Scientific realism begins with the assumption that induction alone is suspect, and that the ability of a hypothesis to explain empirical facts strengthens it. This is called abductive inference. Realists can go further to suggest that if a particular hypothesis, H, explains a set of facts, F, better than any competing hypothesis then this could warrant belief in the resulting theory.

Objections to this idea are numerous and varied. The most obvious and pressing objection, proposed by Mill when he was arguing against wave theories of light in the nineteenth century, is as follows: abductive inference is used where the subject matter isn't experientially accessible, e.g. to predict that a particular phenomenon is causing various effects even though that phenomenon cannot be observed. As such, no matter how important and useful the theory, and no matter how bad the rival theories, it cannot be accepted as a basis for belief. Hypotheses can only be of instrumental use for scientists in trying to predict outcomes and construct models.

Instrumentalism therefore states that all theoretical language used in abductive inference is itself instrumental, bearing no descriptive or referential properties (to observations). Where observation statements are insufficient to support a theory, instrumental propositions can be used, and judged according to their efficacy, however that is to be decided.

Objections to abductive inference

1. Quine - Any given body of evidence can have more than one instrumental explanation (Thesis of Empirical Equivalence). Therefore no body of evidence supports one theory to the exclusion of all others, and there are no epistemic grounds on which one can decide between equivalent theories. The only move open to realists is to choose between theories on a purely pragmatic basis, which becomes the domain of scientists not philosophers. In other words, abductive inference ceases to have any epistemic import, and becomes a mere methodology based on pragma.

2. van Fraasen - If we rely on pragma as a basis for analysis of explanation then we confront the problem of context. Our hypothesis, H, may have explanatory power in one context and none in another, e.g. because of factors unaccounted for in the hypothesis. We have no objective perspective from which to judge when a hypothesis is successful in this respect, and so H can only be considered instrumentally useful given a specific context, including our interests and the direction of its arguments. Worse still, a change in context may reverse the status of explanation and outcome, such that we cannot say, without context, what has explanatory power. The result of this is that, decontextualised, explanation has no epistemic import, making explanatory status itself pragmatic.

3. Laudan - Any explanatory virtues may have failed in the past, or may support theories that are incompatible with current science. Thus, though a theory with greater explanatory scope and power than any other may be instrumentally useful, this success cannot, for historical reasons, warrant belief. Only observations can be believed. Again, this extreme form of instrumentalism is fine for scientists, but still gives us no reason to believe the outcome of science except where theories are based wholly on observations.

4. Fine - There are two levels of reasoning involved here. On the first level we have scientific reasoning, on which theories are proposed and evaluated in light of experimental evidence and, according to realism, abductive inference. On the second level we have philosophical reasoning, where general theories about the epistemic status of the conclusions reached on the first level are evaluated. If philosophical theories purport to judge the epistemic import of modes of inference used in science then it cannot use those modes itself, and the methods of philosophy must be deemed stronger. However, abductive inference violates this rule by legitimising itself according to it's own criteria -- it states that it has epistemic import because it is successful. It is therefore a circular argument.

One might reply that this shouldn't concern us. After all, any epistemic theory will itself need to be justified, and we could adopt a stance similar to coherentism or reliabilism and say that the success of any given theory has epistemic import.