Empirical Evidence Claims Are A Priori

And consider how this bears on the following hypothesis:

h: Mary will not become pregnant during a period in which she takes a pill of type P every day.

Now imagine we learn something like:

e1: The 1000 patients referred to in e were all infertile before taking P.

Achinstein (1995, p. 450) concludes that such facts:

should make us withdraw, or at least modify, our original claim about the evidential efficacy [of e with respect to h]

In part, it is possible to deal with this problem by declaring that P(h, e) = 0.5, say on the basis of objective Bayesian equivocation norms, because e neither confirms nor disconfirms h. But this leaves the worry that it is difficult to see what kind of evidence could confirm h.

Enter the move that Achinstein does not anticipate. Consider P(h, eb), where b is ‘The study was not flawed’. Clearly, the advocate of the a priori theory will say, this conditional probability is positive, and indeed equal to unity. After all, h is entailed by e and b.

Now if we allow scientists to use claims such as b as working hypotheses, the a priori thesis is saved. To illustrate this, let us represent e and b as E. All that matters for determining whether E is evidence for h is the truth of E. So all that Achinstein (1995) shows, it may be said, is that we often have great difficulty in determining the truth of the evidence by empirical means.

This is just a cursory sketch of the main argument in the paper. It also considers Achinstein’s argument for his empirical view of evidence from the history of science, based on the view that we should take what scientists say literally, and shows how this fails. It concludes by suggesting that the a priori view lies at the core of progressive research programmes in artificial intelligence and computational social epistemology.