Climate scientists often apply statistical tools to a set of different estimates generated by an “ensemble” of models. In this paper, I argue that the resulting inferences are justified in the same way as any other statistical inference: what must be demonstrated is that the statistical model that licenses the inferences accurately represents the probabilistic relationship between data and target. This view of statistical practice is appropriately termed “model-based,” and I examine the use of statistics in climate fingerprinting to show how the difficulties that climate scientists encounter in applying statistics to ensemble-generated data are the practical difficulties of normal statistical practice. The upshot is that whether the application of statistics to ensemble-generated data yields trustworthy results should be expected to vary from case to case.

A number of philosophers of science have argued that there are important differences between robustness in modeling and experimental contexts, and—in particular—many of them have claimed that the former is non-confirmatory. In this paper, I argue for the opposite conclusion: robust hypotheses are confirmed under conditions that do not depend on the differences between and models and experiments—that is, the degree to which the robust hypothesis is confirmed depends on precisely the same factors in both situations. The positive argument turns on the fact that confirmation theory doesn’t recognize a difference between different sources of evidence. Most of the paper is devoted to rebutting various objections designed to show that it should. I end by explaining why philosophers of science have (often) gone wrong on this point.

Pierre Duhem's influential argument for holism relies on a view of the role that background theory plays in testing: according to this still common account of "auxiliary hypotheses," elements of background theory serve as truth-apt premises in arguments for or against a hypothesis. I argue that this view is mistaken. Rather than serving as truth-apt premises in arguments, auxiliary hypotheses are employed as (reliability-apt) "epistemic tools": instruments that perform specific tasks in connecting our theoretical questions with the world but that are not (or not usually) premises in arguments. On the resulting picture, the acceptability of an auxiliary hypothesis depends not on its truth but on contextual factors such as the task or purpose it is put to and the other tools employed alongside it. [I]n saying these words, we are doing something ... rather than reporting something. Austin (1962, 13) 0 Introduction Pierre Duhem's incredibly influential The Aim and Structure of Physical Theory (1914/1951) is usually and rightly remembered for its holism. Central to the argument for holism-and perhaps more influential than the holist conclusion itself-is a view of the role that background theory plays in testing. For Duhem, theories or models other than the hypothesis that we aim to test serve primarily as premises in the derivation of consequences. These consequences are then compared to the world; since all the premises have the same logical relationship to the consequence, either all of the premises pass as a conjunction, or they fail as the same. Hence holism. Call these background theories and models "auxiliary hypotheses." The central thesis of this paper is that the Duhemian view of the role that auxiliary hypotheses play in testing is mistaken. The theories, models, and analogies employed in testing a hypothesis rarely serve as premises. Instead, they serve as "epistemic tools": instruments that perform specific tasks in connecting our theoretical questions with the world but that are not (or not usually) premises in arguments. For the purposes of testing, therefore, it's largely irrelevant whether these theories or models represent truthfully. What we care about is whether they reliably perform their task.

William Whewell's account of induction differs dramatically from the one familiar from twentieth-century debates. I argue that Whewell's induction can be usefully understood by comparing the difference between his views and more standard accounts to contemporary debates between semantic and syntactic views of theories: rather than understanding inductive inference as capturing a relationship between sentences or propositions, Whewell understands it as a method for constructing a model of the world. The difference between this ("semantic") view and the more familiar ("syntactic") picture of induction is reflected in other aspects of Whewell's philosophy of science, particularly his treatment of consilience and the order of discovery. What is meant by success in these cases? To this we reply, that our inquiry must be, whether the facts have the same relation in the hypothesis which they have in reality. (William Whewell 1840/2014b, 210

I argue that the appropriateness of an assertion is sensitive to context—or, really, the “common ground”—in a way that hasn’t previously been emphasized by philosophers. This kind of context-sensitivity explains why some scientific conclusions seem to be appropriately asserted even though they are not known, believed, or justified on the available evidence. I then consider other recent attempts to account for this phenomenon and argue that if they are to be successful, they need to recognize the kind of context-sensitivity that I argue for.