Discussion Note: McCain on Weak Predictivism and External World Scepticism
- First Online:
- Cite this article as:
- Harker, D.W. Philosophia (2013) 41: 195. doi:10.1007/s11406-012-9409-y
- 90 Views
In a recent paper McCain (2012) argues that weak predictivism creates an important challenge for external world scepticism. McCain regards weak predictivism as uncontroversial and assumes the thesis within his argument. There is a sense in which the predictivist literature supports his conviction that weak predictivism is uncontroversial. This absence of controversy, however, is a product of significant plasticity within the thesis, which renders McCain’s argument worryingly vague. For McCain’s argument to work he either needs a stronger version of weak predictivism than has been defended within the literature, or must commit to a more precise formulation of the thesis and argue that weak predictivism, so understood, creates the challenge to scepticism that he hopes to achieve. The difficulty with the former is that weak predictivism is not uncontroversial in the respect that McCain’s argument would require. I consider the prospects of saving McCain’s argument by committing to a particular version of weak predictivism, but find them unpromising for several reasons.
KeywordsExternal world scepticismWeak predictivismKevin McCainCartesian demon
In a recent paper McCain (2012) argues that weak predictivism undermines a familiar method for motivating external world scepticism. Consider the much discussed possibilities that I might be a brain in a vat, or am being deceived by a Cartesian demon. In discussions of external world scepticism these are advanced as rivals to our common sense belief that there is a mind-independent world of which we can achieve significant knowledge. The sceptical hypotheses account for all our sensory experiences. Thus, since there is nothing in our experiences that enables us to dismiss them in favour of the commonsense view, the sceptic concludes that our commonsense beliefs about the world (CS) are not justified. A central component of McCain’s objection is the idea that CS grounds many correct predictions concerning both future experiences and experiences involving other senses. My current, gustatory experiences with a cup of coffee lead to expectations concerning the taste of future gulps from the same cup, for example; my observation of steam rising from the coffee encourages convictions about how the cup will feel in my hand, how the steam will feel on my upper lip, and how the coffee will feel in my mouth. More generally, according to McCain, CS predicts certain kinds of continuity and coherence that many radical, sceptical hypotheses do not, including that of the Cartesian demon; the latter can only accommodate much of what CS correctly predicts.1 The significance of this observation occurs when McCain combines it with a thesis from the philosophy of science. According to predictivists, scientific theories that correctly predict data are better confirmed by those consequences than competing theories that can only accommodate the data. McCain claims to invoke only a weak version of predictivism to conclude that commonsense beliefs, in virtue of their predictive successes, are better supported by our sensory experiences than these infamous, radical, sceptical hypotheses. The latter might account for all our sensory experiences; it doesn’t follow that they are equally plausible in light of our experiences. McCain concludes that commonsense beliefs are better confirmed than many sceptical alternatives.
McCain concedes that his argument doesn’t refute all forms of scepticism, but does assert that the argument is effective “against sceptical hypotheses as they are typically formulated”, and “that CS is better supported by our sensory experiences than a large class of sceptical hypotheses … [including] the sceptical hypotheses that are typically put forward in arguments for scepticism.” McCain is thus pursuing what Huemer (2000) describes as an aggressive, rather than defensive, response to scepticism. Aggressive responses, for Huemer, require arguments that the sceptic’s position is misleading; defensive responses, by contrast, argue that the sceptic’s argument is unpersuasive from the perspective of at least some epistemic attitudes. McCain is pursuing the aggressive strategy, arguing that the sceptic is mistaken in assuming that our sensory experiences provide no reason for preferring CS over many standard sceptical hypotheses.
McCain’s argument is intriguing, but his use of weak predictivism is problematic, I’ll argue, and critically so. He describes weak predictivism as uncontroversial. However, to assume weak predictivism in the sense that would support his argument is not uncontroversial; the sense in which weak predictivism is, plausibly, uncontroversial doesn’t support the argument that McCain wishes to defend, at least not without significant further argumentation that McCain doesn’t provide. I explore ways in which McCain’s argument might be strengthened, but find the prospects of developing the argument along such lines unpromising for several reasons.
“Uncontroversial” Weak Predictivism
Within the predictivist literature a distinction has been drawn between strong and weak versions of the thesis.2 Predictivists contend that verified predictions provide better confirmation for a theory than the consequences of that theory which are not predicted, but might in all other respects be equivalent. The consequences of a theory that are not predicted are usually described as having been accommodated by the theory.3 Strong predictivist theses suppose there is something intrinsically more valuable about predictions, for purposes of confirmation. For example, if only a theory’s predictions provide a genuine test of the theory, then the strong thesis might seem justified. Weak predictivist theses propose that verified predictions are more reliable indicators (than accommodations of data) of further virtues or reasons, which are themselves relevant to our assessments of available hypotheses. Mayo (1996) suggests that predictive success seems more valuable only because it is positively correlated with more severe tests of a theory or hypothesis.4 Lipton (2004) argues that predictive success is an indicator that a scientist has not ‘fudged’ the theory to make it fit available data.5 Lange’s (2001) suggestion is that successful predictions imply a more unified, rather than arbitrary, hypothesis. Hitchcock and Sober (2004) argue that in certain well-defined circumstances predictive success is a useful and reliable indication that theorists aren’t guilty of overfitting available data. Barnes (2008) suggests that predictions can indicate the availability of further evidence, for a particular theory, that some agents might be unaware of. In all these cases the advocate for weak predictivism anchors predictive success to a further, apparent, theoretical virtue, scientific goal, or cause for optimism. The excess value of predictive success, beyond that of mere accommodations, is thus understood as a product of the former’s greater but imperfect correlation with a further quality that itself has evidentiary value. Now, one might wonder why the additional evidence, fudging, unification, over-fitting and severity of tests are not evaluated directly, if these are salient for purposes of evaluating a theory or hypothesis. The response from some weak predictivists is that we may not always be positioned to make direct evaluations of these qualities, and in such circumstances might benefit from indirect means of assessing them.6 Whether such circumstances often, rarely, or never present themselves is typically not much addressed by weak predictivists.
Given the different ways in which weak predictivism has been understood, it is perhaps better thought of as a cluster of theses. These have in common the idea that predictive success, in certain circumstances, is evidence for the presence or absence of further virtues, but what the implied virtues are, in what contexts they matter, and how often we rely on such proxies, could differ significantly from one version to another. These predictivist theses of Mayo, Lipton, Lange, Hitchcock and Sober, and Barnes haven’t been much challenged within the literature. It is reasonable to describe their shared ground as uncontroversial. However, the suggestion that in some circumstances predictive successes are a more reliable indicator of some further theoretical virtue, and thus that it might sometimes be appropriate to utilize predictive success to assess those further considerations, is a weaker thesis than McCain requires. Weak predictivism understood in this sense is compatible with there being very few circumstances in which predictive success confers greater confirmation than accommodation of data.7 It is quite consistent with there being circumstances in which accommodated data provide better support for a theory than predictions. McCain quotes Hitchcock and Sober (2004) approvingly, but these authors argue explicitly that there are circumstances when predictions count more, circumstances when it doesn’t matter whether evidence was predicted or accommodated, and circumstances when accommodated data provide greater support. They also explicitly take no stance on how frequently these distinct types of circumstance arise. Similarly, Lipton (2004) admits that if we can establish that fudging has not occurred, then the advantage of predictions over accommodations washes out. Barnes (2008) describes a variety of circumstances where predictions will carry more weight than accommodations, for particular evaluators of scientific theories. The details are important, however. For many evaluators, particular predictions might have no excess value beyond the accommodation of the same data.
The general point is that, although a variety of plausible weak predictivist theses have been advanced, none imply that predictive success will always, or even typically, reliably gesture us towards the better confirmed hypotheses, from among empirical equivalents. Weak predictivists restrict the domain of their thesis to certain contexts. Thus, if McCain is assuming weak predictivism holds in all, or even most, circumstances where empirically equivalent hypotheses are being evaluated, then he is making a sufficiently strong claim that it can no longer be admitted as uncontroversial. From the sense in which weak predictivism is uncontroversial it doesn’t follow that predictive success confers a favourable epistemic status on CS over sceptical alternatives. In fact, within the context of this discussion there appears little difference between declaring that predictive success is intrinsically more valuable than accommodating (i.e. strong predictivism), and assuming that predictive success typically implies some further but unspecified virtue, which has evidentiary import.8 Although McCain attempts to distance himself from strong predictivism, his argument’s reliance on merely weak predictivism is questionable.9 If the appeal to weak predictivism is to serve as more than lip-service, then McCain must be more specific about how he understands the thesis. In summary, being vague about the details weakens McCain’s argument, because every advocate for any weak predictivist theses argues only that in some circumstances it is reasonable to weight predictive success more heavily, and we have been given no reason to suppose that any of those circumstances include our assessment of CS and sceptical hypotheses.
Particular Weak Predictivisms
The problems rehearsed thus far motivate an obvious question: can we strengthen McCain’s argument through greater specificity, by appealing to one of the aforementioned weak predictivist theses? The first problem for this strategy is that weak predictivism holds interest only because of the possibility that there might be circumstances in which we have incomplete access to the actual support a theory deserves, rendering it reasonable to use imperfect indicators of such support. In the case of complex, scientific theories, such circumstance seem quite possible. We might lack expertise to evaluate whether a given theory has been gerrymandered to fit certain observations. We might not know whether a curve was fit to the data in a manner that properly balanced the number of adjustable parameters against the closeness of fit. We might have reason to believe that experts are in possession of evidence that we are unaware of. In the case of CS and the sceptical hypotheses, however, we can readily access the details of the hypotheses and the available evidence. We can assess directly whether such hypotheses has been gerrymandered, for example. We also have no reason to suppose that the sceptic is ignorant of certain evidence for CS that others possess.10 We don’t need predictive successes to serve as an imperfect guide to the virtues and vices of CS and sceptical alternatives.11
[t]he scientist knows the answer she must get, and she does whatever it takes to get it. The result may be an unnatural choice or modification of the theory and auxiliaries that results in a relatively poor explanation and so weak support, a choice she might not have made if she did not already know the answer she ought to get.
The sceptic didn’t fudge, as Lipton seems to define the term, their hypotheses or auxiliaries. Modifications aren’t needed. The sceptic doesn’t confront choices concerning how the theory should be developed, or what data should be ignored. Every sensory experience is explained by, and entailed by, the claim that the Cartesian demon created that experience for us. Admittedly, we likely feel some tendency towards regarding the sceptical hypotheses as providing poorer explanations, but this isn’t the result of Lipton-style fudging.
Similarly, consider the connection Hitchcock and Sober draw between predictive success and the problem of overfitting. Data sets are very often noisy; curves that are fit to available data can do a better or worse job of ignoring that noise; Hitchcock and Sober show that in certain circumstances, and in the absence of more direct information concerning how a curve was generated, predictive success can provide helpful information about whether the data were overfit. The Cartesian demon hypothesis, however, doesn’t overfit anything. McCain can’t plug this particular version of weak predictivism into his argument, for the thesis simply has nothing to say about the value of predictive success outside questions of fitting curves to available data.
Barnes’ (2008) weak predictivist thesis departs from the predictivist tradition in two key respects. First, Barnes defines a prediction as a certain act, performed by humans who are sufficiently confident in some theory that they endorse its consequences, rather than an implication of a theory. Second, the significance of predictive success can vary from one agent to another, depending on various factors. Barnes’ discussion is significant for present purposes because he briefly considers the possibility of utilizing predictivism to answer sceptical challenges.12
For Barnes, there are two ways that predictions can confer better support than accommodations.13 The first, roughly, involves the plausible inference that if a competent scientist is willing to predict certain phenomena, on the basis of some theory, then they must have good reasons for doing so. Consequently, for others, unsure about endorsing that particular prediction, the implication that there are good reasons for doing so might increase their confidence in the underlying theory. This role for predictive success can’t help McCain. We all share the commonsense beliefs that McCain is concerned with. There are no experts on CS that might have additional reasons for endorsing CS that we can only indirectly infer. The second distinctive role that predictive successes play, for Barnes, is the way they redound to the credit of background assumptions that are peculiar to those who endorse certain predictions. To illustrate the idea, Barnes provides an extended discussion of Dmitri Mendeleev and his prediction of the discovery of new chemical elements on the basis of his periodic table. Barnes describes several assumptions that Mendeleev held that were not commonly held by his peers. These rendered the discovery much more probable. The importance of Mendeleev’s predictive success was that his background assumptions were better confirmed than those of his contemporaries. But this doesn’t help McCain either, because all our experiences are entailed by the sceptical hypotheses just as they are entailed by commonsense beliefs. We might concede that CS predicts what sceptics only accommodate, but our sensory experience are not less probable, given the sceptical hypotheses. New sensory experiences thus don’t provide any reason to prefer one set of background assumptions over another.
Our experiences don’t provide a more severe test for CS than sceptical hypotheses, in Mayo’s sense, nor is the sceptic guilty of advancing “an arbitrary conjunction”, as Lange defines it. Weak predictivist theses posit a correlation between predictive success and further virtues, rendering predictive success potentially useful if we hope to evaluate hypotheses with respect to those virtues. Consequently, however, such predictivist theses only have value, for purposes of evaluating competing hypotheses, when the hypotheses under consideration can better or worse fulfill the associated virtues. With CS and radical, sceptical hypotheses this condition is not met, given available, weak predictivist theses. My objection here has nothing to do with the merits of the various weak predictivist theses themselves, only with their value for purposes of evaluating external world scepticism. Of course it’s possible that McCain might devise a new weak predictivism, relating predictive success to considerations that are relevant to evaluating CS and sceptical hypotheses. Available theses are not up to the task.
Finally, suppose we can overcome the problems described thus far. It might require yet further argumentation to convince the sceptic that CS is therefore more likely to be correct. Hitchcock and Sober, for example, are explicit in adopting an instrumentalist perspective throughout their discussion. Predictive success indicates the absence of overfitting, but the latter is not regarded by Hitchcock and Sober as reason to suppose that the curve under consideration is more likely to be true. Concerning Lipton’s and Lange’s theses, suppose we admit that the predictive success of CS suggests less fudging or a less arbitrary conjunction. For McCain’s argument to unsettle the sceptic, the latter would also have to concede that these features are indicative of more plausible hypotheses. But there seems little reason to suppose that this is anything the sceptic would admit. The epistemic significance of both unification and simplicity, and the prohibition on ad hoc hypotheses, have received much attention. Lipton’s concern with fudging and Lange’s preference for non-arbitrary conjunctions are related to these concepts. The problem, however, is that for all the attention these ideas have received no consensus has emerged concerning either how we should understand the suggestion that one theory is more unified, or less ad hoc, or why these are relevant to our evaluations. Thus, even supposing that we cannot evaluate directly the extent to which sceptical hypotheses are fudged, for example, and thus must rely on imperfect indicators; supposing furthermore that we can provide evidence that the predictive success of CS indicates less fudging, in some sense that would require spelling out; still, the sceptic might require a further argument that the absence of fudging, appropriately defined, is reason to regard a hypothesis as more likely to be true. I’ve now identified three independent ways in which weak predictivism, as it’s currently understood, falls short of challenging external world scepticism.
The sense in which weak predictivism does appear uncontroversial is insufficient for us to infer that CS is better confirmed than empirically equivalent, sceptical hypotheses. Weak predictivist these asserts that in some circumstances, predictive success can more reliably indicate the presence of additional, epistemically relevant, considerations. This is a weak reason for supposing that the predictive successes of CS is reason to regard sceptical hypotheses as less well confirmed by our sensory experiences. Second, weak predictivist theses are argued to have value, only because we might not always be situated to evaluate certain features of a theory directly. With CS and the sceptical hypotheses we are not hindered in this respect, so there is no motivation for using predictive success as an imperfect guide, rather than assess the implied virtues directly. Finally, those virtues that are correlated with predictive success, according to the theses of Mayo, Lipton, and so on, appear both hard to apply to our assessment of external world scepticism, and unlikely to persuade. Weak predictivism, I conclude, creates no challenge for sceptics.
Within this note I’ll concede this suggestion of McCain’s, although I suspect it’s less straightforward than he admits. For example, if sceptical hypotheses are offered as evidence that the content of my memories cannot be trusted, then it is unclear whether I can be sure that I predicted anything at all.
Exactly how we are to distinguish predictions from accommodations is a matter for debate. McCain is not explicit, but I suspect he understands predictivism heuristically—what distinguishes predictions from accommodations is something about how each are used in the construction or development of a theory. Heuristic predictivism has emerged as the most influential version of predictivism within philosophy of science in recent years, so it is a reasonable version for McCain to adopt. My objections don’t rely on attributing to McCain this interpretation.
The concept of a severe test is central to Mayo’s ideas on confirmation.
Lipton identifies his own predictivist thesis as strong, rather than weak. On what has become the standard means of distinguishing strong from weak versions, however, Lipton’s qualifies only as a weak predictivist thesis.
McCain’s presentation of weak predictivism includes neither the observation that advocates for the thesis suggest only that predictive success is important in some circumstances, nor that distinct predictivist theses relate predictive success to different qualities.
Suppose, to illustrate, scientists A and B disagree about which of two rival hypotheses are best supported by available evidence. Scientist A notes a feature of one hypothesis that is not shared by the second hypothesis and appeals to this feature as reason to justify their ranking. Scientist B responds that this feature has no intrinsic value, for purposes of evaluating these particular hypotheses. It would be unsatisfying for A to concede the point but add that, nevertheless, this feature implies some further, unspecified, quality that does have evidentiary value. McCain commits the same mistake as scientist A.
It is beyond the scope of this paper to review the literature on strong predictivist theses. It is worth noting that the majority of predictivist defenses in recent years have been of the weak variety. For critiques of the strong version see Barnes (2008) and Harker (2008).
An important advantage of predictions, for Barnes, is their implication that a competent judge must have good reasons if she predicts some outcome on the basis of a particular theory. Evaluators of that theory who are unsure of its merits might regard the competent predictor’s endorsement of the predictions as evidence for evidence, i.e. evidence that there are further reasons (albeit unknown to the evaluator) for being favourably disposed towards the theory. The evaluator might therefore increase their degree of confidence in the theory. If the significance of the predictions of CS is to be understood in the same terms, however, then the sceptic must be ignorant of reasons that we possess for making certain predictions. But in this case the burden is clearly on us to point out the reasons themselves for our confidence in CS, rather than relying on an indirect argument that because we predict certain experiences we probably have good reasons for doing so.
There is again here the suspicion that McCain is guilty of conflating strong and weak predictivist theses. If predictive success is evidence that the successful theory is better confirmed than empirical equivalents that accommodate data, then McCain’s argument works. However, this is stronger than weak predictivism as it is typically understood. If predictivism is evidence of further considerations that are themselves epistemically relevant, then McCain owes an explanation for why we can’t evaluate those considerations directly, in the case of CS.
Barnes distinguishes virtuous from unvirtuous predictors. His analysis of unvirtuous predictors is similar, as he acknowledges, to Lipton’s fudging explanation, so I’ll focus here on his original contributions concerning virtuous predictors.
My thanks to two anonymous referees from this journal for helpful comments on an earlier draft.