What variables should be used to get explanations (of AI systems) that are easily interpretable? The challenge to find the right degree of abstraction in explanations, also called the ‘variables problem’, has been actively discussed in the philosophy of science. The challenge is striking the right balance between specificity and generality. Concepts such as proportionality and exhaustivity are investigated and discussed. We propose a new and formal definition based on Kolmogorov complexity and argue that this corresponds to our intuitions about the right level of abstraction. First, we require that variables are uniform, so that they cannot be decomposed into less abstract variables without increasing the Kolmogorov complexity. Next, uniform variables are optimal for an explanation if they can compose the explanation without increasing its Kolmogorov complexity. For this, the concepts K-decomposability and K-composability of sets are defined. Explanations of a certain instance should encompass a maximal set of instances without being K-decomposable. Although Kolmogorov complexity is uncomputable and depends on the choice of programming language, we show that it can be used effectively to evaluate and reason about explanations, such as in the evaluation of XAI methods.@en

ChatGPT is a powerful language model from OpenAI that is arguably able to comprehend and generate text. ChatGPT is expected to greatly impact society, research, and education. An essential step to understand ChatGPT’s expected impact is to study its domain-specific answering capabilities. Here, we perform a systematic empirical assessment of its abilities to answer questions across the natural science and engineering domains. We collected 594 questions on natural science and engineering topics from 198 faculty members across five faculties at Delft University of Technology. After collecting the answers from ChatGPT, the participants assessed the quality of the answers using a systematic scheme. Our results show that the answers from ChatGPT are, on average, perceived as “mostly correct”. Two major trends are that the rating of the ChatGPT answers significantly decreases (i) as the educational level of the question increases and (ii) as we evaluate skills beyond scientific knowledge, e.g., critical attitude.

@en

Relevancy is a prevalent term in value alignment. We either need to keep track of the relevant moral reasons, we need to embed the relevant values, or we need to learn from the relevant behaviour. What relevancy entails in particular cases, however, is often ill-defined. The reasons for this are obvious, it is hard to define relevancy in a way that is both general and concrete enough to give direction towards a specific implementation. In this paper, we describe the inherent difficulty that comes along with defining what is relevant to a particular situation. Simply due to design and the way an AI system functions, we need to state or learn particular goals and circumstances under which that goal is completed. However, because of both the changing nature of the world and the varied wielders and users of such implements, misalignment occurs, especially after a longer amount of time. We propose a way to counteract this by putting contestability front and centre throughout the lifecycle of an AI system, as it can provide insight into what is actually relevant at a particular instance. This allows designers to update the applications in such a manner that they can account for oversight during design.@en

With the widespread use of artificial intelligence, it becomes crucial to provide information about these systems and how they are used. Governments aim to disclose their use of algorithms to establish legitimacy and the EU AI Act mandates forms of transparency for all high-risk and limited-risk systems. Yet, what should the standards for transparency be? What information is needed to show to a wide public that a certain system can be used legitimately and responsibly? I argue that process-based approaches fail to satisfy, as knowledge about the development process is insufficient to predict the properties of the resulting system. Current outcome-based approaches [Mitchell et al., 2019; Loi et al., 2021] are also criticized for a lack of attention to the broader socio-technical system and failure to account for empirical results that show that people care about more than just the outcomes of a process [as reported by Meyerson et al. (Procedural justice and relational theory: Empirical, philosophical, and legal perspectives, Taylor & Francis, 2021)]. Instead, I propose value-based transparency, on which the information we need to provide is what values have been considered in the design and how successful these have been realized in the final system. This can handle the objections to other frameworks, matches with current best practices on the design of responsible AI and provides the public with information on the crucial aspects of a system’s design.@en

Machine learning techniques are driving — or soon will be driving — much of scientific research and discovery. Can they function as models similar to more traditional modeling techniques in scientific contexts? Or might they replace models altogether if they deliver sufficient predictive accuracy? These questions cut across at least two types of applications of machine learning models. First, machine learning models are used to study the brain, where neural networks might represent aspects of neural activity. The principal question here is: Can the uses of neural networks provide scientific explanations and models for neuroscience? Second, machine learning models are applied in science more broadly, where representational links are less clear. What are the epistemic implications of machine learning in those areas? Can they replace more traditional scientific models of phenomena, or is mere predictive accuracy sufficient? These two strands are finally brought together to create an overview of the (epistemic) role machine learning can play in scientific modeling.

@en

This chapter explores the principles and frameworks of human-centered artificial intelligence (AI), specifically focusing on user modeling, adaptation, and personalization. It introduces a four-dimensional framework comprising paradigms, actors, values, and levels of realization that should be considered in the design of human-centered AI systems. This framework highlights a perspective-taking approach with four lenses of technology-centric, user-centric, human-centric, and future-centric perspectives. Ethical considerations, transparency, fairness, and accountability, among other aspects, are highlighted as values when developing and deploying AI systems. The chapter further discusses the corresponding human values for each of these concepts. Opportunities and challenges in human-centered AI are examined, including the need for interdisciplinary collaboration and the complexity of addressing diverse perspectives. Human-centered AI provides valuable insights for designing AI systems that prioritize human needs, values, and experiences while considering ethical and societal implications.@en

Technologies have all kinds of impacts on the environment, on human behavior, on our society and on what we believe and value. But some technologies are not just impactful, they are also socially disruptive: they challenge existing institutions, social practices, beliefs and conceptual categories. Here we are particularly interested in technologies that disrupt existing concepts, for example because they lead to profound uncertainty about how to classify matters. Is a humanoid robot - which looks and even acts like a human - to be classified as a person or is it just an inert machine? Conceptual disruption occurs when the meaning of concepts is challenged, and such challenges may potentially lead to a revision of concepts. We illustrate how technologies can be conceptually disruptive through a range of examples, and we argue for an intercultural outlook in studying these socially disruptive technologies and conceptual disruption. Such an outlook is needed to avoid a Western bias in labeling technologies socially or conceptually disruptive, as this outlook takes inspiration from a broad range of philosophical traditions.

@en

Machine learning is used more and more in scientific contexts, from the recent breakthroughs with AlphaFold2 in protein fold prediction to the use of ML in parametrization for large climate/astronomy models. Yet it is unclear whether we can obtain scientific explanations from such models. I argue that when machine learning is used to conduct causal inference we can give a new positive answer to this question. However, these ML models are purpose-built models and there are technical results showing that standard machine learning models cannot be used for the same type of causal inference. Instead, there is a pathway to causal explanations from predictive ML models through new explainability techniques; specifically, new methods to extract structural equation models from such ML models. The extracted models are likely to suffer from issues though: they will often fail to account for confounders and colliders, as well as deliver simply incorrect causal graphs due to ML models tendency to violate physical laws such as the conservation of energy. In this case, extracted graphs are a starting point for new explanations, but predictive accuracy is no guarantee for good explanations.

@en

Process reliabilist accounts claim that a belief is justified when it is the result of a reliable belief-forming process. Yet over what range of possible token processes is this reliability calculated? I argue against the idea that all possible token processes (in the actual world, or some other subset of possible worlds) are to be considered using the case of a user acquiring beliefs based on the output of an AI system, which is typically reliable for a substantial local range but unreliable when all possible inputs are considered. I show that existing solutions to the generality problem imply that these cases cannot be solved by a more fine-grained typing of the belief-forming process. Instead, I suggest that reliability is evaluated over a range restricted by the content of the actual belief and by the similarity of the input to the actual input.@en

AI systems are increasingly being used to support human decision making. It is important that AI advice is followed appropriately. However, according to existing literature, users typically under-rely or over-rely on AI systems, and this leads to sub-optimal team performance. In this context, we investigate the role of stated system accuracy by contrasting the lack of system information with the presence of system accuracy in a loan prediction task. We explore how the degree to which humans understand system accuracy influences their reliance on the AI system, by investigating numeracy levels and with the aid of analogies to explain system accuracy in a first of its kind between-subjects study (N=281). We found that explaining the stated accuracy of a system using analogies failed to help users rely on the AI systemappropriately (i.e., the tendency of users to rely on the system when the system is correct, or on themselves otherwise). To eliminate the impact of subjective attitudes towards analogy domains, we conducted a within-subjects study (N=248) where each participant worked on tasks with analogy-based explanations from different domains. Results from this second study confirmed that explaining stated accuracy of the system with analogies was not sufficient to facilitate appropriate reliance on the AI system in the context of loan prediction tasks, irrespective of individual user differences. Based on our findings from the two studies, we reason that the under-reliance on the AI system may be a result of users' overestimation of their own ability to solve the given task. Thus, although familiar analogies can be effective in improving the intelligibility of stated accuracy of the system, an improved understanding of system accuracy does not necessarily lead to improved system reliance and team performance.

@en

Why should we explain opaque algorithms? Here four papers are discussed that argue that, in fact, we don’t have to. Explainability, according to them, isn’t needed for trust in algorithms, nor is it needed for other goals we might have. I give a critical overview of these arguments, showing that there is still room to think that explainability is required for responsible AI. With that in mind, the second part of the paper looks at how we might achieve this end goal. I proceed not from technical tools in explainability, but rather highlight accounts of explanation in philosophy that might inform what those technical tools should ultimately deliver. While there is disagreement here on what constitutes an explanation, the three accounts surveyed offer a good overview of the current theoretical landscape in philosophy and of what information might constitute an explanation. As such, they can hopefully inspire improvements to the technical explainability tools.@en

Explaining the behaviour of Artificial Intelligence models has become a necessity. Their opaqueness and fragility are not tolerable in high-stakes domains especially. Although considerable progress is being made in the field of Explainable Artificial Intelligence, scholars have demonstrated limits and flaws of existing approaches: explanations requiring further interpretation, non-standardised explanatory format, and overall fragility. In light of this fragmentation, we turn to the field of philosophy of science to understand what constitutes a good explanation, that is, a generalisation that covers both the actual outcome and, possibly multiple, counterfactual outcomes. Inspired by this, we propose CHIME: a human-in-the-loop, post-hoc approach focused on creating such explanations by establishing the causal features in the input. We first elicit people's cognitive abilities to understand what parts of the input the model might be attending to. Then, through Causal Discovery we uncover the underlying causal graph relating the different concepts. Finally, with such a structure, we compute the causal effects different concepts have towards a model's outcome. We evaluate the Fidelity, Coherence, and Accuracy of the explanations obtained with CHIME with respect to two state-of-the-art Computer Vision models trained on real-world image data sets. We found evidence that the explanations reflect the causal concepts tied to a model's prediction, both in terms of causal strength and accuracy.@en

Users of sociotechnical systems often have no way to independently verify whether the system output which they use to make decisions is correct; they are epistemically dependent on the system. We argue that this leads to problems when the system is wrong, namely to bad decisions and violations of the norm of practical reasoning. To prevent this from occurring we suggest the implementation of defeaters: information that a system is unreliable in a specific case (undercutting defeat) or independent information that the output is wrong (rebutting defeat). Practically, we suggest to design defeaters based on the different ways in which a system might produce erroneous outputs, and analyse this suggestion with a case study of the risk classification algorithm used by the Dutch tax agency.@en

With recent advances in explainable artificial intelligence (XAI), researchers have started to pay attention to concept-level explanations, which explain model predictions with a high level of abstraction. However, such explanations may be difficult to digest for laypeople due to the potential knowledge gap and the concomitant cognitive load. Inspired by recent work, we argue that analogy-based explanations composed of commonsense knowledge may be a potential solution to tackle this issue. In this paper, we propose analogical inference as a bridge to help end-users leverage their commonsense knowledge to better understand the concept-level explanations. Specifically, we design an effective analogy-based explanation generation method and collect 600 analogy-based explanations from 100 crowd workers. Furthermore, we propose a set of structured dimensions for the qualitative assessment of analogy-based explanations and conduct an empirical evaluation of the generated analogies with experts. Our findings reveal significant positive correlations between the qualitative dimensions of analogies and the perceived helpfulness of analogy-based explanations. These insights can inform the design of future methods for the generation of effective analogy-based explanations. We also find that the understanding of commonsense explanations varies with the experience of the recipient user, which points out the need for further work on personalization when leveraging commonsense explanations. @en

Explainable artificial intelligence (XAI) aims to help people understand black box algorithms, particularly of their outputs. But what are these explanations and when is one explanation better than another? The manipulationist definition of explanation from the philosophy of science offers good answers to these questions, holding that an explanation consists of a generalization that shows what happens in counterfactual cases. Furthermore, when it comes to explanatory depth this account holds that a generalization that has more abstract variables, is broader in scope and/or more accurate is better. By applying these definitions and contrasting them with alternative definitions in the XAI literature I hope to help clarify what a good explanation is for AI.@en

In recent years philosophers have used results from cognitive science to formulate epistemologies of arithmetic (e.g. Giaquinto in J Philos 98(1):5–18, 2001). Such epistemologies have, however, been criticised, e.g. by Azzouni (Talking about nothing: numbers, hallucinations and fictions, Oxford University Press, 2010), for interpreting the capacities found by cognitive science in an overly numerical way. I offer an alternative framework for the way these psychological processes can be combined, forming the basis for an epistemology for arithmetic. The resulting framework avoids assigning numerical content to the Approximate Number System and Object Tracking System, two systems that have so far been the basis of epistemologies of arithmetic informed by cognitive science. The resulting account is, however, only a framework for an epistemology: in the final part of the paper I argue that it is compatible with both platonist and nominalist views of numbers by fitting it into an epistemology for ante rem structuralism and one for fictionalism. Unsurprisingly, cognitive science does not settle the debate between these positions in the philosophy of mathematics, but I it can be used to refine existing epistemologies and restrict our focus to the capacities that cognitive science has found to underly our mathematical knowledge.@en

Clarke and Beck argue that the ANS doesn't represent non-numerical magnitudes because of its second-order character. A sensory integration mechanism can explain this character as well, provided the dumbbell studies involve interference from systems that segment by objects such as the Object Tracking System. Although currently equal hypotheses, I point to several ways the two can be distinguished. @en