonJour’s recent a priori defense of the inductive principle.
2. BonJour’s A Priori Defense of Induction
In a recent book which makes a fresh assessment of the status of a priori reasoning, BonJour sets out to build a case for an a priori solution to the problem of induction.Footnote 3 BonJour’s commitment to an a priori defense derives from his acceptance of an assumption about how justification is transferred from epistemically foundational beliefs to non-basic beliefs, especially beliefs whose content goes beyond direct empirical experience and observation. He holds that warrant for such beliefs requires some form of inference from the underlying foundational beliefs. Observation-transcendent inference is typically carried out by forms of inductive reasoning. Since inferred beliefs about the external world go beyond the content of immediate experience, it is impossible to directly justify those inferences solely on the basis of experience. Where we extend our conclusions beyond observed cases by generalizing from experience, we tacitly employ basic inductive principles, and hence we need to justify those principles. If our attempted justification involves reasoning from experience, then we beg the question by assuming the truth of those very principles, as Hume pointed out. Thus, if extreme forms of skepticism are to be avoided, justification of observation-transcendent beliefs requires an a priori justification of the underlying inductive principles.Footnote 4
Given BonJour’s overall project, it is natural for him to attempt to answer the Humean skeptical argument outlined above by countering the second premise. With this end in mind, BonJour constructs and defends an a priori argument for the truth of IP. In order to dispel certain misconceptions about what an a priori justification of induction would require, he offers some preliminary remarks aimed at clarifying what is meant by an a priori argument.Footnote 5 An a priori approach is not a subtle attempt to turn inductive reasoning into deductive reasoning. No claim is made to the effect that conclusions of particular inductive arguments follow from their premises with deductive certainty. Nor should an a priori solution involve the implausible claim that a principle such as IP is itself a self-evident a priori truth. BonJour recognizes that we cannot rule out on a priori grounds the possibility that ours is a chaotic universe which lacks the order and regularity required for the truth-preserving character of inductive inference. Rather, an a priori justification of induction or IP is a justificatory argument whose reasoning and key premises are known or justified strictly a priori. Roughly, a basic non-empirical proposition (basic in the sense that it is not inferred from other propositions) is known or justified a priori if, and only if, its truth and necessity is grasped or apprehended in a direct and immediate way through an act of rational insight.Footnote 6 While the comprehension of certain concepts may depend on the subject’s having had sensory experiences of a certain sort, the act of rational insight itself does not depend on experience in any further way. BonJour allows that an a priori argument may contain one or more empirical premises, but stipulates that the core premises and reasoning linking the empirical propositions to the conclusion must be known or strongly justified a priori.
BonJour goes on to highlight the basic ingredients that are required for an a priori solution to the problem of induction: an a priori reason for thinking that the conclusion of a standard inductive argument is likely to be true if the premises are true, which consists of two a priori claims, a) there is some explanation for why the proportion of observed As that are Bs converges on some relatively constant value m/n, and b) there is some sort of objective regularity which best accounts for the phenomenon described in (a). This forms the basis of a three-step a priori argument for induction, which BonJour proceeds to lay out and defend in some detail. His first premise is:
(I-1) In a situation in which a standard inductive premise obtains, it is highly likely that there is some explanation (other than mere coincidence or chance) for the convergence and constancy of the observed proportion.
(I-1) asserts that it is highly probable a priori that there is a non-chance explanation for the truth of a standard inductive premise, an explanation for why m/n of the observed As are Bs. The empirical evidence that informs the inductive premise consists of a set of observations taken over time which have the cumulative effect of showing a general constancy in the observed proportion or a gradual convergence on m/n. Chance occurrences are event tokens that are uncaused, or do not conform to appropriate law-like regularities, or fail to be the result of some causal propensity. By a “non-chance” explanation, then, what BonJour presumably has in mind is an explanation of the observed proportion of evidence in terms of antecedent factors in the world which are instances of appropriate causal regularities or result from certain causal propensities. Contrary to what many philosophers have assumed, BonJour finds no compelling reason why such a meta-thesis, about the likelihood of a certain other thesis, cannot be an a priori truth. Indeed, there might be possible worlds (including the actual world itself) in which a chance explanation could regularly be found for the truth of standard inductive premises. Yet as long as this situation is infrequent within the total class of possible worlds, it would still remain true in every world that it is likely that there is a non-chance explanation for the truth of a standard inductive premise. Hence (I-1) would hold in every possible world and would thereby be true necessarily.Footnote 7
BonJour’s second step in the argument involves articulating the sort of non-chance explanation for the observed proportion which he deems most plausible a priori:
(I-2) [Excluding the possible influence of observation] the most likely explanation for the truth of a standard inductive premise is the straight inductive explanation, namely that the observed proportion m/n accurately reflects a corresponding objective regularity in the world.
The objective regularities in nature, which are reflected in inductive evidence, refer to uniform patterns of events and states of affairs which hold across relevant spatio-temporal boundaries. They involve objective, lawful correlations in the world, deriving from relevant underlying causal processes and mechanisms, and which, in turn, are responsible for the fact that approximately m/n of all As are Bs. BonJour considers the possibility that certain factors involving observation itself might affect the proportion that is actually observed, making it turn out that the observed sample does not accurately reflect the overall proportion of As that are Bs in the world.Footnote 8 However, as he sees it, that is a different question, and the problem of induction simply does not address the issue. The classical problem of induction is about whether generalization from observed to unobserved cases is justified when such observational influences are absent; and to this problem BonJour thinks he has a solution.
In defense of his second premise, BonJour considers what other possible explanations, besides the straight inductive explanation, could account for the inductive evidence in question. He calls such possible explanations normal non-inductive explanations. In the simplest case, the relation between the presence of two objects or properties A and B holds as a lawful regularity, but there is some further characteristic or factor C that combines with the As and Bs to produce a situation in which i) m/n of observed As are Bs, but ii) the presence or absence of C affects the proportion of As that are Bs, so iii) it is false that even approximately m/n of all As are Bs.Footnote 9 For instance, it might be the case that there is a certain overall proportion of As that are Cs, which leads to a certain overall proportion of As that are Bs; but that the actual observations of A involve a higher (or lower) proportion of C cases as compared to non-C cases, thus resulting in an observed proportion of As that are Bs which is significantly different from the overall true proportion. Or perhaps the occurrence of C in relation to A does not, on the whole, fall into a regular pattern, with the result that there is no objectively correct proportion of As that are Bs. Nonetheless, observations of A might include a relatively uniform proportion of Cs, resulting in a certain observed proportion of As that are Bs. In either case, the observed proportion will fail to reflect the actual overall proportion in such a way as to falsify the standard inductive conclusion. BonJour contends that it is a priori highly unlikely that either of these two situations be realized through sheer coincidence or chance. So a normal non-inductive explanation is extremely unlikely to be true. It follows, then, that the best explanation for the observed constant proportion of As that are Bs, the explanation most likely to be true, is the straight inductive explanation.
Once the likelihood of the straight inductive explanation as claimed in (I-2) is established, BonJour’s a priori justification of induction is complete. Theses (I-1) and (I-2) conjoined together constitute an a priori justification of induction. If it is highly likely a priori that there is a non-chance explanation for the obtaining of standard inductive evidence, and if the explanation that is a priori most likely to be true is the straight inductive explanation, then it is probable a priori that given the truth of a standard inductive premise, the corresponding inductive conclusion is true. So from the above two theses, BonJour concludes:
(I-C) Therefore, it is a priori likely that if a standard inductive premise is true, then the corresponding standard inductive conclusion is true also.
(I-C) is equivalent to the inductive principle (IP) mentioned in the previous section: given that m/n observed instances or events of type A are or have been B, we can correctly infer that m/n As are (or have) B. BonJour’s argument constitutes a defense of this principle and thus purports to establish the cogency of basic inductive inference. It is the fact that the objective regularities and patterns referred to in (I-2) are observation-transcendent—they hold across unobserved spatio-temporal boundaries—that licenses a given standard inductive conclusion. If the above argument is successful, then BonJour has provided a way of answering the Humean challenge.
3. Clarifying the Argument
Anthony Brueckner develops a two-pronged approach in responding to BonJour’s argument.Footnote 10 On the one hand, he argues that an a priori approach is not the only way to justify induction, for induction can have epistemic warrant in the context of certain theories of justification without appeal to any formal argument at all. He develops this point in some detail by outlining how a pure coherentist approach would deem inductive inferences justified and undermine P1 of the skeptical argument.Footnote 11 I will set aside this aspect of Brueckner’s reply in what follows, for my aim here is to focus on the cogency of BonJour’s strategy and to assess whether an a priori justificatory argument for induction might be successful. On the other hand, Brueckner raises several doubts about the plausibility of (I-1) and (I-2). These mainly take the form of direct challenges to the effect that BonJour has not shown how either (I-1) or (I-2) is a priori justified.Footnote 12 In discussing these criticisms in the next two sections, I will suggest how BonJour might counter them based on claims implicit in his explication of the a priori argument. This will pave the way for my main line of criticism of BonJour, which contrasts with and is much more thoroughgoing than the individual challenges and worries pressed by Brueckner. For I will show why the peculiar nature of probability claims like (I-1) and (I-2) are such that they cannot be given a suitable a priori justification, something Brueckner has not attempted to do. BonJour’s argument for induction and IP fails, not because the claims expressed by the premises are themselves implausible, but because if they are justified at all, they are by nature such that they cannot be not known or justified a priori. Hence (I-C) lacks an adequate a priori defense.
In utilizing the notions of ‘likelihood’ and ‘best explanation’ throughout the argument, BonJour opts not to give a formal characterization of probability, Bayesian or otherwise. In the discussion that follows, while I will generally stick with BonJour’s rather loose and informal notion of probability, at times it will be helpful to clarify certain matters by appeal to a few simple Bayesian formulas.Footnote 13 Both (I-1) and (I-2) can be recast as conditional probability statements: (I-1) says that given the occurrence of a set of inductive evidence e, the probability that there is some non-chance cause or explanation of the e-type pattern or distribution in the evidence (or the convergence on the pattern) is high. Given the evidence e, there is a high probability that some non-chance hypothesis h n with respect to the cause of e is true: P(h n/e) is high.Footnote 14 Likewise, (I-2) claims that the standard inductive premise that the observed proportion m/n of As are Bs renders it highly probable that there is a corresponding objective regularity in the world, reflected in the evidence. Where h u is the hypothesis that there is an objective, uniform regularity among events and phenomena in the world which is reflected in an e-type pattern or distribution of a range of inductive evidence e, then given the occurrence of such evidence, h u is extremely likely to be true. In other words, P(h u/e) is high.Footnote 15
Notice how (I-1) and (I-2) are related. On one possible reading, (I-2) entails (I-1). For if a standard inductive premise renders likely or probable the straight inductive explanation—that is, renders probable the existence of an objective regularity as its explanans—then the standard inductive premise ipso facto renders probable the existence of some non-chance phenomenon or occurrence as its explanans, since an objective regularity constitutes a type of non-chance phenomenon. On this reading, (I-2) alone would entail IP, and there would be no need to separately specify (I-1) in the argument; it would play no genuine role in the reasoning. This is not the way BonJour intends the relationship between the premises to be understood. BonJour construes (I-2) in such a way that he takes the probability of h u to be conditional on the inductive evidence together with the assumption of a non-chance explanation of the evidence (h n). Using our notation, P(h u/h n&e) is high. On this construal, (I-2) does not entail (I-1), and the corresponding probabilities are to be assessed independently. The probability of the straight inductive explanation on the evidence (P(h u/e)) will be the product of the probabilities expressed by (I-1) and (I-2): P(h u/e) = P(h n/e) x P(h u/h n&e). P(h u/e) expresses the likelihood that there is an objective regularity of the sort which licenses IP and (I-C), and so we can assume that this equals the probability that (I-C) is true.Footnote 16
According to his argument, BonJour thinks that it is an a priori truth that the occurrence of a range of inductive evidence renders it highly likely that the world manifests a corresponding systematic and uniform regularity (h u). Let us call this principle of the objective uniformity in nature which underwrites the inference to (I-C) the uniform regularity hypothesis. What degree of probability would be sufficient for the uniform regularity hypothesis to have, given the inductive evidence, so that inductive inference would be rendered trustworthy? This will be determined by the degree of likelihood we think (I-C) should have, in order for standard cases of inductive reasoning to be considered cogent. A probability of two-thirds (.67) would arguably be too low for many cases of inductive reasoning, but let’s assume this as our target. A probability of .67 assigned to (I-C) or IP would require that the probability of each of the premises of BonJour’s argument average slightly greater than .81. Of course, if the likelihood of one of them could be shown to be higher by a certain amount, then the likelihood of the other could be correspondingly lower. What good a priori reason do we have for thinking that the probability of either (I-1) or (I-2) approaches .81? I submit that we cannot know a priori whether either of (I-1) or (I-2) has a probability that approaches .81. The likelihood of claims such as (I-1) and (I-2) is, by the very nature of the claims, a priori undecidable. If so, then BonJour hasn’t shown that the standard inductive evidence renders IP or IC likely enough to warrant standard inductive inferences. Or so I will argue.
4. The A Priori Status of (I-1)
Why does BonJour think that the likelihood of a chance explanation of the standard inductive evidence is ruled out a priori? He concedes that considered in the abstract, a chaotic world in which everything happens by chance is no less likely a priori than a world exhibiting an objective and systematic order. In some of those chance worlds, the inductive evidence obtains, and it is even possible that ours is such a world. But although possible, BonJour thinks that such a world is a priori unlikely. Once inductive evidence is introduced, it becomes highly probable that there is a non-chance explanation of the evidence, and that the world does manifest an objective and uniform order. The crucial question is: why should the assumption of the truth of a given range of inductive evidence be a game-changer and render it likely that the world does exhibit such an order? BonJour’s answer is that it is simply evident upon reflection that it is unlikely that mere chance or coincidence is at work; possible worlds in which inductive evidence obtains but induction generally fails “involve the repeated occurrence of an unlikely situation—and hence they are quite rare and unlikely within the total class of possible worlds.”Footnote 17
In his chief objection to (I-1), Brueckner charges that BonJour has offered no serious rationale for holding that the truth of a standard inductive premise has a non-chance explanation. He briefly considers two ways in which such alleged truths might be grounded—by appealing to knowledge of frequencies, or by appealing to some “probabilistically qualified principle of sufficient reason.”Footnote 18 Such a principle would claim that for every contingent truth t, it is likely that there is a correct explanation for t. Brueckner rightly rejects both possibilities. He dismisses the former because empirical knowledge of frequencies would be a posteriori rather than a priori. He further points out that BonJour’s own claims and defense of his core premises implies that one can have a priori knowledge of probabilistic statements that are in no way based on frequencies. Regarding the second possibility, Brueckner claims that there is nothing in BonJour’s defense that would provide backing for a probabilistic principle of sufficient reason; hence relying on such a principle would be baseless.
BonJour neither does nor should rely on such a principle. To claim that it is probable that every contingent truth has a genuine explanation is not the same thing as saying that it is probable that recurring patterns of inductive evidence have a genuine (non-chance) explanation. It is this latter claim which BonJour thinks we can know a priori. In attempting to provide a defense of (I-1), BonJour seems to allude to quite a different rationale from the options Brueckner considers. In effect, BonJour argues that the epistemic probability of a chance explanation of inductive evidence is low when figured from the total space of logical possibilities available.Footnote 19 Although BonJour’s reasoning on this point is not as explicit as one would hope, we can elaborate on his idea somewhat and fill in the details where needed. BonJour thinks that among all the possible worlds in which inductive evidence patterns obtain, the majority of those worlds contain non-chance causes of the evidence patterns, that is, they are worlds in which the observed proportion of evidence is explained by antecedent factors arising from the causal dispositions of objects. Since such worlds tend to manifest regular patterns of interaction among their constituents which reflect the nature of the causal dispositions involved, the vast majority of these worlds will contain inductive evidence. So suppose 80% of these non-chance worlds exemplify patterns of inductive evidence. On the assumption that half of the total number of possible worlds are non-chance worlds and the other half are chance worlds, and even supposing that inductive evidence obtains in as many as half of the chance worlds, it turns out that the probability of h n with respect to the evidence (that is, P(h n/e)) will be well over .5.Footnote 20 So contrary to Brueckner’s contention, some simple reasoning about the space of possible worlds and the character of these worlds can lead us to estimate the epistemic probability of the truth of a non-chance hypothesis given a standard inductive premise, and this, in turn, gives us a priori grounds for thinking that (I-1) is likely to be true. If so, then perhaps BonJour’s rationale for (I-1) can be upheld in a manner that Brueckner fails to consider.
What are we to make of this sort of response? On closer inspection, it is doubtful that the epistemic probability of h n with respect to e is assessable in the way BonJour’s rationale requires. Think about what would be needed in order to determine whether a given pattern of evidence is more likely to be the result of chance, or whether it is more likely to be the result of some non-chance phenomenon. According to Bayes’ rule, the probability of a hypothesis on a given set of evidence is proportional to the prior probability of the hypothesis itself and inversely proportional to the prior probability of the occurrence of the evidence. More formally: P(h/e) = P(e/h) x P(h) / P(e). So in order to assess the likelihood of P(h n/e), one would first have to know two things: the prior likelihood of the truth of a non-chance hypothesis (which explains the evidence) and the prior probability of the occurrence of the given type of evidence pattern. For standard cases in which we already know or can estimate to a rough approximation the values of P(h) and P(e), an assessment of the probability of the hypothesis given the evidence can be readily carried out. If I roll a die one hundred consecutive times, and each time it comes up a six, naturally I’ll think there is a high likelihood that the die has been somehow rigged. On the one hand, prior to the occurrence of the evidence, the probability of this hypothesis being true is extremely low, as is the probability of the evidence pattern itself. I reason that if it were a fair die, on average I should roll a six (or any other selected number) on one out of every six rolls. On the other hand, the probability that the evidence pattern occurs given the truth of the hypothesis is high. I can give a rough estimate of the likelihood that the die is rigged in favour of sixes because I have prior knowledge of P(h) and P(e).
Notice also that my belief that both P(h) and P(e) are low tacitly presupposes that a system of underlying causal laws and objective regularities already holds in the world; these laws, in turn, determine the types of behaviour and activity that macro-objects are able to undergo. If a solid object with such-and-such a shape and uniform weight distribution is propelled with a certain magnitude of force onto a flat level surface, it will land a certain way and eventually come to rest on one facing surface (and won’t come to rest on one of its edges). If I didn’t assume that a system of uniform physical laws holds, then I would lack the necessary backdrop for predicting that a die has a one in six chance of turning up a six, or for forming a correct judgment about the likelihood of a fair or unfair die turning up a six on one hundred consecutive rolls. And lacking that information, I would be unable to determine, even to an approximation, the degree to which the hypothesis of an unfair die is probable on the given set of evidence.
In the same way, all real-life probability calculations make use of background knowledge of prior probabilities as well as assumptions about the kinds of systematic regularities that obtain. In estimating the likelihood that a smoker will develop cancer, for example, we need to take into account the antecedent probability that any person develops cancer, along with the prior probability of being a smoker. We further factor in the likelihood that someone is a smoker given that he or she has cancer. Using this information in Bayes’ rule, we can estimate the probability that someone will contract cancer given that he or she is a smoker. Moreover, the antecedent probability that a person will, at some point in the future, develop cancer presupposes that certain causal laws and uniform, systematic regularities hold in relevant areas—in relation to human beings and their environment, as well as in relation to their physiological and genetic makeup. In the absence of such an assumption, we would have no good reason to think that sample data from the past can be extrapolated to cover future or unknown cases.
As these examples illustrate, typically when we ascertain the likelihood of a hypothesis or the likelihood of certain evidence for a hypothesis, we do so with respect to an already assumed backdrop of systematic regularity and order. Perhaps we can say that standard cases of probability assessment, as illustrated in the examples above, involve a certain kind of framework from which probability is viewed and assessed from the inside, as it were; let’s call epistemic probabilities figured in this way I-assessments. I-assessments presuppose a system of uniform, objective regularities in the world and make use of those regularities in figuring the antecedent probabilities of events (or hypotheses about events). In assessing probability from the inside, we assign antecedent probabilities to events in two ways: by calculating frequencies based on information gathered by direct observation of occurrences of the events, or by estimating their probabilities based on other available background information. In either case, an essential factor in computing these prior probabilities is background information about the kinds of underlying laws or regularities that we assume to hold in the world. The central problem with BonJour’s (I-1) and (I-2) is two-fold: first, we lack information about the relevant prior probabilities; and second, we cannot presuppose any framework of objective laws and regularities, because it is the existence of such regularities that is at issue. If assessing (I-1) and (I-2) as highly probable were to implicitly appeal to the existence of an overarching and uniform order in nature, that would amount to assuming a key part of what is being argued for; to assume the existence of such an order would beg the question in favour of something like the uniform regularity hypothesis.
Since BonJour’s argument is an attempt to establish the very framework of objective and systematic order which licenses inductive inference in particular cases, his judgment of the likelihood of the premises cannot rely on that framework. His only option is to set out to determine the probability of the uniform regularity hypothesis from the outside. The probability of a hypothesis or phenomenon is assessed from the outside—or given an O-assessment—when there is, in principle, no suitable empirical background information that can be appealed to in order to determine relevant antecedent probabilities, and there is no overarching framework of laws or regularities relative to which probabilities can legitimately be measured. That is, in offering an O-assessment of a probability, we don’t have the necessary background information for figuring prior probabilities, and there is no assumed underlying backdrop of more fundamental regularities on the basis of which we can compute or estimate those probabilities. Without such a framework, how is BonJour to assess the relevant prior probabilities of the standard inductive evidence (e), the non-chance hypothesis (h n), and the uniform regularity hypothesis (h u)? I won’t take a stand here on the more general question of whether it is ever possible to give an accurate O-assessment of a hypothesis, but rather I will confine my present arguments to the specific claims BonJour makes about induction.Footnote 21 Let’s examine (I-1) first. Bayes’ rule tells us that in order to know P(h n/e), we need to know P(h n) and P(e), as well as P(e/h n). Since h n refers to a non-chance hypothesis which specifically explains the occurrence of e, and not just any non-chance hypothesis about the world, every world in which h n holds will be a world in which e occurs. This allows us to set the probability of e/h n to 1, thus simplifying our calculations somewhat. Our remaining task is to estimate P(h n) and P(e). Articulating the possibilities in terms of possible worlds vocabulary, we can assume that there are many non-chance worlds—worlds in which phenomena and events for the most part have non-chance causes and explanations—as well as many chance worlds. We can locate several points at which an attempt to acquire a priori justification for P(h n/e) ≥ .81 breaks down, because the required information is simply not available. For example, should we assume that the number of chance worlds is approximately equal to the number of non-chance worlds? If the former are significantly more numerous than the latter, that could result in there being more e worlds that are not h n worlds, lowering the ratio of h n to e worlds, and thus lowering the probability of h n/e. There doesn’t seem to be any strong a priori reason for positing that the number of non-chance worlds is roughly equal to or significantly greater than the number of chance worlds. But let’s assume for the moment that we have reason to believe that the number of possible chance worlds and non-chance worlds is roughly equal. That still would not give us a good reason for thinking that P(h n/e) ≥ .81. For we don’t know the ratio of non-chance worlds in which e occurs to chance worlds in which e occurs. The lower the ratio of the former to the latter, the lower P(h n/e) will be. A priori reflection on the matter simply does not impart any compelling rational conviction regarding how many metaphysically possible chance hypotheses there are or what proportion of them include the inductive evidence. With regard to these questions, the Bayesian method for estimating probabilities appears to break down, leaving us with no objective procedure for assigning prior probabilities to e or h n, and thus rendering these probabilities a priori inaccessible.
But, for the sake of argument, let’s assume that 90% of all e worlds are non-chance worlds. Could we then infer that P(h n/e) is .9? Such an inference assumes that each logically possible world has an equal ‘chance’ of being actualized. Suppose that there are 100 metaphysically possible worlds, 100 combinations of maximally possible states of affairs. Call one of these maximally possible states of affairs Wn. It doesn’t follow from the fact that there are 100 possible worlds that Wn has an antecedent chance of .01 being actual. Even if we could accurately estimate the number of metaphysical possibilities or scenarios, and estimate the proportion of those scenarios in which the evidence in question obtains, that wouldn’t tell us the actual chance or probability of any particular one being instantiated. For there is no further backdrop or underlying metaphysical framework which would provide warrant for the assumption that each (metaphysical) possibility has an ‘equal chance’ of occurring. The chance of any given possible world becoming actual is not like the frequency chance of a fair die rolling a six, or a roulette wheel spinning a 22. These probabilities are what they are because of measurable background conditions which determine objective chance frequencies. We simply don’t have epistemic access to such conditions when attempting to assess what the antecedent probability would be of Wn being actualized. For all we know, it may be a brute fact about certain worlds that they have a greater ‘chance’ than certain other worlds of being actualized. Even if we knew that there were a total of 1000 possible worlds, of which 500 were chance worlds and 500 non-chance worlds, and that of the 50 total e worlds 45 of them were non-chance worlds and only 5 of them chance worlds, we nevertheless could not correctly infer that P(h n/e) is .9. For all we know, it might simply be an inexplicable fact that non-chance worlds without e tend to become actual, or that chance worlds containing e have a greater chance than others of being actualized. Alternatively, it might be an inexplicable fact that non-chance e worlds have a greater antecedent chance of becoming actual than chance e worlds. The point is that in figuring P(h n/e) one is attempting to provide an O-assessment which, lacking an underlying framework of regularities or other set of objective rules for assigning a prior probability to h n or to the occurrence of e, affords one no intuitive grip on these probabilities and thus renders it impossible to objectively assign a conditional probability to h n/e. Attempting an accurate O-assessment fails in the case of evaluating the plausibility of (I-1), since we lack access to information necessary for deciding the issue. The result is that P(h n/e) is a priori undecidable, and hence BonJour’s case for the apriority of (I-1) is dubious.
5. The A Priori Status of (I-2)
Upon examining the alleged a priori status of BonJour’s second premise, we discover that the justification of (I-2) is beset by a similar problem. Above I claimed that in order for an application of (I-C) and IP to have even a .67 likelihood of being true, BonJour needs to establish a .81 probability for the truth of each of the argument’s premises. In Bayesian terminology, with respect to (I-2), BonJour needs to demonstrate a priori that P(h u/h n&e) = .81. Since P(h u/h n&e) = P(h u/h n), as argued above, our present task is to judge whether there is strong a priori reason for thinking that P(h u/h n) is at least .81. Given the truth of a non-chance cause and explanation of the inductive evidence, what is the probability of the hypothesis that there is an objective, uniform regularity among events and phenomena in the world which is reflected in the pattern or distribution of that evidence?
At this point BonJour considers certain non-chance alternative hypotheses to the uniform regularity hypothesis h u that might be thought to be capable of explaining the inductive evidence. These alternative scenarios—or “normal non-inductive explanations”, as he calls them—posit systematic and uniform regularities governing events and phenomena in the world, but which are not accurately reflected in the actual inductive evidence patterns.Footnote 22 A simple example of this sort of case involves the presence of some (unknown) factor C which causes the observed proportion of As that are Bs to deviate in a uniform way from the actual overall distribution. There is an objective regularity involving C, A, and B, such that when C occurs in a disproportionately high number of cases in the observation sample, the sample doesn’t reflect the actual proportion of As that are Bs. BonJour claims that it is, on the whole, unlikely for the epistemic situation implied by the normal non-inductive explanation to arise or persist.
In his critique of the argument, Brueckner raises two main objections to (I-2), both of which warrant further discussion. The first objection is that (I-2)’s claim that the most likely explanation of the truth of a standard inductive premise is the straight inductive explanation derives from the claim that the straight inductive explanation is a better explanation than all other rival (normal non-inductive) hypotheses. So in order to infer that the straight inductive explanation is probably true, we need to assume a principle of inference to the best explanation (BE): if a hypothesis H is the best explanation of a given phenomenon (when weighed against other competing hypotheses), then H is likely to be true and should be accepted. Brueckner objects that in order to be justified in believing the straight inductive explanation over rival normal non-inductive explanations, we first need a good a priori reason for accepting the BE principle itself—and BonJour has failed to provide us with one.Footnote 23
It would be wise, at this point, to take a step back and think for a moment about what BonJour is trying to accomplish with his a priori defense of induction. In setting out an a priori argument for the inductive principle, BonJour is certainly not attempting to give an a priori defense of each of the underlying logical principles involved in the reasoning. What he is claiming to do is to give an a priori defense of the two core substantive premises of his argument. With this in mind, a few comments are in order. First, BonJour takes the two locutions ‘best explanation’ and ‘highly likely to be true explanation’ to mutually entail one another in a purely analytic sense, and he employs them interchangeably in this way throughout his discussion of his core premises.Footnote 24 Given this usage of the two expressions, BonJour’s defense of the apriority of (I-2) need not provide a rationale for BE in addition to the one he offers in support of the premise itself. Furthermore, the considerations he adduces in favour of the straight inductive explanation as the ‘most likely explanation’ of standard inductive evidence provide direct support for that hypothesis and (I-2) without relying on a distinct best explanation principle. This is because BonJour’s a priori rationale derives directly from informally figuring the conditional epistemic probability of the hypothesis (straight inductive explanation) based on knowledge of the space of logical possibilities. Whether his reasoning ultimately succeeds here is a separate question which I will consider shortly.
Even if BE is regarded as a distinct underlying principle from which (I-2) is derived, BonJour can rightly view his project as one of establishing an a priori basis for the inductive principle (IP) given the a priori status of certain fundamental logical-explanatory principles. Logical arguments typically make use of an array of basic inference rules, such as Modus Ponens and BE. The epistemic status of such logical and explanatory principles is something that BonJour takes as a given in exploring the possibility of an a priori defense of IP. In the earlier chapters of his general defense of a priori truth, BonJour lays out his case for the a priori status of fundamental non-inductive logical and explanatory principles.Footnote 25 Granting this general framework of the a priori, BonJour can then go on to ask whether inductive principles such as IP can be given a priori justification. In furnishing such a justification, BonJour has a right to employ underlying principles which are part and parcel of the general a priori reasoning which he has previously argued is warranted. By flagging the best explanation principle implicit in (I-2), Brueckner’s worry unwittingly serves to pinpoint and refine the crucial issue facing BonJour: granted a general framework of a priori truth, can this framework then be extended to include a justification of basic inductive procedures? (I’ll argue shortly that it can’t; the nature of (I-2) itself is such that it doesn’t admit of a priori justification.) Brueckner’s present objection doesn’t address this question; the plausibility of the core claim expressed by (I-2) is left untouched. If we replace the expression ‘most likely explanation’ in BonJour’s wording of (I-2) with ‘best explanation’, the substantive claim of the argument remains—the straight inductive explanation is the best explanation of the inductive evidence.
This leaves Brueckner with one substantive objection to (I-2). He contends that BonJour dismisses the normal non-inductive explanation too quickly, and thus BonJour hasn’t shown that we have good a priori reason to believe the straight inductive explanation to be a better explanation than the normal non-inductive explanation. According to Brueckner, the way BonJour rejects any normal non-inductive hypothesis is by considering two possible explanations for it and then finding them both implausible. The conclusion to be drawn is that the straight inductive explanation is more likely to be true than the non-inductive hypothesis. The two possibilities BonJour considers for explaining a normal non-inductive hypothesis are (i) the existence of a connection between factor C and the observation of As, leading to a distribution of observed As that are Bs that is skewed in one direction, and (ii) a pure chance explanation of the skewed distribution. Brueckner objects that it is unfair to demand an explanation for the non-inductive hypothesis, and furthermore that BonJour offers no explanation for the truth of the straight inductive explanation itself. Perhaps the presence and distribution of the distorting factor C in the observation pool is a brute fact with no explanation at all; if so, then BonJour is not warranted in judging the normal non-inductive explanation to be unlikely solely because his two target possibilities are deemed to be implausible.Footnote 26
Brueckner’s worry loses its force once we realize that BonJour is not seeking an explanation for why the distorting factor C obtains or what the nature of C involves, and then ruling it out as implausible. In fact, he does not consider option (i)—the influence of observation on the presence of C—to be wholly implausible. What BonJour says is that any justificatory argument or defense of induction must face the possibility of observational influences on the evidence, whereas the classical problem of induction concerns the justification of generalizations from the observed to the unobserved “even when such influence by observation is assumed or stipulated not to be present.”Footnote 27 For the purpose of focusing on the distinctively Humean problem, BonJour brackets the issue of observational influence and sets it aside as a separate skeptical worry.
Turning to possibility (ii), when BonJour surmises that the non-inductive hypothesis might be true in virtue of the distribution of Cs in the observed cases deviating from the overall distribution of Cs purely as a result of chance, he is in essence conceding the possibility that such a hypothesis lacks a genuine explanation. He thinks that what is important in deciding between the straight inductive explanation and the proposed non-inductive hypothesis is not our ability to identify what further facts might explain why one or the other is true, but rather our ability to discern directly by a priori reflection that given a range of inductive evidence, a non-inductive explanation is inherently unlikely to be true. BonJour claims that apart from observational influences, regardless of what factor C represents, it will be unlikely, in general, that in the observed inductive sample the distribution of A, B, and C diverges uniformly from their overall distribution.Footnote 28 And this can be discerned from the armchair, so to speak, by simple reflection on hypothetical cases, given the truth of (I-1)—a non-chance explanation of the inductive evidence in terms of the causal properties and dispositions of objects and their relations.
To illustrate the a priori considerations that come into play in weighing the straight inductive explanation against the normal non-inductive explanation, let As represent a collection of manufactured toy dolls of a certain type, let B be the feature of being slightly disfigured, and let C be the (as yet undiscovered) factor of having been produced in a factory with a faulty mold.Footnote 29 Suppose that in a particular observation sample 30% of As are Cs, but only 5% of As are Cs in the total collection of dolls. This situation might arise, say, because the samples selected all came from the same factory, or it might arise for some other reason. Extrapolating from this sample would result in drawing the incorrect inductive conclusion that roughly 30% of As are Bs, because the distribution of As that are Cs in the sample collection is skewed and is significantly higher than in the overall distribution. BonJour insists that it is evident a priori that while this situation is certainly possible, it is inherently unlikely to the extent that most observation samples would tend to contain a more randomized sample of dolls. His reason for thinking so is that, apart from observational influences, the observation sample will tend over time to reflect the overall pattern of distribution of relevant features. Eventually, either the observed proportion of As that are Bs will gradually converge on the actual proportion as we expand the range of our inductive sample, or we will come to modify the inductive premise to include information about C in light of accumulating new knowledge of C and the lawful interactions in which C participates. If increasing and broadening the sample base fails to alter the proportion of As that are Cs and with it the proportion of As that are Bs, then the sample base likely reflects a distribution of As, Bs, and Cs which obtains in the realm “in which we live, move, and have our cognitive being.”Footnote 30 For most of the problems to which induction is applied, the inductive conclusion will, in effect, turn out to be true. Due to the fact that one of these situations will likely prevail, BonJour can claim that it is, on the whole, unlikely for the epistemic situation envisioned by the normal non-inductive explanation to arise or enjoy any prolonged duration.
What is especially pertinent here is that Brueckner’s remarks regarding (I-2) say nothing to counter this line of thought. The upshot is that Brueckner misidentifies the primary grounds to which BonJour appeals in rejecting the non-inductive explanation as improbable. Contrary to Brueckner’s complaint, the rationale which BonJour provides for rejecting the normal non-inductive hypothesis as improbable is not based on singling out certain possible explanations of this sort of hypothesis and then eliminating them. Rather, his case for the straight inductive explanation over normal non-inductive ones is direct, based on a direct a priori apprehension of the character of the two competing explanations.
Nevertheless, BonJour’s defense of (I-2) confronts a fundamental problem once we recognize that the straight inductive explanation and normal non-inductive explanation are not the only candidates for types of competing explanation. BonJour’s normal non-inductive explanation presupposes that an underlying system of uniform regularities and lawful interactions holds between elements such as A, B, and C. That is, he presumes that once we narrow down the range of possible explanations of the inductive premise to non-chance explanations, the only remaining explanations there are to consider are ones invoking some form of systematic regularity and order holding across all relevant spatial and temporal boundaries. And given a backdrop of such an objective and uniform order, we will deem it highly probable that our observed samples of inductive evidence reflect the actual distribution of phenomena in the world. Normal non-inductive types of explanation are ruled out as improbable because they assume an overarching framework of regularities with respect to which the likelihood of the occurrence of patterns of inductive evidence (or the occurrence of phenomena which deviate from expected patterns) can be discerned from the inside in accordance with an I-assessment of probability.
What BonJour’s discussion of (I-2) and normal non-inductive explanation is missing is a consideration of possible non-chance hypotheses or explanations of the inductive evidence which do not invoke uniform systematic regularities of the kind required to underwrite inductive inference. There are many broad classes or kinds of non-chance hypotheses of this sort. Such possibilities are not considered at all in BonJour’s discussion of normal non-inductive explanations, which views probability assessment from the inside. Taking into account such alternative explanations will significantly affect our a priori justification for assigning a sufficiently high probability to h u/h n. One type of explanation posits that patterns of inductive evidence observed are the result of the volitions of intentional agents situated within or outside the space-time universe. The choices and actions of these agent(s) are directly or indirectly responsible for the observed patterns, but nothing about what we observe or experience gives us a priori reason to think that the patterns of volition of these agents will persist beyond our current spatio-temporal range of observation. Call the set of hypotheses which accord with this type of explanation I-hypotheses. Another broad class of hypotheses postulates that within the natural world there exist certain built-in causal mechanisms or propensities responsible for the evidence patterns, for the fact that m/n As are Bs, but these propensities will change in the future or beyond the current observed spatial range. Thus we have a family of P-hypotheses which rival h u. For example, a propensity or mechanism might cease to exist or evolve into a different propensity, resulting in new regularities emerging. The inductive evidence exhibits a pattern (m/n proportion of As are Bs) up to the present time t, but due to chance or some further causal factor C which influences the As or Bs in a systematic way, the regularity is altered after t and ceases to be reflected in the original evidence set. After t, the world instantiates a different overarching pattern of order and regularities. Still another sort of possibility is that the relevant causal propensities or mechanisms cease to hold altogether beyond a certain spatial region or approaching temporal point, and from there chance takes over. Call these C-hypotheses. In all of these scenarios, robust non-chance regularities and propensities hold within a limited spatial and temporal region, for example within spatio-temporal region R, but not beyond R.
Consideration of these possibilities underscores the problem of assessing the probability of h ufrom the outside, that is, the problem of providing an O-assessment of (I-2) that renders it justified a priori. BonJour’s argument claims that (I-2) is known a priori in the sense that its truth is evident on rational reflection. But certain types of alternative explanations, such as the ones briefly mentioned above, imply that while the distribution of the inductive evidence reflects a corresponding objective regularity which held in the past (or which held during the extended time interval in which the evidence occurred), the regularity will not continue to hold in the future. Each of these types of scenarios or hypotheses offers a non-chance explanation of the observed distribution of evidence, but, by avoiding invoking a system of overarching and uniform regularities which holds across sufficiently broad spatio-temporal boundaries, falls short of the kind of explanation—encapsulated by the uniform regularity hypothesis—which is required to license an inference to IP or (I-C). These possibilities invite the question of precisely what sort of claim is correctly implied by standard inductive evidence. Granting, for the sake of argument, the plausibility of (I-1) along with BonJour’s preemptive attempt to show that certain types of normal non-inductive explanations are a priori unlikely, the most that one can conclude a priori, given the occurrence of a range of inductive evidence, is that the observed proportion m/n reflects an objective regularity that existed in the observed past (and inside the observed spatial region). Let’s define an inductive spatio-temporal world segment (STWS) as a certain tightly defined spatial region and segment of the temporal order, whose outer boundaries are demarcated by either (i) the range of spatial and temporal regions in which the particular events and phenomena referred to by a given standard inductive premise occur, or (ii) the relevant spatio-temporal region(s) referred to by a corresponding standard inductive conclusion. And let’s call an STWS which satisfies specification (i) a P-bounded STWS, and an STWS which satisfies (ii) a C-bounded STWS. We shall say that inductive evidence obtains when certain observations are made and empirical data gathered which come to constitute standard inductive evidence. Now what appears evident a priori, on reflection, given BonJour’s standard inductive premise, is not (I-2), but rather this revised thesis:
(I-2*) The most likely explanation for the truth of a standard inductive premise is that the observed proportion m/n accurately reflects a corresponding objective pattern in the inductive spatio-temporal world segment in which the standard inductive evidence obtains, that is, in the relevant P-bounded STWS.
We can see that (I-1) together with (I-2*) clearly do not entail either (I-2) or (I-C). For BonJour has given us no purely a priori reason for thinking that an objective regularity of the sort allegedly implied by an inductive premise, even if such a regularity existed in the observed past, will continue to hold in the unobserved future. And if there is no a priori reason for thinking that the regularity extends into the future, then BonJour’s argument has not established that there is an overarching system of order or regularities sufficient to ground induction.
In order to derive the desired conclusion (I-C), BonJour’s argument requires an additional a priori premise, which, in conjunction with (I-2*), will yield (I-2). The required premise is something along the lines of:
(I-3) It is (a priori) likely that objective patterns or regularities which hold in a P-bounded STWS will continue to hold in its corresponding C-bounded STWS.
What good reason do we have for thinking that (I-3) is true? Its truth doesn’t appear to be evident a priori, for there are explanations for the objective regularities which hold in the given P-bounded STWS—the I, P, and C hypotheses we considered above—which rival the straight inductive explanation and are incompatible with (I-3). We have no a priori reason for regarding these explanations as less likely than the straight inductive explanation BonJour favours. Since these other hypotheses are not known to be false or improbable a priori—their falsity or unlikelihood is not evident to reason alone—and they are strictly incompatible with (I-3), it follows that (I-3) is not justified a priori. The existence of a certain sort of objective order or regularity which is uniform across space-time is no more a priori likely given a standard inductive premise than is the existence of an objective order which holds up to the present time t but fails to hold beyond t. Nor can we marshal support for (I-3) by pointing out that objective regularities which have held for P-bounded STWSs in the past have tended to hold for their corresponding C-bounded STWSs, and then go on to infer that such regularities will continue to hold for other C-bounded STWSs; for that would constitute a form of inductive reasoning which presupposes the truth of IP, hence begging the question.
As we did earlier with (I-1), we can articulate the case against the a priori status of (I-3), and by implication (I-2), by stating our objection in terms of possible worlds. Within the total set of possible worlds in which standard inductive evidence obtains, there are innumerable worlds that contain certain features which make it unreliable and inappropriate for us to continue to follow the sorts of inductive procedures with which we are familiar. These are worlds in which reasoning by straight inductive explanations would be, on the whole, unsuccessful, although such reasoning would succeed up to a point. In these worlds, the standard inductive evidence has some non-chance explanation which is due to some unknown feature which allows standard inductive explanations to be successful up to a certain limit but no further. (In fact, there are possible worlds in which following certain normal non-inductive practices are in the long run more successful). There are many worlds in which a given objective system of regularities R holds in the relevant P-bounded STWSs but fails to hold in the corresponding C-bounded STWSs, such as, for example, worlds in which the I, P, and C classes of hypotheses are true. Now, if there are many such possible worlds (even if ours is not), then, since the standard inductive premise is true in all of them, it is not the case that the occurrence of the inductive evidence tells us a priori which sort of world ours is likely to be. So since it is not likely, based on any compelling a priori reason or argument I can think of, that observed patterns or regularities will hold in C-bounded STWSs, neither is it justified a priori that the best explanation for the truth of an inductive premise is the straight inductive explanation.
In discussing (I-1) above, I questioned the assumption that we can simply equate the prior probability of a given hypothesis with the fraction of the total number of metaphysically possible worlds in which the hypothesis is true. But were we to grant this assumption for the sake of argument, we would still lack a priori grounds for assigning a sufficiently high probability to (I-3) or h u/h n. To see this, consider what happens when we attempt to compare the likelihood of h u given h n with the likelihood of our alternative rival hypotheses given h n. Given k metaphysically possible worlds, the antecedent probability of any particular world being actual would be 1/k. If there are n worlds in which h u is true, then the antecedent probability of h u is n/k. The prior probabilities of our alternative sets of hypotheses can be represented in the same way, by delineating various classes of non-chance worlds—for example, I-worlds in which the volitions and actions of intentional agents are responsible for the distribution of the inductive evidence, and other classes of worlds (P-worlds and C-worlds) in which objective regularities hold within the region(s) of space-time of the inductive evidence but cease to hold beyond that region. Keep in mind that I, P, and C hypotheses represent classes of hypotheses, and I, P, and C worlds are classes of possible worlds in which the particular I, P, and C hypotheses are true. There are undoubtedly other non-chance possibilities that we haven’t considered here, but to simplify the argument, we can stipulate that the worlds in which h n is true include h u-worlds, I-worlds, P-worlds, and C-worlds. The latter three classes of possibilities are worlds in which the correct explanation of a given standard inductive premise is something other than the straight inductive explanation. They are all worlds in which the truth of the inductive premise is accounted for without there being the type of overarching and uniform system of regularities which is required in BonJour’s argument in order to ground (I-3) and, in turn, to buttress support for (I-2). So what we want to know is whether we have good a priori reason for thinking that the number of h u-worlds far exceeds the number of I, P, and C worlds, such that P(h u/h n) approaches .81. Since we have stipulated that P(h n/h u) = 1, our applied Bayesian formula states that P(h u/h n) = P(h u) / P(h n).Footnote 31 So speaking in terms of fractions or ratios of possible worlds, we can ask whether we have an a priori reason for thinking that the fraction of h u-worlds over h n-worlds is .81.
The a priori status of this claim is dubious at best. In order for P(h u/h n) to approach .81, the class of h u-worlds must contain at least four times as many members as the total class which is the union of the I-worlds, P-worlds, and C-worlds. Does this ratio hold? It doesn’t seem clear, on the basis of rational reflection alone, what the answer should be. Call an affirmative answer to this question AA. AA is what has to be the case in order for the likelihood of h u/h n to approach .81. Since AA is nothing more than a simple possible-worlds elaboration of BonJour’s probability claim encapsulated in (I-2), then if (I-2) were immediately evident on rational reflection, so too would AA be evident in this way. But even if AA is true, its truth is not plainly evident to reason alone; hence AA is not justified a priori. And from this it follows that (I-2) does not possess adequate a priori justification. To repeat the point in a slightly different way: BonJour hasn’t provided a good a priori reason for thinking that within the total class of possible worlds (in which the standard inductive evidence obtains) the ratio of h u-worlds to I-worlds ∪ P-worlds ∪ C-worlds is at least 4:1. For all we know from a purely a priori standpoint, there could be just as many I-worlds, P-worlds, or C-worlds as there are worlds which instantiate the appropriate uniform and objective order. BonJour has not established a priori that worlds in which the straight inductive explanation is successful are more numerous, or are more likely to be instantiated, than worlds in which the straight inductive explanation fails. So for all we know a priori, the likelihood of the straight inductive explanation involving h u could fall well short of .81. The upshot is that BonJour hasn’t provided us with any good a priori reason for supposing that the sort of objective order that licenses drawing standard inductive inferences is likely to obtain given the existence of the inductive evidence. Hence, contrary to what BonJour alleges, (I-2) does not constitute an a priori fact, even if it turns out to be a fact at all.
In order to establish (I-2) and get BonJour’s argument to work, we need to find some a priori reason that supports (I-3). We need a good reason for thinking that objective regularities observed in P-bounded STWS—regularities that hold in roughly the whole spatio-temporal region or segment spanning the observations—continue to hold in spatio-temporal segments referred to in the conclusions of inductive arguments (C-bounded STWS). One way to do this would be to set forth some well-supported metaphysical thesis which gives an account of a robust conception of objective regularity that is something stronger than mere Humean constant conjunction, and which at the same time implies that the objective regularities won’t change or differ outside our P-bounded STWS. An adequate metaphysical theory which explicates a robust conception of objective order or necessary connection would avoid skepticism regarding whether the relevant regularities will continue to hold in the future, and would thereby establish the straight inductive explanation as the most likely explanation of the standard inductive premise. In short, BonJour needs to establish something like the following thesis:
(MT) There is some true metaphysical thesis M which entails (or renders it highly probable) that objective regularities which hold in any P-bounded STWS will probably hold in the corresponding C-bounded STWS.
In a brief paragraph, BonJour alludes to a preemptive reply of just this sort when he surmises that elucidating a non-Humean, metaphysically robust conception of objective regularity does not seem insurmountable (1998, p 214). But BonJour needs to do more than just raise the possibility of developing such a reply; he also needs to provide an a priori justification for MT or its kin. For in claiming that (I-2) or (I-3) is known or justified a priori, BonJour means that their truth is evident on reflection or that the reasons adduced in support of them are also known or justified a priori. I have already shown why I think that neither thesis is immediately evident on rational reflection. To the extent that the justification of (I-2) or (I-3) involves appeal to a substantive metaphysical thesis such as MT, then that thesis must be justified a priori. The underlying principle at work here is that if S is justified a priori in believing that q, and S’s sole justificatory reason or basis for believing q is p, then S is justified a priori in believing that p. This is a principle that BonJour himself would accept. I won’t attempt here to defend this principle in general, but this particular instance of it is surely plausible.Footnote 32
So in order to claim a priori status for (I-3), (I-2), and for the overall argument, MT itself must be known or strongly justified a priori in BonJour’s sense. However, even if the content of such a thesis can be clearly explicated, making the case for the a priori true status of the thesis—a metaphysical account which supports the required system of robust, objective regularities—proves to be no easy task. It does not seem to be the case that we know by a priori means, either self-evidently or by a priori argument and reflection, that the truth of a thesis like MT is even remotely probable. In order to establish (I-3) (or (I-2)) and for the argument to succeed, not just any plausible metaphysical thesis will do, but one which a) is a priori likely to be true, in the sense BonJour accepts that it is either a priori evident to reason itself, or based on premises which are a priori evident to reason; b) gives an account of objective regularities that is a priori likely to be true, and c) entails that these objective regularities hold (for the most part) in the unobserved past, present, and future. It is doubtful that there is a metaphysical theory that meets these conditions—in which case MT does not enjoy strong a priori warrant. Furthermore, if we compare IP or (I-C) with MT and ask which is more epistemically basic, the correct answer would seem to be that IP and (I-C) are more epistemically basic, on par with principles of logical reasoning and other principles of rationality. If so, then to argue for IP and (I-C) on grounds such as MT would be to employ a misguided strategy. The upshot is that without establishing the a priori likelihood of MT, then from a purely a priori standpoint, the truth of (I-3) (and therefore (I-2)) cannot be established. And we are left with the additional implication that I, P, and C-type hypotheses appear a priori to be equally plausible explanations of the standard inductive premise as h u is. If so, then (I-2) is a priori undecidable. I conclude that, with respect to establishing the apriority of (I-1) and (I-2), BonJour’s argument for (I-C) fails on both counts.
6. Concluding Reflections
A priori reasoning of the sort exemplified in BonJour’s argument for IP is highly problematic. Arguments which attempt to identify an a priori discernible connection between propositions about observed patterns of evidence and propositions or beliefs whose content goes beyond direct empirical experience appear to be mistaken. They are mistaken because the plausibility or probability of observation-transcendent inferences cannot be assessed ‘from the outside’, apart from some presumed framework which already includes considerable observation-transcendent knowledge. Hence the plausibility of an alleged connection between the observation-sentences and the propositions or beliefs which are extrapolated from those sentences is a priori undecidable. Hume is therefore vindicated with respect to at least part of his reasoning for P2. There are other types of a priori arguments for induction, such as the linguistic argument set forth by Strawson, which defends the rationality of induction based on an analysis of the ordinary meaning of a family of interrelated concepts.Footnote 33 My arguments do not address ‘analytic’ a priori attempts to argue for induction.
The failure of BonJour’s sort of argument does not mean that no successful a priori argument for induction is possible, let alone imply that induction is not known or justified in some way. Given a more modest foundationalism, the justification of induction may not require an argument to show that inductive processes will continue to be truth-conducive. We do not require this kind of warrant for other types of cognitive processes which we take to be reliable, such as perception or memory, none of which can be argued for non-circularly. A case can be made that belief in IP and other inductive principles are among our basic beliefs, whose warrant derives from their describing certain reliable belief forming processes. If so, then BonJour’s a priori strategy is motivated by the fundamentally flawed assumption that justification of induction requires rational argument. Perhaps we would do well to abandon that assumption.
