Tim Räz

Recently, watermarking schemes for large language models (LLMs) have been proposed to distinguish text generated by machines and by humans. The present paper explores philosophical, political, and ethical ramifications of implementing and using watermarking schemes. A definition of authorship that includes both machines (LLMs) and humans is proposed to serve as a backdrop. It is argued that private watermarks may provide private companies with sweeping rights to determine authorship, which is incompatible with traditional standards of authorship determination. Then, possible ramifications of the so-called entropy dependence of watermarking mechanisms are explored. It is argued that entropy may vary for different, socially salient groups. This could lead to group dependent rates at which machine generated text is detected. Specifically, groups more interested in low entropy text may face the challenge that it is harder to detect machine generated text that is of interest to them.

Future AI systems may need to provide medical professionals with explanations of AI predictions and decisions. While current XAI methods match these requirements in principle, they are too inflexible and not sufficiently geared toward clinicians’ needs to fulfill this role. This paper offers a conceptual roadmap for how XAI may be integrated into future medical practice. We identify three desiderata of increasing difficulty: First, explanations need to be provided in a context- and user-dependent manner. Second, explanations need to be created through a genuine dialogue between AI and human users. Third, AI systems need genuine social capabilities. We use an imaginary stroke treatment scenario as a foundation for our roadmap to explore how the three challenges emerge at different stages of clinical practice. We provide definitions of key concepts such as genuine dialogue and social capability, we discuss why these capabilities are desirable, and we identify major roadblocks. Our goal is to help practitioners and researchers in developing future XAI that is capable of operating as a participant in complex medical environments. We employ an interdisciplinary methodology that integrates medical XAI, medical practice, and philosophy.

The present paper examines an episode from the historiography of the genesis of general relativity. Einstein rejected a certain theory in the so-called “Zurich notebook” in 1912–13, but he reinstated the same theory for a short period of time in the November of 1915. Why did Einstein reject the theory at first, and why did he change his mind later? The group of Einstein scholars who reconstructed Einstein’s reasoning in the Zurich notebook disagree on how to answer these questions. According to the “majority view”, Einstein was unaware of so-called “coordinate conditions”, and he relied on so-called “coordinate restrictions”. John Norton, on the other hand, claims that Einstein must have had coordinate conditions all along, but that he committed a different mistake, which he would repeat in the context of the famous “hole argument”. After an account of the two views, and of the reactions by the respective opponents, I will probe the two views for weaknesses, and try to determine how we might settle the disagreement. Finally, I will discuss emerging methodological issues.

In computer science, there are efforts to make machine learning more interpretable or explainable, and thus to better understand the underlying models, algorithms, and their behavior. But what exactly is interpretability, and how can it be achieved? Such questions lead into philosophical waters because their answers depend on what explanation and understanding are—and thus on issues that have been central to the philosophy of science. In this paper, we review the recent philosophical literature on interpretability. We propose a systematization in terms of four tasks for philosophers: (i) clarify the notion of interpretability, (ii) explain the value of interpretability, (iii) provide frameworks to think about interpretability, and (iv) explore important features of it to adjust our expectations about it.

In a recent paper, Aidan Lyon and Mark Colyvan have proposed an explanation of the structure of the bee's honeycomb based on the mathematical Honeycomb Conjecture. This explanation has instantly become one of the standard examples in the philosophical debate on mathematical explanations of physical phenomena. In this critical note, I argue that the explanation is not scientifically adequate. The reason for this is that the explanation fails to do justice to the essentially three-dimensional structure of the bee's honeycomb

We respond to two kinds of skepticism about integrated history and philosophy of science: foundational and methodological. Foundational skeptics doubt that the history and the philosophy of science have much to gain from each other in principle. We therefore discuss some of the unique rewards of work at the intersection of the two disciplines. By contrast, methodological skeptics already believe that the two disciplines should be related to each other, but they doubt that this can be done successfully. Their worries are captured by the so-called dilemma of case studies: On one horn of the dilemma, we begin our integrative enterprise with philosophy and proceed from there to history, in which case we may well be selecting our historical cases so as to fit our preconceived philosophical theses. On the other horn, we begin with history and proceed to philosophical reflection, in which case we are prone to unwarranted generalization from particulars. Against worries about selection bias, we argue that we routinely need to make explicit the criteria for choosing particular historical cases to investigate particular philosophical theses. It then becomes possible to ask whether or not the selection criteria were biased. Against worries about unwarranted generalization, we stress the iterative nature of the process by which historical data and philosophical concepts are brought into alignment. The skeptics’ doubts are fueled by an outdated model of outright confirmation versus outright falsification of philosophical concepts. A more appropriate model is one of stepwise and piecemeal improvement.

The Lotka–Volterra predator-prey-model is a widely known example of model-based science. Here we reexamine Vito Volterra’s and Umberto D’Ancona’s original publications on the model, and in particular their methodological reflections. On this basis we develop several ideas pertaining to the philosophical debate on the scientific practice of modeling. First, we show that Volterra and D’Ancona chose modeling because the problem in hand could not be approached by more direct methods such as causal inference. This suggests a philosophically insightful motivation for choosing the strategy of modeling. Second, we show that the development of the model follows a trajectory from a “how possibly” to a “how actually” model. We discuss how and to what extent Volterra and D’Ancona were able to advance their model along that trajectory. It turns out they were unable to establish that their model was fully applicable to any system. Third, we consider another instance of model-based science: Darwin’s model of the origin and distribution of coral atolls in the Pacific Ocean. Darwin argued more successfully that his model faithfully represents the causal structure of the target system, and hence that it is a “how actually” model.

Parameterization and parameter tuning are central aspects of climate modeling, and there is widespread consensus that these procedures involve certain subjective elements. Even if the use of these subjective elements is not necessarily epistemically problematic, there is an intuitive appeal for replacing them with more objective (automated) methods, such as machine learning. Relying on several case studies, we argue that, while machine learning techniques may help to improve climate model parameterization in several ways, they still require expert judgment that involves subjective elements not so different from the ones arising in standard parameterization and tuning. The use of machine learning in parameterizations is an art as well as a science and requires careful supervision.

Machine learning methods have recently created high expectations in the climate modelling context in view of addressing climate change, but they are often considered as non-physics-based ‘black boxes’ that may not provide any understanding. However, in many ways, understanding seems indispensable to appropriately evaluate climate models and to build confidence in climate projections. Relying on two case studies, we compare how machine learning and standard statistical techniques affect our ability to understand the climate system. For that purpose, we put five evaluative criteria of understanding to work: intelligibility, representational accuracy, empirical accuracy, coherence with background knowledge, and assessment of the domain of validity. We argue that the two families of methods are part of the same continuum where these various criteria of understanding come in degrees, and that therefore machine learning methods do not necessarily constitute a radical departure from standard statistical tools, as far as understanding is concerned.

I raise an objection to Stewart Shapiro's version of ante rem structuralism: I show that it is in conflict with mathematical practice. Shapiro introduced so-called ‘finite cardinal structures’ to illustrate features of ante rem structuralism. I establish that these structures have a well-known counterpart in mathematics, but this counterpart is incompatible with ante rem structuralism. Furthermore, there is a good reason why, according to mathematical practice, these structures do not behave as conceived by Shapiro's ante rem structuralism

The paper presents, and discusses, four candidate explanations of the structure, and construction, of the bees’ honeycomb. So far, philosophers have used one of these four explanations, based on the mathematical Honeycomb Conjecture, while the other three candidate explanations have been ignored. I use the four cases to resolve a dispute between Pincock and Baker about the Honeycomb Conjecture explanation. Finally, I find that the two explanations focusing on the construction mechanism are more promising than those focusing exclusively on the resulting, optimal structure. The main reason for this is that optimal structures do not uniquely determine the relevant optimization leading to the optimal structure

Some machine learning models, in particular deep neural networks (DNNs), are not very well understood; nevertheless, they are frequently used in science. Does this lack of understanding pose a problem for using DNNs to understand empirical phenomena? Emily Sullivan has recently argued that understanding with DNNs is not limited by our lack of understanding of DNNs themselves. In the present paper, we will argue, _contra_ Sullivan, that our current lack of understanding of DNNs does limit our ability to understand with DNNs. Sullivan’s claim hinges on which notion of understanding is at play. If we employ a weak notion of understanding, then her claim is tenable, but rather weak. If, however, we employ a strong notion of understanding, particularly explanatory understanding, then her claim is not tenable.

The present paper provides an analysis of Euler’s solutions to the Königsberg bridges problem. Euler proposes three different solutions to the problem, addressing their strengths and weaknesses along the way. I put the analysis of Euler’s paper to work in the philosophical discussion on mathematical explanations. I propose that the key ingredient to a good explanation is the degree to which it provides relevant information. Providing relevant information is based on knowledge of the structure in question, graphs in the present case. I also propose computational complexity and logical strength as measures of relevant information.

This paper argues that a notion of statistical explanation, based on Salmon’s statistical relevance model, can help us better understand deep neural networks. It is proved that homogeneous partitions, the core notion of Salmon’s model, are equivalent to minimal sufficient statistics, an important notion from statistical inference. This establishes a link to deep neural networks via the so-called Information Bottleneck method, an information-theoretic framework, according to which deep neural networks implicitly solve an optimization problem that generalizes minimal sufficient statistics. The resulting notion of statistical explanation is general, mathematical, and subcausal.

Michael Weisberg and Kenneth Reisman argue that the Volterra Principle can be derived from multiple predator-prey models and that, therefore, the Volterra Principle is a prime example for robustness analysis. In the current article, I give new results regarding the Volterra Principle, extending Weisberg’s and Reisman’s work, and I discuss the consequences of these results for robustness analysis. I argue that we do not end up with multiple, independent models but rather with one general model. I identify the kind of situation in which this generalization approach may occur, and I analyze the generalized Volterra Principle from an explanatory perspective.

We outline a framework for analyzing episodes from the history of science in which the application of mathematics plays a constitutive role in the conceptual development of empirical sciences. Our starting point is the inferential conception of the application of mathematics, recently advanced by Bueno and Colyvan. We identify and discuss some systematic problems of this approach. We propose refinements of the inferential conception based on theoretical considerations and on the basis of a historical case study. We demonstrate the usefulness of the refined, dynamical inferential conception using the well-researched example of the genesis of general relativity. Specifically, we look at the collaboration of the physicist Einstein and the mathematician Grossmann in the years 1912--1913, which resulted in the jointly published ``Outline of a Generalized Theory of Relativity and a Theory of Gravitation,'' a precursor theory of the final theory of general relativity. In this episode, an independently developed mathematical theory, the theory of differential invariants and the absolute differential calculus, was applied in the process of physical theorizing aiming at finding a relativistic theory of gravitation. We argue that the dynamical inferential conception not only provides a natural framework to describe and analyze this episode, but it also generates new questions and insights. We comment on the mathematical tradition on which Grossmann drew, and on his own contributions to mathematical theorizing. We argue that the dynamical inferential conception allows us to identify both the role of heuristics and of mathematical resources as well as the systematic role of problems and mistakes in the reconstruction of episodes of conceptual innovation and theory change