The emergence of logical positivism led to the loss of precious, hard-won, and deep insights of Kant into the nature of human knowledge. Because of the complexity of Kant, many interpretations have emerged. Our discussion below follows Gardner (1999). Aspects relevant to the fact/value distinction can be summarized as follows.
Reality generates signals which impact on our physiological equipment for detecting our environment as sensory data (A). This sense data is interpreted (by our mind) to create a model of reality (B). The process of interpretation also involves some prior knowledge represented in (C). According to Kant, a central concern of traditional metaphysics was the correspondence between our models of reality and reality itself, labeled D in the diagram. The question “Do electrons, charges, gravitational fields, energy exist?” reflects this concern – do these terms in our models of physics correspond to objects in reality out there? Two key insights of Kant which he termed a “Copernican Revolution” in philosophy are
1: [Negative] It is impossible to assess whether our representation of reality is a faithful and accurate representation of reality. This is because we have no independent access to reality, other than by our models of reality. We can and do construct, compare and evaluate different models of reality along many different dimensions. However, we cannot judge these models on the crucial dimension of which is a more accurate representation of reality, because models are all we have.
2: [Positive] One can make progress in epistemology by focusing on (B) and (C), the process by which we transform the chaotic jumble of sense data about the real world into a coherent model of reality.
In accordance with this Kantian insight, “Do electrons exist?” is the wrong question – we can never know whether our models of reality accurately represent what is out there. A more modest question is: “do electrons help in the process of sorting our sense data into a coherent model of reality?” Here the answer is clearly “yes, currently they do.” But a later theory may come along which dispenses with electrons and creates a more “interesting, informative, appealing, or elegant” picture of reality. At which point, electrons will blink out of existence, like ether. This pragmatic approach to ontology became the accepted resolution of debates about the existence of unicorns after Russell parsed them out of existence.
The facts/value distinction is based on the positivist attempt to solve the “impossible” problem of (D). Positivists claim that facts reflect features of the real world. Since we cannot point to any feature of the world out there that is a “value,” values don’t exist. The Kantian answer is that electrons and values are both useful as devices to create a coherent picture of reality. A disturbing implication is that there are no “facts” if we understand facts as sentences which directly describe features of reality out there. These were the observation sentences of the positivists. Attempts to clearly define “facts” along positivist lines – unmediated sense data, providing clear sharp and accurate information about reality – ran into trouble in many different ways, and forced many alternative reformulations before being eventually abandoned. Putnam (2002) has given a detailed discussion of the problems faced by positivists in defining “facts” clearly.
This does not mean that we cannot distinguish between facts and values, but only that we cannot do so along the lines suggested by the positivists: one category is real and objective, while the other is subjective and has no correspondent observables.
Another important insight from Kant is the necessity of prior knowledge in creating a coherent model of reality – the path (C) . This conflicts with the positivist idea that science is purely objective knowledge of the real world, without any underlying subjective or prior judgements. An interesting confirmation of Kant’s idea is furnished by the fundamental theorem of statistical decision theory, according to which all admissible decision rules are approximately Bayes; see Ferguson (1967, Chapter 2) for a clear exposition. This means that all valid statistical inference procedures mix information from data with information from prior judgments to arrive at a decision. There are no purely objective ways of looking at the world.