The Structure of Empirical Knowledge: Theories of Justification

So far we’ve discussed how to characterise knowledge. We now come to a different issue: the structure of what we know, or at least believe. What we have in mind are the relations between the different components of our belief systems, the way we argue or infer from one to another. Some beliefs, after all, provide support for others. Questions about the structure of our belief systems are important because they are also questions about justification. We’ve seen already how the concept of justification is used in the JTB analysis of knowledge. Even if it doesn’t appear in the correct philosophical analysis of knowledge, justified belief is still a phenomenon epistemology has to deal with.

The Regress Argument

Much discussion of the structure of epistemic justification starts by considering the regress argument. This is by far the best-known argument for the doctrine known as ‘foundationalism’. It starts with the idea that some of one’s beliefs are justified by reference to others. For example, if I already believe that Socrates is a person, and that all people are mortal, I might infer from these the new belief that Socrates is mortal. If my two existing beliefs were themselves justified, then via this inference I’ve acquired a new, different, justified belief. Let’s call this an inferentially justified belief. The regress argument simply claims that not all of my beliefs can be inferentially justified beliefs. As well as beliefs which are justified in this way there must be others, which are justified non-inferentially.

The regress argument proceeds by reductio ad absurdum. Start by assuming the negation of the conclusion, i.e. that all justification is inferential. Inference is a matter of moving along a path from premises to conclusion. Conclusions can only be justified if they follow from justified premises. If my conclusion R is justified, that must be because it follows from some premise or set of premises. Let’s call that (possibly complex) premise Q. If R is to be justified, then premise Q must itself be justified. It’s obviously not enough, to be justified in believing R, to infer R from any old set of premises. But if my belief that Q is justified, that must be because it follows from some other premise or set of premises. Let’s call that premise or set of premises P. Then the same issue arises again over the justification of P. The possibilities for inferential justification here are just two: either there’s no end to the chain of justification, or the chain loops back on itself, P justifying Q, Q justifying R, and R justifying P.

Neither of these possibilities looks attractive. If the first holds, we have an infinite regress of justification. Each belief is justified by virtue of its connection to some other belief, and so on ad infinitum. This is supposed to be unsatisfying because it offers no prospect of anything being unconditionally justified. Inferential justification also requires us to actually have an infinity of actual beliefs, which might be thought problematic.

If the second possibility holds, we have a circularity. We are, in effect, claiming that our justification for believing that p stems, ultimately, from p itself. This is unsatisfying because circular arguments are just too easy to construct. What’s more, it doesn’t seem successful anyway. As long as a particular belief is in need of justification, it cannot provide justification for other beliefs, especially those beliefs needed to support the original belief itself". "The loop", as Jonathan Dancy puts it, "will never succeed in removing the conditionality" (Introduction to Contemporary Epistemology, p.56). It doesn’t seem to matter how wide the circle is. One can justify any belief if one’s allowed to justify it by reference to itself.

The regress, then, is a vicious infinite regress: if all justification is inferential, no belief can actually be justified at all (to any extent). This completes our reductio ad absurdum. Its sceptical conclusion that no belief can ever have any positive justification is absurd. Therefore, foundationalists reason, we should reject the premise that led us to it, the idea that all justification is inferential. We’re forced to accept that one’s premises can be justified in some other way, some non-inferential way.


This means that some of one’s beliefs must be basic, that is, non-inferentially justified. This is exactly the conclusion embraced by the foundationalist: there are two forms of justification, inferential and non-inferential.

Foundationalists hold that some parts of the structure which we call human knowledge, and of the structures that each of us counts as our own belief system, are more basic than others. This relationship of basic and derived beliefs is not one of temporal precedent: it’s not that some things have to be learned before others. Rather, as philosophers, we’re interested in epistemic relationships between different parts of our belief system, relationships of justification. Some of our beliefs (the ‘Basic Beliefs’) are more basic than others, says the foundationalist, in the sense that they cannot be justified by reference to other beliefs (call them the ‘Derived Beliefs’) whereas derived beliefs can be justified by reference to basic beliefs. It’s characteristic of a foundationalist view, then, at least that:

It posits the existence of at least two distinct classes of beliefs.

Beliefs in one of these classes (the basic class) are privileged in that they form at least part of the grounds or justification for holding a belief from the other class (the non-basic class).

But basic beliefs are not justified by reference to beliefs of any other class. Instead, they’re foundational (in that they are either self-justifying or unjustified justifiers, or justified by reference to something other than beliefs).

To give an account of any structure which is deemed to have foundations, one is obliged to explain at least two things:

[1] what the foundations of the structure are and why they count as foundational,


[2] what relationship the foundations bear to the superstructure.

These demands are reflected in epistemological foundationalism.

Classical Foundationalism

Foundationalist theories are one type of account of the structure of knowledge and the notion of justification. They were perhaps the only such account seriously considered until the nineteenth century. The picture of knowledge as a structure with foundations was a popular metaphor going back through Descartes, to the Greek philosophers. Foundationalism is associated, strongly, with empiricism, but this doesn’t rule out the possibility of non-empiricist foundationalisms (e.g. Descartes). Nevertheless the foundationalisms we’ll consider are empiricist: they add to the structural distinction between basic and non-basic beliefs the idea that the basic beliefs have a special kind of content, experiential content. Basic beliefs, it’s said, are about our own ‘immediate experience’ or ‘sensory states’. Empiricists think that they’re going to be perceptual beliefs, beliefs about one’s immediate perceptual experience. The empiricist idea is that knowledge has a specific content, experiential content, and that it does so in virtue of having a special origin, in perception.

The constraints we’ve mentioned so far still allow a wide latitude for variation among foundationalist doctrines. The kind we’ll examine has been historically the most important. This is ‘classical’, or ‘radical’ foundationalism. According to classical foundationalism, Dancy says,

epistemology is... a research programme which sets out to show how our beliefs about an external world, about science, about a past and a future, about other minds, etc., can be justified on a basis of infallible beliefs about our own sensory states (ibid., p.54).

The classical foundationalist takes the most uncompromising foundationalist position. He or she thinks the foundations of knowledge must be absolutely secure, in fact that they must be sceptic-proof. Thus he or she claims that basic beliefs have some especially valuable epistemic property: they are infallible or incorrigible or indubitable or certain. Any infallible belief would have to be non-inferentially justified. Non-inferentially justified, infallible beliefs will stop the threat of infinite regress. (Having said this, we must point out that ‘classical foundationalism’ is more like an ideal type than an historically-existing entity. The truth is that forms of foundationalism approach the ideal of classical foundationalism to varying degrees). Classical foundationalism says that beliefs about immediate experience are non-inferentially justified because they’re infallible.

Objections to Classical Foundationalism, I: Problems with Basic Beliefs

We said that every foundationalist is obliged to give an account of at least two things: the basic beliefs, and the relationships between the basic beliefs and the derived beliefs. Let’s see how the classical foundationalist fares against objections.

The problems start when we consider the supposedly infallible beliefs demanded by Classical Foundationalism. Why should the foundationalist require infallibility anyway? An historically important reason has been that if beliefs about our own sensory states were infallible, we would have a guarantee that they were always true, and thus a sure-fire reply to some kinds of scepticism. But then we must ask exactly what form of infallibility is at issue, what does being ‘infallible’ mean? Some suggestions that have been made are:

Certainty. (p has a probability of 1 relative to S’s evidence).

Incorrigibility. (If S believes p, p is true, or p does not go beyond what S is directly aware of, or p can’t be falsified by subsequent experience, it has no predictive content).

Indubitability. (If S thinks of p, S knows p). (For these categories, see B.A.O.Williams, Descartes: The Project of Pure Enquiry, (Harmondsworth: Penguin, 1978), p.306, and W.P.Alston, ‘Varieties of Privileged Access’ in his Epistemic Justification: Essays in the Theory of Knowledge, (Ithaca: Cornell University Press, 1989)).

Are our beliefs about our own sensory states infallible in any of these ways? And would it be a vindication of the classical foundationalist doctrine if they were? Let’s take the second question first.

The Problem of Empirical Content

The trouble with most suggested kinds of basic beliefs is that they fall between two stools. One the one hand we have the requirement that they be infallible, and other the other hand we must require that they have adequate empirical content. These requirements seem to be in tension. Empirical beliefs, e.g. beliefs about medium-sized physical objects, are paradigm cases of beliefs with substantial empirical content. But paradigm cases of beliefs which are infallible, such a mathematical beliefs, simply don’t have much (if any) empirical content. (One criterion of empirical content is how much one can derive from a belief, how predictively fruitful it is, or how many possible states of affairs it rules out). Empirical content is inversely proportional to infallibility! To aim at the one is necessarily to fall short of the other. To say something substantial (informative) is necessarily to stick one’s epistemic neck out. One doesn’t have to hold either that we can give examples of beliefs which are ‘pure cases’ of each type (the analytic/synthetic distinction), or that each belief has its own neatly identifiable empirical content (semantic atomism), in order to use the above objection.

Are Basic Beliefs Infallible?

Beliefs about sensory experience, then, are either too weak to support inferences to other, richer beliefs, or they are such that under certain circumstances we can doubt them: they are not infallible. Those who contend that these beliefs are infallible of course allow that we may make a mistake in describing our sensory states, but they dismiss this as merely verbal error. They argue that being mistaken about the meanings of the words I use to describe my experience doesn’t show that I have any mistaken beliefs about that experience, and that verbal mistakes can be corrected in standard ways. Let’s ask exactly what kinds of mental phenomena one could have infallible beliefs about.

Could a person be mistaken about their beliefs about their sensations? Here’s an argument to the conclusion that mistakes about sensations are possible, due to David Armstrong. Consider what would happen if science advanced to the stage where we could establish via an instrument, call it a ‘cerebroscope’, that a person has a certain sensation, say a visual sensation of red, when and only when he is in a certain brain state, say state 1279. Suppose this has been confirmed so far as a reliable correlation. And then say we give this person a drug which will have the effect of making him tell the truth, in a laboratory experiment. Then suppose that he reports a sensation of red, but that the cerebroscope reports some brain state other than 1279. On the basis of the evidence that he is drugged we should conclude that he believes what he says, he is not lying. But on the basis of the cerebroscope report we should conclude that he is not having a sensation of red. Surely, Armstrong argues, we might feel free to accept the verdict of the cerebroscope over that of the experimental subject, especially if the correlation was reliably established? It must then be logically possible that the person believes that he has a sensation of red when in fact he does not.

Unfortunately, the argument fails. The defender of infallibility is able to reply that the experimenter, instead of being convinced that the cerebroscope has revealed that the sensation is not taking place, should have concluded that the experiment reveals that the cerebroscope has itself failed to reveal a genuinely law-like correlation. This is because the experimenter fails to realise that the fact that the man is not having the sensation would logically imply that he does not believe he has the sensation. "It would only be unreasonable for the scientist to continue to insist that he was [not] appeared to redly if such beliefs were not incorrigibly justified, and that is the very question at issue" (J.Pollock, Contemporary Theories of Knowledge, (Totowa, NJ: Rowman & Littlefield, 1986), p.34).

Armstrong’s cerebroscopist therefore begs the question. But maybe it’s possible, for example, to believe that one is having a particular sensation when one is in fact having a different one (and yet where the error is not merely verbal). Perhaps the most plausible example of this is the case of a person being tortured, who is unexpectedly given ice cubes on his back instead of electric shocks. He might say "For the first few seconds, I thought it was a pain, but I was wrong, it was a cool sensation".

Could a person have mistaken beliefs about their own beliefs? This rather depends on what we think beliefs are. As long as we’re persuaded by a simple dispositional model of beliefs, it would seem that one certainly can be mistaken, since one may not be a reliable indicator of one’s own dispositions. But could you have a mistaken belief about your occurrent thoughts, the things that ‘run through one’s head’? And if not, would that rescue the classical foundationalist?

Objections to Classical Foundationalism, II: The Relationship Between Basic and Derived Beliefs

The second major task of the foundationalist is to explain how the basic beliefs justify the derived beliefs. We’ve already suggested that beliefs about sensory experience might well be too weak, too uninformative, to support claims about physical objects. But this needs investigating more thoroughly, since several possible accounts of the supporting relationship have been suggested.

The strongest such relationship would be logical entailment. When a proposition p logically entails a proposition q, p can’t be true without q being true. To get from an object’s looking red to its being red we would have to find some further premise which, when conjoined with the fact that one is ‘being appeared to redly’ (as epistemologists sometimes put it), would logically entail the conclusion that there is a red object present. So the proposal is to find entailments of the form:

Being appeared to redly under circumstances of type C entails that there’s a red object present.

But our specification of the ‘Circumstances C’ mustn’t involve reference to the states of material objects, otherwise we will be presupposing perceptual knowledge of material objects in order to explain how perceptual knowledge of material objects is possible. So ‘Circumstances C’ can only involve reference to appearances. The phenomenalist proposal is that material object statements can be analysed as statements about appearances. C.I.Lewis, for example, proposed that ‘x is red’ can be analysed as a conjunction of a (possibly infinite) number of conditionals of the form "if I were to do A in circumstances C then I would be appeared to redly".

What are the problems with this proposal? A major one is that of specifying the circumstances C without any reference to material objects. It does seem that no purely phenomenal circumstances will suffice. Appropriate conditions will ensure that if x is red and I look at it then it will appear red to me. But no matter what we propose for C "we can always imagine elaborate conditions under which the putative entailment will be falsified" (Pollock, ibid., p.41).

If the relationship between basic and derived beliefs isn’t deductive then, what else could it be? What about inductive relationships? Surely a basic belief to the effect that I am appeared to redly might count as very good inductive evidence for the supposition that there is red object in front of me? The problem here, though, is that we could only legitimate such discoveries if we had independent access to the colours of objects and the colours they look to us, and can compare them. (This is a problem for the representative theory of perception).

Another stab would be to say that the inference to material objects from sensations is what philosophers of science call an inference to the best explanation. "Given a set of observations, we often take a hypothesis to be confirmed because it is the best explanation of those observations" (Pollock ibid., p.42). Under ordinary circumstances, surely the best explanation for the fact that I am appeared to redly is that there is a red object in front of me? And the best explanation for why something looks red is that it is red! Unfortunately for this proposal, an object’s being red "only explains its looking the way it does to me if I already know that that is the way red things ordinarily look to me, but of course, I cannot know that without first acquiring some perceptual knowledge" (ibid., p.43).

So explaining the relationship between basic and derived beliefs is a thorny problem for the foundationalist. Its not that anyone thinks there is no relationship, but the classical foundationalist needs there to be a relationship which will support an infallible and entirely one-way relation of justification. If the principles of inference which take us from basic to non-basic beliefs are themselves fallible, why should the foundationalist bother to insist that the basic beliefs which act as input should be infallible?