For this second edition, the editors have returned to Austin's original lecture notes, amending the printed text where it seemed necessary.

For evidence-based practice and policy, randomised controlled trials (RCTs) are the current gold standard. But exactly why? We know that RCTs do not, without a series of strong assumptions, warrant predictions about what happens in practice. But just what are these assumptions? I maintain that, from a philosophical stance, answers to both questions are obscured because we don't attend to what causal claims say. Causal claims entering evidence-based medicine at different points say different things and, I would suggest, failure to (...) attend to these differences makes much current guidance about evidence for medical and social policy misleading. (shrink)

There is a takeover movement fast gaining influence in development economics, a movement that demands that predictions about development outcomes be based on randomized controlled trials. The problem it takes up—of using evidence of efficacy from good studies to predict whether a policy will be effective if we implement it—is a general one, and affects us all. My discussion is the result of a long struggle to develop the right concepts to deal with the problem of warranting effectiveness predictions. Whether (...) I have it right or not, these are questions of vast social importance that philosophers of science can, and should, help answer. (shrink)

This article considers the issue of opacity as a problem for socially consequential mechanisms of classification and ranking, such as spam filters, credit card fraud detection, search engines, news trends, market segmentation and advertising, insurance or loan qualification, and credit scoring. These mechanisms of classification all frequently rely on computational algorithms, and in many cases on machine learning algorithms to do this work. In this article, I draw a distinction between three forms of opacity: opacity as intentional corporate or state (...) secrecy, opacity as technical illiteracy, and an opacity that arises from the characteristics of machine learning algorithms and the scale required to apply them usefully. The analysis in this article gets inside the algorithms themselves. I cite existing literatures in computer science, known industry practices, and do some testing and manipulation of code as a form of lightweight code audit. I argue that recognizing the distinct forms of opacity that may be coming into play in a given application is a key to determining which of a variety of technical and non-technical solutions could help to prevent harm. (shrink)

Advanced AI systems are rapidly making their way into medical research and practice, and, arguably, it is only a matter of time before they will surpass human practitioners in terms of accuracy, reliability, and knowledge. If this is true, practitioners will have a prima facie epistemic and professional obligation to align their medical verdicts with those of advanced AI systems. However, in light of their complexity, these AI systems will often function as black boxes: the details of their contents, calculations, (...) and procedures cannot be meaningfully understood by human practitioners. When AI systems reach this level of complexity, we can also speak of black-box medicine. In this paper, we want to argue that black-box medicine conflicts with core ideals of patient-centered medicine. In particular, we claim, black-box medicine is not conducive for supporting informed decision-making based on shared information, shared deliberation, and shared mind between practitioner and patient. (shrink)

Understanding is a central aim of science and highly important in present-day society. But what precisely is scientific understanding and how can it be achieved? This book answers these questions, through philosophical analysis and historical case studies, and presents a philosophical theory of scientific understanding that highlights its contextual nature.

This is the introduction to the Routledge Companion to Thought Experiments.

Many of the computing systems programmed using Machine Learning are opaque: it is difficult to know why they do what they do or how they work. Explainable Artificial Intelligence aims to develop analytic techniques that render opaque computing systems transparent, but lacks a normative framework with which to evaluate these techniques’ explanatory successes. The aim of the present discussion is to develop such a framework, paying particular attention to different stakeholders’ distinct explanatory requirements. Building on an analysis of “opacity” from (...) philosophy of science, this framework is modeled after accounts of explanation in cognitive science. The framework distinguishes between the explanation-seeking questions that are likely to be asked by different stakeholders, and specifies the general ways in which these questions should be answered so as to allow these stakeholders to perform their roles in the Machine Learning ecosystem. By applying the normative framework to recently developed techniques such as input heatmapping, feature-detector visualization, and diagnostic classification, it is possible to determine whether and to what extent techniques from Explainable Artificial Intelligence can be used to render opaque computing systems transparent and, thus, whether they can be used to solve the Black Box Problem. (shrink)

Claims pertaining to understanding are made in a variety of contexts and ways. As a result, few in the philosophical literature have made an attempt to precisely characterize the state that is y understanding x. This paper builds an account that does just that. The account is motivated by two main observations. First, understanding x is somehow related to being able to manipulate x. Second, understanding is a mental phenomenon, and so what manipulations are required to be an understander must (...) only be mental manipulations. Combining these two insights, the paper builds an account (URM) of understanding as a certain representational capacity—specifically, understanding x involves possessing a representation of x that could be manipulated in useful ways. By tying understanding to representation, the account correctly identifies that understanding is a fundamentally cognitive achievement. However, by also demanding that which representations count as understanding-conferring be determined by their practical effects, URM captures the insight that understanding is vitally connected to practice. URM is fully general, and can apply equally well to understanding states of affairs, understanding events, and even understanding people and works of art. The ultimate test of URM is its applicability in actual scientific and philosophical discourse. To that end the paper discusses the importance of understanding in the philosophy of science, psychology, and computer science. (shrink)

This paper focuses on two questions: Is understanding intimately bound up with accurately representing the world? Is understanding intimately bound up with downstream abilities? We will argue that the answer to both these questions is “yes”, and for the same reason-both accuracy and ability are important elements of orthogonal evaluative criteria along which understanding can be assessed. More precisely, we will argue that representational-accuracy and intelligibility are good-making features of a state of understanding. Interestingly, both evaluative claims have been defended (...) by philosophers in the literature on understanding as the criterion of evaluation. We argue that proponents of both approaches have important insights and that, drawing on both their own observations and a few novel arguments, we can construct a more complete picture of understanding evaluation. We thus posit the theory of there being Multiple Understanding Dimensions. The main thing to note about our dualism regarding the evaluative criteria of understanding is that it accounts for the intuitions about cases underlying both previously held positions. (shrink)

In this paper, I argue that explanations just ARE those sorts of things that, under the right circumstances and in the right sort of way, bring about understanding. This raises the question of why such a seemingly simple account of explanation, if correct, would not have been identified and agreed upon decades ago. The answer is that only recently has it been made possible to analyze explanation in terms of understanding without the risk of collapsing both to merely phenomenological states. (...) For the most part, theories of explanation were for 50 years held hostage to the historical accident that they far outstripped in sophistication corresponding accounts of understanding. (shrink)

Philosophers of science increasingly recognize the importance of idealization: the intentional introduction of distortion into scientific theories. Yet this recognition has not yielded consensus about the nature of idealization. e literature of the past thirty years contains disparate characterizations and justifications, but little evidence of convergence towards a common position.

Simple idealized models seem to provide more understanding than opaque, complex, and hyper-realistic models. However, an increasing number of scientists are going in the opposite direction by utilizing opaque machine learning models to make predictions and draw inferences, suggesting that scientists are opting for models that have less potential for understanding. Are scientists trading understanding for some other epistemic or pragmatic good when they choose a machine learning model? Or are the assumptions behind why minimal models provide understanding misguided? In (...) this paper, using the case of deep neural networks, I argue that it is not the complexity or black box nature of a model that limits how much understanding the model provides. Instead, it is a lack of scientific and empirical evidence supporting the link that connects a model to the target phenomenon that primarily prohibits understanding. (shrink)

There is considerable agreement among epistemologists that certain abilities are constitutive of understanding-why. These abilities include: constructing explanations, drawing conclusions, and answering questions. This agreement has led epistemologists to conclude that understanding is a kind of know-how. However, in this paper, I argue that the abilities constitutive of understanding are the same kind of cognitive abilities that we find in ordinary cases of knowledge-that and not the kind of practical abilities associated with know-how. I argue for this by disambiguating between (...) different senses of abilities that are too often lumped together. As a consequence, non-reductionists about understanding—those that claim that understanding-why is not reducible to knowledge-that—need to find another way to motivate the view. In the end, the fact that abilities are constitutive of understanding-why does not give us reason to conclude that understanding is a kind of know-how. (shrink)

The ontological dependence of one domain on another is compatible with the explanatory autonomy of the less basic domain. That autonomy results from the fact that the relationship between two domains can be very complex. In this paper I distinguish two different types of complexity, two ways the relationship between domains can fail to be transparent, both of which are relevant to evolutionary biology. Sometimes high level explanations preserve a certain type of causal or counterfactual information which would be lost (...) at the lower level; I argue that this is central to the proper understanding of the adaptationist program. Sometimes high level kinds are multiply realised by lower level kinds: I argue that this is central to the understanding of macroevolution. (shrink)

Scientific explanations must bear the proper relationship to the world: they must depict what, out in the world, is responsible for the explanandum. But explanations must also bear the proper relationship to their audience: they must be able to create human understanding. With few exceptions, philosophical accounts of explanation either ignore entirely the relationship between explanations and their audience or else demote this consideration to an ancillary role. In contrast, I argue that considering an explanation’s communicative role is crucial to (...) any satisfactory account of explanation. (shrink)

I argue for the causal character of modeling in data-intensive science, contrary to widespread claims that big data is only concerned with the search for correlations. After discussing the concept of data-intensive science and introducing two examples as illustration, several algorithms are examined. It is shown how they are able to identify causal relevance on the basis of eliminative induction and a related difference-making account of causation. I then situate data-intensive modeling within a broader framework of an epistemology of scientific (...) knowledge. In particular, it is shown to lack a pronounced hierarchical, nested structure. The significance of the transition to such “horizontal” modeling is underlined by the concurrent emergence of novel inductive methodology in statistics such as non-parametric statistics. Data-intensive modeling is well equipped to deal with various aspects of causal complexity arising especially in the higher level and applied sciences. (shrink)

Recent claims, mainly from computer scientists, concerning a largely automated and model-free data-intensive science have been countered by critical reactions from a number of philosophers of science. The debate suffers from a lack of detail in two respects, regarding the actual methods used in data-intensive science and the specific ways in which these methods presuppose theoretical assumptions. I examine two widely-used algorithms, classificatory trees and non-parametric regression, and argue that these are theory-laden in an external sense, regarding the framing of (...) research questions, but not in an internal sense concerning the causal structure of the examined phenomenon. With respect to the novelty of data-intensive science, I draw an analogy to exploratory as opposed to theory-directed experimentation. (shrink)

Has the rise of data-intensive science, or ‘big data’, revolutionized our ability to predict? Does it imply a new priority for prediction over causal understanding, and a diminished role for theory and human experts? I examine four important cases where prediction is desirable: political elections, the weather, GDP, and the results of interventions suggested by economic experiments. These cases suggest caution. Although big data methods are indeed very useful sometimes, in this paper’s cases they improve predictions either limitedly or not (...) at all, and their prospects of doing so in the future are limited too. (shrink)

Although decision‐making algorithms are not new to medicine, the availability of vast stores of medical data, gains in computing power, and breakthroughs in machine learning are accelerating the pace of their development, expanding the range of questions they can address, and increasing their predictive power. In many cases, however, the most powerful machine learning techniques purchase diagnostic or predictive accuracy at the expense of our ability to access “the knowledge within the machine.” Without an explanation in terms of reasons or (...) a rationale for particular decisions in individual cases, some commentators regard ceding medical decision‐making to black box systems as contravening the profound moral responsibilities of clinicians. I argue, however, that opaque decisions are more common in medicine than critics realize. Moreover, as Aristotle noted over two millennia ago, when our knowledge of causal systems is incomplete and precarious—as it often is in medicine—the ability to explain how results are produced can be less important than the ability to produce such results and empirically verify their accuracy. (shrink)

The usefulness of machine learning algorithms has led to their widespread adoption prior to the development of a conceptual framework for making sense of them. One common response to this situation is to say that machine learning suffers from a “black box problem.” That is, machine learning algorithms are “opaque” to human users, failing to be “interpretable” or “explicable” in terms that would render categorization procedures “understandable.” The purpose of this paper is to challenge the widespread agreement about the existence (...) and importance of a black box problem. The first section argues that “interpretability” and cognates lack precise meanings when applied to algorithms. This makes the concepts difficult to use when trying to solve the problems that have motivated the call for interpretability. Furthermore, since there is no adequate account of the concepts themselves, it is not possible to assess whether particular technical features supply formal definitions of those concepts. The second section argues that there are ways of being a responsible user of these algorithms that do not require interpretability. In many cases in which a black box problem is cited, interpretability is a means to a further end such as justification or non-discrimination. Since addressing these problems need not involve something that looks like an “interpretation” of an algorithm, the focus on interpretability artificially constrains the solution space by characterizing one possible solution as the problem itself. Where possible, discussion should be reformulated in terms of the ends of interpretability. (shrink)

I consider recent strategies proposed by econometricians for extrapolating causal effects from experimental to target populations. I argue that these strategies fall prey to the extrapolator’s circle: they require so much knowledge about the target population that the causal effects to be extrapolated can be identified from information about the target alone. I then consider comparative process tracing as a potential remedy. Although specifically designed to evade the extrapolator’s circle, I argue that CPT is unlikely to facilitate extrapolation in typical (...) econometrics and evidence-based policy applications. To argue this, I offer a distinction between two kinds of extrapolation, attributive and predictive, the latter being prevalent in econometrics and evidence-based policy. I argue that CPT is not helpful for predictive extrapolation when using the kinds of evidence that econometricians and evidence-based policy researchers prefer. I suggest that econometricians may need to consider qualitative evidence to overcome this problem. (shrink)

The literature on the nature of understanding can be divided into two broad camps. Explanationists believe that it is knowledge of explanations that is key to understanding. In contrast, their manipulationist rivals maintain that understanding essentially involves an ability to manipulate certain representations. The aim of this paper is to provide a novel knowledge based account of understanding. More specifically, it proposes an account of maximal understanding of a given phenomenon in terms of fully comprehensive and maximally well-connected knowledge of (...) it and of degrees of understanding in terms of approximations to such knowledge. It is completed by a contextualist semantics for outright attributions of understanding according to which an attribution of understanding is true of one just in case one knows enough about it to perform some contextually determined task. It is argued that this account has an edge over both its explanationist and manipulationist competitors. (shrink)

Many of the things that we try to explain, in both our common sense and our scientific engagement with the world, are capable of being explained more or less finely: that is, with greater or lesser attention to the detail of the producing mechanism. A natural assumption, pervasive if not always explicit, is that other things being equal, the more finegrained an explanation, the better. Thus, Jon Elster, who also thinks there are instrumental reasons for wanting a more fine-grained explanation, (...) assumes that in any case the mere fact of getting nearer the detail of production makes such an explanation intrinsically superior: “a more detailed explanation is also an end in itself”. Michael Taylor agrees: “A good explanation should be, amongst other things, as fine-grained as possible.”. (shrink)

I draw upon speech act theory to understand the speech acts appropriate to the multiple aims of scientific practice and the role of nonepistemic values in evaluating speech acts made relative to those aims. First, I look at work that distinguishes explaining from describing within scientific practices. I then argue speech act theory provides a framework to make sense of how explaining, describing, and other acts have different felicity conditions. Finally, I argue that if explaining aims to convey understanding to (...) particular audiences rather than describe literally across all contexts, then evaluating explanatory acts directed to the public or policymakers involves asking nonepistemic questions. (shrink)

We argue that artificial networks are explainable and offer a novel theory of interpretability. Two sets of conceptual questions are prominent in theoretical engagements with artificial neural networks, especially in the context of medical artificial intelligence: Are networks explainable, and if so, what does it mean to explain the output of a network? And what does it mean for a network to be interpretable? We argue that accounts of “explanation” tailored specifically to neural networks have ineffectively reinvented the wheel. In (...) response to, we show how four familiar accounts of explanation apply to neural networks as they would to any scientific phenomenon. We diagnose the confusion about explaining neural networks within the machine learning literature as an equivocation on “explainability,” “understandability” and “interpretability.” To remedy this, we distinguish between these notions, and answer by offering a theory and typology of interpretation in machine learning. Interpretation is something one does to an explanation with the aim of producing another, more understandable, explanation. As with explanation, there are various concepts and methods involved in interpretation: Total or Partial, Global or Local, and Approximative or Isomorphic. Our account of “interpretability” is consistent with uses in the machine learning literature, in keeping with the philosophy of explanation and understanding, and pays special attention to medical artificial intelligence systems. (shrink)

Inferences like these are known as extrapolations.

Various publications claim that medical AI systems perform as well, or better, than clinical experts. However, there have been very few controlled trials and the quality of existing studies has been called into question. There is growing concern that existing studies overestimate the clinical benefits of AI systems. This has led to calls for more, and higher-quality, randomized controlled trials of medical AI systems. While this a welcome development, AI RCTs raise novel methodological challenges that have seen little discussion. We (...) discuss some of the challenges arising in the context of AI RCTs and make some suggestions for how to meet them. (shrink)