Certainty, Probability and the Correction of Evidence

Abstract

What is certain extends well beyond the a priori, the testimony of our senses and the records of our memory. No one engaged in deciding between alternative courses of action or in determining the correct answer to a given question proceeds without making background assumptions which include universal generalizations, statistical hypotheses and theoretical presuppositions. Nor is this sociological fact an indication of imperfection in the human condition from which an ideally rational man ought to be enjoined to escape. To regard a hypothesis h as certain in the sense I am using, is to be committed to discounting risk of error involved in acting on the truth of h. h is accorded probability 1. h is part of the evidential corpus in the light of which the probabilities attached to nonevidential hypotheses are appraised. Many writers of Bayesian persuasion find the view I am advocating philosophically insupportable. Roughly speaking, they insist that, with the possible exception of observation reports of certain kinds, no empirical judgments can be counted as certain; for the truth of universal generalizations, statistical hypotheses and theoretical assumptions cannot be infallibly secured. I yield to no man in my devotion to the thesis that human knowledge is fallible. Indeed, unlike some Bayesians, I regard observation reports as fallible and hold the same to be true of logical