ABSTRACT

Assume that, when the observation is made, E turns out to be false. All that logic guarantees is that at least one of the premises is false – it does not dictate which one and in particular it does not dictate that it is the central theory. Those scientists whom kuhn describes as treating recalcitrant data as “anomalies” are just taking it that, at least as a first move, the “blame” for getting the data wrong lies either with an auxiliary theory or with one of the specific assumptions rather than with any theory basic to the paradigm. There are many cases in the history of science showing that this type of move, far from being under suspicion of possible “irrationality”, has produced some of the greatest scientific breakthroughs. Perhaps the most famous was the discovery of Neptune: by holding on to Newton’s theory despite its apparent clash with the facts about Uranus’s orbit, Adams and Leverrier were led successfully to predict the existence of a hitherto unknown planet. Treating a negative result as an anomaly is, therefore, sometimes good science. But in other cases it seems to be the very essence of pseudoscience. Consider, for example,

creation “scientists” defending their basic theory that god created the universe in 4004 BC against the evidence of the fossil record by assuming, as Gosse famously did, that god created the rocks with the fossils already in them. And even within science, such defenses of an entrenched theory often seem to be clearly bad science. When, for example, the wave theory of light made impressive predictions about the results of various diffraction experiments, some corpuscularists, just as kuhn would suggest, “held out” for their preferred theory and claimed that these results were merely “anomalies” for their theory: eventually, by making the right (and clearly quite complex) assumptions about the “diffracting forces” that affect the particles as they pass the edges of opaque objects, these results could be given a corpuscularist account. Duhem’s analysis shows that such a move is always logically possible. However, although corpuscularists might produce tailor-made assumptions about diffracting forces to accommodate, say, the outcome of the two-slit experiment, the strong intuition remains that this is a telling result in favor of the wave theory. If we are to show that theory-change in science has been rational in the precise sense that later theories are invariably better empirically supported than their predecessors, then we shall need an account of empirical support that underwrites this intuition. An obvious distinguishing feature in these cases is that the newer theory standardly predicts the empirical results, while the defenders of the older theory accommodate those results after the fact. So Fresnel’s theory predicted the white spot at the centre of the geometrical shadow of a small opaque disc; corpuscularists suggested after the event that this result might be accounted for within their approach by making suitable assumptions about “diffracting forces”. Darwinian theory predicts (in a way) the fossil record; creationists only accommodate the facts after the event by supposing that god chose to draw pretty pictures in some rocks when creating them. If then there were a general defensible rule of empirical support that predictions count more then we would have the rationale we are seeking. The issue of prediction vs. accommodation is a long-running one that continues to be hotly debated. There seem, however, to be two obvious problems with the suggestion that predictions carry more supportive weight than explanations of (otherwise equivalent but) already established facts. The first is that while the suggestion yields the intuitively correct judgments in some cases, it does not do so in all. The facts about the precession of Mercury’s perihelion were, for example, well known before the general theory of relativity was articulated, and yet all serious commentators regard that theory’s explanation of Mercury’s orbit as constituting important empirical support for it – at least as strong support as it received from the prediction of any temporally novel fact. The second problem is more general: the suggestion seems to stand without any epistemic justification – why on earth should the time-order of theory and evidence have any epistemological import? It seems then that for all its sharpness, the predictions-count-more view cannot be the correct solution to the Duhem problem. And in fact the main defect of the creationist account of the fossil record, for example, is surely not that the facts were already known when the specific theory that captures them was first formulated, but

rather that they had to be known since they were used in the construction of that specific theory. The basic idea of creationism gives no indication whatsoever that there should be particular “pictures” found in particular rocks – the specific theory that has them as part of creation is based entirely on the observations themselves. Similarly, in the optics case, the basic idea that light consists of material particles subject to forces gives no indication whatsoever that the particular “diffracting forces” emanating from a small disc should be such as to draw the particles passing the edge so that they hit the center of the geometrical shadow: that fact had to be given and to form the starting point of the construction of some force function that would do the job. On the other hand, those cases in which some already-known result seems to supply strong empirical support to a theory are characterized by the fact that the result follows from the central theory concerned, using only natural auxiliaries – not special assumptions that are tailored to the fact concerned. For example, planetary stations and retrogressions fall out naturally from the Copernican theory as straightforward consequences of the fact that we are making observations of the other planets from a moving observatory: a given planet’s stations occur when we overtake or are overtaken by it. The issue is not about prediction versus accommodation, unknown vs. known facts, but rather all about non-ad hoc vs. ad hoc accounts of phenomena whether already known or not (though of course a scientist cannot tailor an assumption to an empirical result she does not yet know about!). This is not to assert that ad hoc maneuvers are automatically scientifically illicit. Adams and Leverrier created a theory specifically so that it would entail the already known (and initially anomalous) details of Uranus’s orbit. Often, indeed, scientists obtain specific theories by deduction from the phenomena – where this really means deduction from the phenomena plus a general theory (or set of such theories) that they already accept. As I argued in “New Evidence for Old” (2002), we need in fact to differentiate two types of empirical support. Deductions from the phenomena supply support for the deduced theory, but only against the already-given background of the general theory: they supply no further support for that general theory. Thus, the creationist theory with the fossils gets (conditional) support from the fossils – they provide a very good reason to hold that particular version of the creation story if you are going to hold any version of that story at all; but the fossils give no (unconditional) support whatsoever to the general story. Similarly, in the Adams and Leverrier case the data from Uranus give very good support to their version of the Newtonian account involving a change in the number of planets presupposed, but the data alone give no unconditional support, I would say, to the general Newtonian theory. The difference in the two cases is, of course, that there is independent evidence in the Newtonian case: the revised theory is read off the Uranian data but then predicts the existence of a new planet, a prediction that can, of course, be checked observationally and which turned out to be true. In the creation case there is patently no such independent testability – writing the fossils into creation simply avoids the initial problem presented by those data but yields no further prediction that can be checked. One important issue is whether the currently most widely held formal account of empirical support – that of the personalist Bayesians – can adequately capture the

intuitive judgments of confirmation. However the merits and demerits of Bayesianism are discussed elsewhere in this collection.