Explainable AI (xAI) methods are important for establishing trust in using black-box models. However, recent criticism has mounted against current xAI methods that they disagree, are necessarily false, and can be manipulated, which has started to undermine the deployment of black-box models. Rudin (2019) goes so far as to say that we should stop using black-box models altogether in high-stakes cases because xAI explanations ‘must be wrong’. However, strict fidelity to the truth is historically not a desideratum in science. Idealizations (...) – the intentional distortions introduced to scientific theories and models – are commonplace in the natural sciences and are seen as a successful scientific tool. Thus, it is not falsehood qua falsehood that is the issue. In this paper, I outline the need for xAI research to engage in idealization evaluation. Drawing on the use of idealizations in the natural sciences and philosophy of science, I introduce a novel framework for evaluating whether xAI methods engage in successful idealizations or deceptive explanations (SIDEs). SIDEs evaluates whether the limitations of xAI methods, and the distortions that they introduce, can be part of a successful idealization or are indeed deceptive distortions as critics suggest. I discuss the role that existing research can play in idealization evaluation and where innovation is necessary. Through a qualitative analysis we find that leading feature importance methods and counterfactual explanations are subject to idealization failure and suggest remedies for ameliorating idealization failure. (shrink)

We argue that the trend toward providing users with feasible and actionable explanations of AI decisions—known as recourse explanations—comes with ethical downsides. Specifically, we argue that recourse explanations face several conceptual pitfalls and can lead to problematic explanation hacking, which undermines their ethical status. As an alternative, we advocate that explanations of AI decisions should aim at understanding.

The description of how individual models in families of models are related to each other is crucial for the general philosophical understanding of model-based scientific practice. We focus on the Capital Asset Pricing Models (CAPM) family, a cornerstone in financial economics, to provide a descriptive analysis of model relations within a family. We introduce the concepts of theoretical and empirical complementarity to characterise model relations. Our complementarity analysis of model relations has two types of payoff. Specifically regarding the CAPM, our (...) analysis reveals why this model family, which has been empirically contested, has yet remained popular and important: the different models that have been added over time have made new empirically and theoretically complementary contributions to the model family. More generally, our analysis reveals the dynamic character of model-based scientific practice. Characterising relations between models as theoretically and/or empirically complementary yields three hitherto underappreciated lessons: (i) actual modelling purposes are not always primarily epistemic, (ii) a model's individual import is relative, not absolute, and (iii) there is an important interplay between theory and data models. Faithfully characterising scientific modelling in this way facilitates subsequent analysis of models' epistemic import. (shrink)

Both ecologists and statistical physicists use a variety of highly idealized models to study active matter and self-organizing critical phenomena. In this paper, I show how universality classes play a crucial role in justifying the application of highly idealized ‘minimal’ models to explain and understand the critical behaviors of active matter systems across a wide range of scales and scientific fields. Appealing to universality enables us to see why the same minimal models can be used to explain and understand behaviors (...) across these different systems despite drastic differences in the causes and mechanisms responsible for the behaviors of interest. After analyzing these cases in detail, I argue that accounts that focus on identifying common causes or mechanisms in order to explain patterns are unable to accommodate these cases. In contrast, I argue that the justification for using these minimal models is that they are within the same universality class as real systems whose causes and mechanisms are known to be different. I also use these cases to identify several different kinds of explanatory autonomy that have important implications for how scientists ought to approach the modeling of multiscale phenomena. (shrink)

This paper proposes an account of accurate scientific representation in terms of techniques that produce data from a target phenomenon. I consider an approach to accurate representation that abstracts from such epistemic factors, justified by a thesis I call Ontic Priority. This holds that criteria for representational accuracy depend on a pre-established account of the nature of the relation between a model and its target phenomenon. I challenge Ontic Priority, drawing on the observation that many working scientists do not have (...) access to information allowing them to describe this relation between model and target or compare them accordingly. I critique the ability of an ontic-first approach to provide accuracy criteria in such cases and present historical support for an alternative according to which a model is accurate if and only if integrating the model into a theory of the data acquisition process yields well-fitting predictions of patterns in the data. (shrink)

Empiricist modal epistemologies can be attractive, but are often limited in the range of modal knowledge they manage to secure. In this paper, I argue that one such account – similarity-based modal empiricism – can be extended to also cover justification of many scientifically interesting possibility claims. Drawing on recent work on modelling in the philosophy of science, I suggest that scientific modelling is usefully seen as the creation and investigation of relevantly similar epistemic counterparts of real target systems. On (...) the basis of experiential knowledge of what is actually the case with the models, one can draw justified conclusions about what is de re possible for the target systems. (shrink)

The notion of understanding occupies an increasingly prominent place in contemporary epistemology, philosophy of science, and moral theory. A central and ongoing debate about the nature of understanding is how it relates to the truth. In a series of influential contributions, Catherine Elgin has used a variety of familiar motivations for antirealism in philosophy of science to defend a non- factive theory of understanding. Key to her position are: (i) the fact that false theories can contribute to the upwards trajectory (...) of scientific understanding, and (ii) the essential role of inaccurate idealisations in scientific research. Using Elgin’s arguments as a foil, I show that a strictly factive theory of understanding has resources with which to offer a unified response to both the problem of idealisations and the role of false theories in the upwards trajectory of scientific understanding. Hence, strictly factive theories of understanding are viable notwithstanding these forceful criticisms. (shrink)

There are two traditions of thinking about idealization offering almost opposite views on their functioning and epistemic status. While one tradition views idealizations as epistemic deficiencies, the other one highlights the epistemic benefits of idealization. Both of these, however, identify idealization with misrepresentation. In this article, we instead approach idealization from the artifactual perspective, comparing it to the distortion-to-reality accounts of idealization, and exemplifying it through the case of the Hodgkin and Huxley model of nerve impulse. From the artifactual perspective, (...) the epistemic benefits and deficiencies introduced by idealization frequently come in a package due to the way idealization draws together different resources in model construction. Accordingly, idealization tends to be holistic in that it is not often easily attributable to just some specific parts of the model. Instead, the idealizing process tightly embeds theoretical concepts and formal tools into the construction of a model. (shrink)

According to an adequacy-for-purpose view, models should be assessed with respect to their adequacy or fitness for particular purposes. Such a view has been advocated by scientists and philosophers...

As the demand for explainable deep learning grows in the evaluation of language technologies, the value of a principled grounding for those explanations grows as well. Here we study the state-of-the-art in explanation for neural models for natural-language processing (NLP) tasks from the viewpoint of philosophy of science. We focus on recent evaluation work that finds brittleness in explanations obtained through attention mechanisms.We harness philosophical accounts of explanation to suggest broader conclusions from these studies. From this analysis, we assert the (...) impossibility of causal explanations from attention layers over text data. We then introduce NLP researchers to contemporary philosophy of science theories that allow robust yet non-causal reasoning in explanation, giving computer scientists a vocabulary for future research. (shrink)

Scientists appeal to models when explaining phenomena. Such explanations are often dubbed model explanations or model-based explanations. But what are the precise conditions for ME? Are ME special explanations? In our paper, we first rebut two definitions of ME and specify a more promising one. Based on this analysis, we single out a related conception that is concerned with explanations that are induced from working with a model. We call them ‘model-induced explanations’. Second, we study three paradigmatic cases of alleged (...) ME. We argue that all of them are MIE, upon closer examination. Third, we argue that this undermines the building consensus that model explanations are special explanations that, e.g., challenge the factivity of explanation. Instead, it suggests that what is special about models in science is the epistemology behind how models induce explanations. (shrink)

Deidealization as a topic in its own right has attracted remarkably little philosophical interest despite the extensive literature on idealization. One reason for this is the often implicit assumption that idealization and deidealization are, potentially at least, reversible processes. We question this assumption by analyzing the challenges of deidealization within a menu of four broad categories: deidealizing as recomposing, deidealizing as reformulating, deidealizing as concretizing, and deidealizing as situating. On closer inspection, models turn out much more inflexible than the reversal (...) thesis would have us believe, and deidealization emerges as a creative part of modeling. (shrink)

Metaphors are found all throughout science: in published papers, working hypotheses, policy documents, lecture slides, grant proposals, and press releases. They serve different functions, but perhaps most striking is the way they enable understanding, of a theory, phenomenon, or idea. In this paper, we leverage recent advances on the nature of metaphor and the nature of understanding to explore how they accomplish this feat. We attempt to shift the focus away from the epistemic value of the content of metaphors, to (...) the epistemic value of the metaphor’s consequences. Many famous scientific metaphors are epistemically good, not primarily because of what they say about the world, but because of how they cause us to think. Specifically, metaphors increase understanding either by improving our sets of representations, or by making it easier for us to encode and process data about complex subjects by changing how we are disposed to conceptualize those subjects. This view hints towards new positions concerning testimonial understanding, factivity, abilities, discovery via metaphor, and the relation between metaphors and models. (shrink)

This paper defends the viability of de-idealization strategies in economic modeling against recent criticism. De-idealization occurs when an idealized assumption of a theoretical model is replaced with a more realistic one. Recently, some scholars have raised objections against the possibility or fruitfulness of de-idealizing economic models, suggesting that economists do not employ this kind of strategy. We present a detailed case study from the theory of industrial organization, discussing three different models, two of which can be construed as de-idealized versions (...) of the first (the so-called Bertrand model of oligopoly). We conclude that recent pessimism about de-idealization in economics is largely unfounded, and that de-idealization strategies are not only possible but also widely employed in economics. (shrink)

Idealization is commonly understood as distortion: representing things differently than how they actually are. In this paper, we outline an alternative artifactual approach that does not make misrepresentation central for the analysis of idealization. We examine the contrast between the Hodgkin-Huxley (1952a, b, c) and the Heimburg-Jackson (2005, 2006) models of the nerve impulse from the artifactual perspective, and argue that, since the two models draw upon different epistemic resources and research programs, it is often difficult to tell which features (...) of a system the central assumptions involved are supposed to distort. Many idealizations are holistic in nature. They cannot be locally undone without dismantling the model, as they occupy a central position in the entire research program. Nor is their holistic character mainly related to the use of mathematical and statistical modeling techniques as portrayed by Rice (2018, 2019). We suggest that holistic idealizations are implicit theoretical and representational assumptions that can only be understood in relation to the conceptual and representational tools exploited in modeling and experimental practices. Such holistic idealizations play a pivotal role not just in individual models, but also in defining research programs. -/- . (shrink)

How can a reflective scientist put forward an explanation using a model when they are aware that many of the assumptions used to specify that model are false? This paper addresses this challenge by making two substantial assumptions about explanatory practice. First, many of the propositions deployed in the course of explaining have a non-representational function. In particular, a proposition that a scientist uses and also believes to be false, i.e. an “idealization”, typically has some non-representational function in the practice, (...) such as the interpretation of some model or the specification of the target of the explanation. Second, when an agent puts forward an explanation using a model, they usually aim to remain agnostic about various features of the phenomenon being explained. In this sense, explanations are intended to be autonomous from many of the more fundamental features of such systems. I support these two assumptions by showing how they allow one to address a number of recent concerns raised by Bokulich, Potochnik and Rice. In addition, these assumptions lead to a defense of the view that explanations are wholly true that improves on the accounts developed by Craver, Mäki and Strevens. (shrink)

This paper defends three claims about concrete or physical models: these models remain important in science and engineering, they are often essentially idealized, in a sense to be made precise, and despite these essential idealizations, some of these models may be reliably used for the purpose of causal explanation. This discussion of concrete models is pursued using a detailed case study of some recent models of landslide generated impulse waves. Practitioners show a clear awareness of the idealized character of these (...) models, and yet address these concerns through a number of methods. This paper focuses on experimental arguments that show how certain failures to accurately represent feature X are consistent with accurately representing some causes of feature Y, even when X is causally relevant to Y. To analyse these arguments, the claims generated by a model must be carefully examined and grouped into types. Only some of these types can be endorsed by practitioners, but I argue that these endorsed claims are sufficient for limited forms of causal explanation. (shrink)

The concept of explanation is central to scientific practice. However, scientists explain phenomena in very different ways. That is, there are many different kinds of explanation; e.g. causal, mechanistic, statistical, or equilibrium explanations. In light of the myriad kinds of explanation identified in the literature, most philosophers of science have adopted some kind of explanatory pluralism. While pluralism about explanation seems plausible, it faces a dilemma Explanation beyond causation, Oxford University Press, Oxford, pp 39–56, 2018). Either there is nothing that (...) unifies all instances of scientific explanation that makes them count as explanations, or there is some set of unifying features, which seems incompatible with explanatory pluralism. Different philosophers have adopted different horns of this dilemma. Some argue that no unified account of explanation is possible. Others suggest that there is a set of necessary features that can unify all explanations under a single account Explanation beyond causation, Oxford University Press, Oxford, pp 74–95, 2018; Strevens in Depth: an account of scientific explanation, Harvard University Press, Cambridge, 2008). In this paper, we argue that none of the features identified by existing accounts of explanation are necessary for all explanations. However, we argue that a unified account can still be provided that accommodates pluralism. This can be accomplished, we argue, by reconceiving of scientific explanation as a cluster concept: there are multiple subsets of features that are sufficient for providing an explanation, but no single feature is necessary for all explanations. Reconceiving of explanation as a cluster concept not only accounts for the diversity of kinds of explanations, but also accounts for the widespread disagreement in the explanation literature and enables explanatory pluralism to avoid Pincock’s dilemma. (shrink)

Relationships of counterfactual dependence have played a major role in recent debates of explanation and understanding in the philosophy of science. Usually, counterfactual dependencies have been viewed as the explanantia of explanation, i.e., the things providing explanation and understanding. Sometimes, however, counterfactual dependencies are themselves the targets of explanations in science. These kinds of explanations are the focus of this paper. I argue that “micro-level model explanations” explain the particular form of the empirical regularity underlying a counterfactual dependency by representing (...) it as a physical necessity on the basis of postulated microscopic entities. By doing so, micro-level models rule out possible forms the regularity could have taken. Micro-model explanations, in other words, constrain empirical regularities and their associated counterfactual dependencies. I introduce and illustrate micro-level model explanations in detail, contrast them to other accounts of explanation, and consider potential problems. (shrink)

Using past episodes of climate change as a source of evidence to inform our projections about contemporary climate change requires establishing the extent to which episodes in the deep past are analogous to the current crisis. However, many scientists claim that contemporary rates of climate change (e.g., rates of carbon emissions or temperature change) are unprecedented, including compared to episodes in the deep past. If so, this would limit the utility of paleoclimate analogues. In this paper, I show how a (...) data adjustment procedure called “temporal scaling,” which must be applied to both contemporary and past rate data, complicates the claim that contemporary rates are truly unprecedented. On top of giving actionable recommendations to scientists, this paper advances the philosophical literature concerning the use of models that are known to be somewhat disanalogous to their target systems. (shrink)

The purpose of this essay is to present and analyse the basic assumptions of Leszek Nowak’s conception of the unity of science. According to Nowak, the unity of science is manifested in the common application of the method of idealisation in scientific research. In accordance with his conception, regardless of the discipline they represent, researchers go through the same stages in building a theory. Two key ones among them are: introducing idealising assumptions into the representation and then their concretisation. In (...) this view, idealisation is the basis of the scientific method, while other cognitive procedures complement it. Nowak’s conception has particular relevance in the context of the dispute between naturalism and anti-naturalism and in the context of the continuing rift between social scientists and natural scientists. It calls into question the anti-naturalist thesis of the ontological uniqueness of social sciences and the resulting methodological consequences. I argue that Nowak’s conception is a cognitively valuable contribution to the contemporary epistemology of science, but it also has weaknesses, mainly due to the limitations of applying the idealisation-concretisation scheme in research practice. For it turns out, as I point out in this essay, that many idealising assumptions are not subject to concretisation and that concretisations do not always condition an increase in the explanatory and/or predictive power of the representations. (shrink)

Science is replete with falsehoods that epistemically facilitate understanding by virtue of being the very falsehoods they are. In view of this puzzling fact, some have relaxed the truth requirement on understanding. I offer a factive view of understanding that fully accommodates the puzzling fact in four steps: (i) I argue that the question how these falsehoods are related to the phenomenon to be understood and the question how they figure into the content of understanding it are independent. (ii) I (...) argue that the falsehoods do not figure into the understanding’s content by being elements of its periphery or core. (iii) Drawing lessons from case studies, I argue that the falsehoods merely enable understanding. When working with such falsehoods, only the truths we extract from them are elements of the content of our understanding. (iv) I argue that the extraction view is compatible with the thesis that falsehoods can have an epistemic value by virtue of being the very falsehoods they are. (shrink)

Generative models have been proposed as a new type of non-representational scientific models recently. A generative model is characterized with the capacity of producing new models on the basis of the existing one. The current accounts do not explain sufficiently the mechanism of the generative capacity of a generative model. I attempt to accomplish this task in this paper. I outline two antecedent accounts of generative models. I point out that both types of generative models function to generate new homogenous (...) models in the sense that the latter is a straightforward derivative of the former, both of which share many similar features. Unfortunately, both accounts are implicit about the generative capacity of generative models. Using a case study, I articulate that a two-staged process of abstraction and idealization in modeling may contribute to the generative capacity of a scientific model. I also demonstrate that this two-staged process may go beyond the capacity of generating new homogenous models to generating new heterogeneous models. (shrink)

Catherine Elgin has recently argued that a nonfactive conception of understanding is required to accommodate the epistemic successes of science that make essential use of idealizations and models. In this paper, I argue that the fact that our best scientific models and theories are pervasively inaccurate representations can be made compatible with a more nuanced form of scientific realism that I call Understanding Realism. According to this view, science aims at (and often achieves) factive scientific understanding of natural phenomena. I (...) contend that this factive scientific understanding is provided by grasping a set of true modal information about the phenomenon of interest. Furthermore, contrary to Elgin’s view, I argue that the facticity of this kind of scientific understanding can be separated from the inaccuracy of the models and theories used to produce it. (shrink)

The debate over the explanatory nature of cognitive models has been waged mostly between two factions: the mechanists and the dynamical systems theorists. The former hold that cognitive models are explanatory only if they satisfy a set of mapping criteria, particularly the 3M/3m* requirement. The latter have argued, pace the mechanists, that some cognitive models are both dynamical and constitute covering-law explanations. In this paper, I provide a minimal model interpretation of dynamical cognitive models, arguing that this both provides needed (...) clarity to the mechanist versus dynamicist divide in cognitive science and also paves the way towards further insights about scientific explanation generally. (shrink)

Abstraction and idealization are the two notions that are most often discussed in the context of assumptions employed in the process of model building. These notions are also routinely used in philosophical debates such as that on the mechanistic account of explanation. Indeed, an objection to the mechanistic account has recently been formulated precisely on these grounds: mechanists cannot account for the common practice of idealizing difference-making factors in models in molecular biology. In this paper I revisit the debate and (...) I argue that the objection does not stand up to scrutiny. This is because it is riddled with a number of conceptual inconsistencies. By attempting to resolve the tensions, I also draw several general lessons regarding the difficulties of applying abstraction and idealization in scientific practice. Finally, I argue that more care is needed only when speaking of abstraction and idealization in a context in which these concepts play an important role in an argument, such as that on mechanistic explanation. (shrink)

Highly idealized models may serve various epistemic functions, notably explanation, in virtue of representing the world. Inferentialism provides a prima facie compelling characterization of what constitutes the representation relation. In this paper, I argue that what I call factive inferentialism does not provide a satisfactory solution to the puzzle of model-based—factive—explanation. In particular, I show that making explanatory counterfactual inferences is not a sufficient guide for accurate representation, factivity, or realism. I conclude by calling for a more explicit specification of (...) model-world mismatches and properties imputation. (shrink)

In scientific modeling, continuum idealizations bridge scales but at the cost of fundamentally misrepresenting the microstructure of the system. This engenders a mystery. If continuum idealizations are dispensable in principle, this de-problematizes their representational inaccuracy, since continuum properties reduce to lower-scale properties, but the mystery of how this reduction could be carried out endures. Alternatively, if continuum idealizations are indispensable in principle, this is consistent with their explanatory and predictive success but renders their representational inaccuracy mysterious. I argue for a (...) deflationary solution, enlisting the applied scientific method of upscaling as demonstrated in a case from soil hydrology. (shrink)

This paper defends the viability of de-idealization strategies in economic modeling against recent criticism. De-idealization occurs when an idealized assumption of a theoretical model is replaced with a more realistic one. Recently, some scholars have raised objections against the possibility or fruitfulness of de-idealizing economic models, suggesting that economists do not employ this kind of strategy. We present a detailed case study from the theory of industrial organization, discussing three different models, two of which can be construed as de-idealized versions (...) of the first. We conclude that recent pessimism about de-idealization in economics is largely unfounded, and that de-idealization strategies are not only possible but also widely employed in economics. (shrink)