Scientific realism regarding a given theory can be understood as a common-sensical position holding that both observable and unobservable entities postulated by said theory exist and behave more or less according to how the theory says they do—that is, that the theory is (approximately) true. Arguably, the most influential argument supporting this form of realism is that it provides us with the best explanation for the success of science. However, there is a gap in the argument between a theory being the most explanatorily and predictively successful and its being (approximately) true. Rather than bridging this gap, John Wright’s An Epistemic Foundation for Scientific Realism offers a novel alternative route to realism that does not make use of inference to the best explanation (IBE).
The book can be naturally divided into three parts. The first (Chapters 2–3) is a rebuttal of sceptical arguments against scientific realism including the problem of induction, the pessimistic meta-induction, the problem of unconceived alternatives, the underdetermination thesis, the experimental regress problem, the theory laden-ness of observation, and so on. In the second part (Chapter 4), Wright considers and rejects IBE as an argument for realism, along with attempts to further defend the argument via the notions of novel predictive success, concordance, simplicity, unification, scope, precision, and so on. Part three (Chapters 5–8) constructs an alternative defence of scientific realism that does not make use of IBE and applies it to examples from science including the atomic theory of matter and modern cosmology. It is in here that Wright’s main novel and positive account appears. I will illustrate the crux of the matter with an example (based on Chapter 7), namely, a route to realism about molecules without IBE. Consider the following:
All observable objects move in accordance with Newton’s laws of motion.
It is likely that there are unobservable objects that move in accordance with Newton’s laws of motion. Call these Mol1.
Why is it likely? If said unobservable objects did not accord with Newton’s laws then the domain of those laws would have happened to have coincided with the limit of perceptibility, which seems a priori unlikely. Wright calls this type of inductive inference an ‘Eddington inference’.1 The basic idea, which I will discuss further below, is that we can extrapolate into the realm of the unobservable given the assumption that our observations were blindly or randomly chosen.
The kinetic-molecular theory of heat conjectures that there are tiny particles too small to see that are responsible for various gas laws and Brownian motion. Call these Mol2. As realists, we want to know if Mol2 actually exist.
Given some other suppositions, we can derive the gas laws and explain Brownian motion with Mol1.
It is likely that the unobservable objects that we believe exist via an Eddington inference are the same as the tiny particles too small to see that are conjectured by the kinetic-molecular theory of heat. Thus, Mol1 = Mol2 = ‘molecules’, which exist and behave in the way that theory describes (for example, they obey Newton’s laws and Laplace’s law of atmospheres).
Why is it likely? If Mol1 and Mol2 are not the same then a highly improbable fluke has occurred: whatever it is that comprises a gas has the same properties as those entities arrived at by an Eddington inference, namely, the properties needed to derive the gas laws and Brownian motion. Wright calls this type of inductive inference a ‘no coincidental agreement’ (NCA) inference.
One can flesh out (4) in a more realistic manner by focusing on the details of the case study. For example, by considering what would happen if a gas contained in a vessel was composed of Mol1 and the rapidity of motion of the particles was doubled, Maxwell was able to derive that Mol1 have the following property: the pressure they exert on the walls of the vessel will be proportional to the square of the velocity. As it turns out, Mol2 of the kinetic-molecular theory of heat share this exact same property. Similarly, one can use Eddington inferences to conclude that there are ‘molecules’ that obey Laplace’s law of atmospheres, and that there are also other ‘molecules’ that give rise to Brownian motion, and a NCA inference (of the kind in (5)) assures us that it is likely that both these molecules are one and the same.
There is a worry regarding (1) and (3). In particular, the underdetermination of theory by actual data will imply that these data can be accounted for by different theories, which in turn will postulate distinct unobservable entities (in relation to (3)) and will identify different possible patterns or regularities found in the data and attributed to observable objects (in relation to (1)). Wright () claims to deal with the underdetermination problem in his book Explaining Science’s Success, and Chapter 6 of the current book is dedicated to outlining that solution. The gist is that in cases of underdetermination, we always prefer theories that are ‘independent from data’, where to be independent from data (to some degree) is to minimize the likelihood that it is merely due to chance that the data accord with theory. In other words, it is more likely that the theory ‘has got on to some genuine tendency […] that exists in nature’ (, p. 152).
As an aspiring realist, I look favourably upon Wright’s admirable attempt to ground realism in purely inductive terms and that neither makes use of IBE (which I take to be suspect as a sui generis universal form of inference) nor is based on some idiosyncratic account of scientific explanation or simplicity (wherein such notions open up a Pandora’s box of other philosophical problems). My main worry with Wright’s project is that he begins his book by taking on the monumental task of solving Hume’s problem of induction and he does this with what I take to be an impoverished account of induction, namely, enumerative induction. In doing so, he bases his defence of scientific realism on claims and principles—such as the principle of indifference and the claim that there are synthetic a priori reasonable beliefs—that many would find more dubious than IBE itself (even with the associated controversial notions of explanation, simplicity, and so on).
For example, consider his justification of inductive inference in Chapter 2, which he takes to be the foundation for his notions of Eddington inferences, NCA inferences, and independence from data. Hume’s problem of induction may be summarized as a kind of trilemma: Inductive (that is, ampliative) inference can be justified inductively, deductively, or by some other method (say, abductively). Inductive justification would be circular. Deductive justification would violate the inductive-ampliative character of the inference in question. Some other justificatory method would itself be in need of justification, thereby engendering an infinite regress or else succumbing to one of the other two horns of the trilemma. Is there an escape? Wright claims that there is: induction can be justified if we have good reasons to believe that some induction will work and we can have a priori reasons to accept some synthetic propositions that power some inductive inferences. Here is the argument in three steps: First, observations or collected data are ‘blindly chosen’ if, prior to choosing to observe and record some data, we have no reason to believe that any particular outcome (associated with our observations or data) is more likely than another. For instance, if I flip a coin 1000 times and you randomly choose to observe a particular set of 100 tosses that all come out heads, then your observation is blindly chosen. Second, we now have a priori reasons to believe the following (alleged) analytic proposition:
AH: If, hypothetically, a coin comes up heads in 100 blindly chosen tosses, then there is good reason to believe that the coin is not fair.
Third, if we have a priori reasons to believe AA, then we have a priori reasons to believe the following synthetic proposition:
AS: If, hypothetically, a coin comes up heads in 100 blindly chosen tosses, then the coin is not fair.
This type of three-step justification of the possibility of inductive inference, when conjoined with the principle of indifference, is then used by Wright to argue for the cogency of enumerative induction. For instance, say I blindly chose to observe crows in Geelong and discover that all crows there are black. It is then likely that all crows everywhere are black since if this were not the case, ‘a highly improbable event would have occurred: the blindly chosen location of our observation would have happened to have coincided with the region of black crows in a sea of non-black crows’ (p. 121).
Needless to say, these claims are highly controversial. For one, it seems to me that whether or not we have good reasons to believe in AA, and hence in AS, will depend on facts about how the world is.2 Consider another example3: A scientist presents us with a glass box, which we observe to be filled with gas in its equilibrium macro-state. Can we say we have a priori reasons to believe that the gas will remain in its equilibrium state if undisturbed? Well, given the way our world is—given certain a posteriori knowledge we have—we have good reason to believe that it is overwhelmingly unlikely that the gas is in one of a minority of micro-states that take it from an equilibrium to a low-entropy state. In short, we have good reasons to believe that the gas will remain in equilibrium. Nonetheless, imagine that we lived in a world where evil scientists regularly set up gasses in boxes in equilibrium states that are badly behaved so that they do not evolve in accordance with the second law of thermodynamics. In that world, it is likely that the gas will not remain in equilibrium and this likeliness claim is grounded by facts that we know about that world. This is so even if the gas-in-a-box contraption was blindly chosen. That is to say, likeliness claims are ultimately grounded in matters of fact pertaining to a specific situation in a particular world and such facts are known a posteriori. Thus, it is worrying that Wright takes his entire project to be based on the plausibility of having a priori reasonable beliefs about synthetic propositions. Many with empiricist intuitions like myself would want to get off the train, so to speak, at this early point of the argument (and that’s while setting aside further worries about embracing the highly contentious principle of indifference). Still, let us move on to Chapter 5 where the notions of Eddington and NCA inferences arise.
Suppose that you are an ichthyologist studying the size of fish in the sea. You catch fish with a net that has holes allowing fish less than four-inches long to slip through. You find that there are fish of various sizes in your catch—ten inches, nine inches, and so forth—all the way down to four-inches long. Now you make the following Eddington inference: it is likely that there are fish in the sea of a size less than four-inches long because if there weren’t, then a highly improbable event would have occurred, namely, the blindly chosen size of holes in the net would have happened to have coincided with the size of the smallest fish in the sea. So, we have an inductive argument for the existence of unobserved small fish that is powered by the fact that the size of the holes in the net were blindly chosen. Such Eddington inferences allow us also to make inferences about unobservable objects like molecules (as is clear in the example above (1)–(5)) that are purely inductive—no appeal to IBE is made and they can be applied universally independent of the local details of a situation. Let us consider some more examples that will make clear the problem with this last point.4
Say we blindly choose some samples of the element bismuth and observe that they melt at 271°C; we conclude that it is likely that all samples of the element bismuth melt at this temperature. According to Wright, what powers the inductive inference is that our samples were ‘blindly chosen’. However, I submit this claim gets things wrong. What powers the inference in this case is a background fact regarding bismuth, which is pertinent to this particular matter of inductive inference. Specifically, samples of bismuth are generally uniform in just that property that determines their melting, namely, their elemental nature. It is this fact that powers the inductive inference and makes it likely that, indeed, all samples of the element bismuth melt at 271°C. That this is so is clear if we consider a different (but formally equivalent) situation: some samples of blindly chosen wax melt at 91°C, therefore it is likely that all samples of wax melt at 91°C. In this case, there is no background fact that can power the inductive inference since ‘wax’ is a generic name for various mixtures of hydrocarbons (Norton , p. 650). In other words, although the samples were blindly chosen, it isn’t likely that all samples of wax melt at 91°C. This shows us that something has gone wrong with Wright’s analysis. In other words, it is worrying that Wright’s account of inductive inference—via the notions of a blindly chosen observation, Eddington inference, and so on—seems to commit him to the claim that there is a single, universal logic of induction, which can be applied in all situations regardless of context. Oddly, Wright (pp. 129–30) seems to embrace this conclusion and its consequences, arguing that Eddington inferences give us (defeasible) reasons to believe that there exist fish down to any arbitrarily small size (a claim that I would interpret as a reductio).
Taking all these reservations into account, it would seem that, strategically, it might have been better for Wright to take our ability to make justified inductive inferences for granted, leaving the task of tackling Hume’s problem of induction for another occasion. After all, it appears reasonable to be optimistic about a solution to Hume’s problem given that induction both works and is the basis of science. Instead, here he might have just focused on adopting a more sophisticated account. My own favoured account is John Norton’s (, [unpublished]) material theory of induction.5 Within this context, we can agree with Wright that notions like Eddington inferences, NCA inferences, and independence from data—along with the agreement of independent methods and the overdetermination of physical constants—are necessary in order to make the case for scientific realism without IBE. But such inference schemas cannot be applied universally in order to warrant inductive inference, and it is not the case that judgements of likelihood are justified a priori. Instead, such inferences are powered by background facts, known a posteriori, and pertinent to the situation at hand. Whether or not, for example, an Eddington inference about the size of fish found in the sea is warranted will depend on facts about fish, the sea, the construction of fishing nets, and so on. Coupled with the material theory, then, Wright’s approach does seem to give us a route to scientific realism without IBE. And if this is the case, his account and the questions that arise strongly merit the attention and scrutiny of philosophers of science. Can empiricists block the inductive inference to the unobservable supplied by Eddington inferences? Are such inferences just appeals to IBE in disguise? And what type of realism is supported by Wright’s approach, global or local? And so on (see Shech [unpublished]).
Thus, though I would challenge some aspects of Wright’s argumentative strategy, his engaging book is rich with philosophical ideas and detailed case studies from the history of science, which make significant contributions to various facets of the realism/anti-realism debate. It is clearly written and well organized, impressive in its scope, and innovative in its novel and sophisticated defence of scientific realism.
Department of Philosophy
Auburn, AL, USA
Eddington, A. : The Philosophy of Physical Science, Cambridge: Cambridge University Press.
Norton, J. : ‘A Material Theory of Induction’, Philosophy of Science, 70 pp. 647–70.
Norton, J. : ‘A Material Dissolution of the Problem of Induction’, Synthese, 191, pp. 671–90.
 Admittedly, Wright (pp. 224–5) considers such an objection but I find his reply unsatisfactory. He claims that we can have a priori reasons to think that some possible worlds are rare while others are not, and suggests that we can make sense of such modal notions in a theory-independent manner.
 For details of this example, see (Shech ). It would seem that if Wright’s proposal regarding a priori synthetic reasonable belief worked (which I don’t think it does), it would have direct implications for foundational issues in statistical mechanics like explaining the second law and justifying the probability measure in Boltzmannian statistical mechanics.
 The following example is from (Norton ).
 That the material theory is able to deal with the problem of induction is suggested in (Norton ) and that it can deal with various historical inductions for/against realism, such as the pessimistic meta-induction and problem of unconceived alternatives, is outlined in (Shech [forthcoming]).