Diego Morales

It is sometimes assumed that large language models (LLMs) know language, or for example that they know that Paris is the capital of France. But what—if anything—do LLMs actually know? In this paper, I argue that LLMs can acquire tacit knowledge as defined by Martin Davies (1990). Whereas Davies himself denies that neural networks can acquire tacit knowledge, I demonstrate that certain architectural features of LLMs satisfy the constraints of semantic description, syntactic structure, and causal systematicity. Thus, tacit knowledge may serve as a conceptual framework for describing, explaining, and intervening on LLMs and their behavior.

Through analyzing examples from the COVID-19 pandemic and legislative strategies for the responsible use of AI, we argue that trustworthy science is an intrinsic value with which every good research practice should align. This means that science has to be conducted in a responsible way. Still, this is a necessary but not sufficient condition for building long-lasting trust in science. The social component plays a significant role in this process. Education is important because it increases the scientific literacy of laypeople and the responsible applications of technology by experts. Moreover, when applying scientific research to a larger population one has to keep in mind the individual, cultural and national context of the measures in question. In order to achieve cross-national trust in science, social and natural sciences have to act in synergy with educational institutions over a long period of time. While the normative questions have to be answered by philosophers and policymakers. Especially in the case of AI, science has to be designed for the benefit of humans and for their use, constituting the so-called human-centered application of AI.

In this article, I investigate whether algorithmic fairness is an inconsistent concept (the inconsistency thesis). Drawing on the work of Kevin Scharp, inconsistent concepts can apply and disapply at the same time (2.). It is shown that paradigmatic issues of algorithmic fairness fit this description (3.). Similarities and differences to received views (4.) and alternatives to the inconsistency thesis are considered (5.). Suggestions are articulated on how the inconsistency thesis might hold ground nevertheless, or at the very least denotes a distinctive option in argumentative space whose status and implications merit further evaluation.

The debate on concept possession in animals has moved at a very abstract level, with scant detailed consideration of case studies in animal behaviour. In this paper, we go against this trend by examining a specific prey defence mechanism, thanatosis or death-feigning, in order to determine what it can tell us about the minds of the predators it targets. We argue that thanatosis gives us evidence of conceptual abilities in predators. In particular, we defend that the best available explanation for the evolution of thanatosis comes from positing that the predators whose cognition acted as selection pressure and shaped thanatosis displays satisfy the concept possession requirements of distance, inferentiality, aspectuality, generality, and a minimal semantic net. Moreover, we argue that the semantic content of the concept targeted by thanatosis likely corresponds to a minimal concept of death. We end on a methodological note, by showing how the consideration of case studies can help move this philosophical debate forward, while also generating concrete empirical predictions that can be put to the test, thus favouring interdisciplinary collaborations.

One of the biggest contemporary challenges in education is the appropriate application of advanced digital solutions. If properly implemented, AI could benefit students, opening the door for personalized study programs. However, we need to ensure that AI in classrooms is used responsibly and that it does not pose a threat to students in any way. More specifically, we need to preserve the moral and epistemic values we wish to pass on to future generations and ensure the inclusion of underprivileged students. Moreover, there is a huge problem of different accessibility to technology and representation in the educational material between the Global South and the Global North. This can cause both global epistemic and moral injustice. On the other hand, AI has the potential to increase global justice in education. Therefore, it is of the utmost importance to use it in accordance with the values of equity, fairness, and inclusion that secure global justice for underprivileged groups.

Recent work in philosophy of technology has come to bear on the question of responsibility gaps. Some authors argue that the increase in the autonomous capabilities of decision-making systems makes it impossible to properly attribute responsibility for AI-based outcomes. In this article we argue that one important, and often neglected, feature of recent debates on responsibility gaps is how this debate maps on to old debates in responsibility theory. More specifically, we suggest that one of the key questions that is still at issue is the significance of the reactive attitudes, and how these ought to feature in our theorizing about responsibility. We will therefore provide a new descriptive categorization of different perspectives with respect to responsibility gaps. Such reflection can provide analytical clarity about what is at stake between the various interlocutors in this debate. The main upshot of our account is the articulation of a way to frame this ‘new’ debate by drawing on the rich intellectual history of ‘old’ concepts. By regarding the question of responsibility gaps as being concerned with questions of metaphysical priority, we see that the problem of these gaps lies not in any advanced technology, but rather in how we think about responsibility.

We argue that the later Wittgenstein’s philosophy of language and mathematics, substantially focused on rule-following, is relevant to understand and improve on the Artificial Intelligence (AI) alignment problem: his discussions on the categories that influence alignment between humans can inform about the categories that should be controlled to improve on the alignment problem when creating large data sets to be used by supervised and unsupervised learning algorithms, as well as when introducing hard coded guardrails for AI models. We cast these considerations in a model of human–human and human–machine alignment and sketch basic alignment strategies based on these categories and further reflections on rule-following like the notion of meaning as use. To sustain the validity of these considerations, we also show that successful techniques employed by AI safety researchers to better align new AI systems with our human goals are congruent with the stipulations that we derive from the later Wittgenstein’s philosophy. However, their application may benefit from the added specificities and stipulations of our framework: it extends on the current efforts and provides further, specific AI alignment techniques. Thus, we argue that the categories of the model and the core alignment strategies presented in this work can inform further AI alignment techniques.

The extent to which normative cognition varies across cultures has implications for a number of important philosophical questions. This chapter examines several striking commonalities and differences in normative cognition across cultures. We focus on cross-cultural commonality and difference in norm typologies (especially the moral-conventional distinction); the externalization of norms; which aspects of life are normativized; and some of the concepts and principles associated with the normative domain. We argue that the distinction between moral and conventional norms is probably not universal and that several concepts and principles that have been suggested as candidate universals may in fact be culturally local. We nevertheless isolate two possible universals: the phenomenon of externalizing norms and the normativization of a preferred class of subject matters.

In this paper, I introduce a “promises and perils” framework for understanding the “soft” impacts of emerging technology, and argue for a eudaimonic conception of well-being. This eudaimonic conception of well-being, however, presupposes that we have something like stable character traits. I therefore defend this view from the “situationist challenge” and show that instead of viewing this challenge as a threat to well-being, we can incorporate it into how we think about living well with technology. Human beings are susceptible to situational influences and are often unaware of the ways that their social and technological environment influence not only their ability to do well, but even their ability to know whether they are doing well. Any theory that attempts to describe what it means for us to be doing well, then, needs to take these contextual features into account and bake them into a theory of human flourishing. By paying careful attention to these contextual factors, we can design systems that promote human flourishing.

Background Big data and AI applications now play a major role in many health contexts. Much research has already been conducted on ethical and social challenges associated with these technologies. Likewise, there are already some studies that investigate empirically which values and attitudes play a role in connection with their design and implementation. What is still in its infancy, however, is the comparative investigation of the perspectives of different stakeholders. Methods To explore this issue in a multi-faceted manner, we conducted semi-structured interviews as well as focus group discussions with patients and clinicians. These empirical methods were used to gather interviewee’s views on the opportunities and challenges of medical AI and other data-intensive applications. Results Different clinician and patient groups are exposed to medical AI to differing degrees. Interviewees expect and demand that the purposes of data processing accord with patient preferences, and that data are put to effective use to generate social value. One central result is the shared tendency of clinicians and patients to maintain individualistic ascriptions of responsibility for clinical outcomes. Conclusions Medical AI and the proliferation of data with import for health-related inferences shape and partially reconfigure stakeholder expectations of how these technologies relate to the decision-making of human agents. Intuitions about individual responsibility for clinical outcomes could eventually be disrupted by the increasing sophistication of data-intensive and AI-driven clinical tools. Besides individual responsibility, systemic governance will be key to promote alignment with stakeholder expectations in AI-driven and data-intensive health settings.

In this reply to van de Poel’s (Philosophy & Technology, 35(3), 82, 2022) commentary on O’Neill (Philosophy & Technology, 35(79), 2022), I discuss two worries about the general contextual integrity approach to evaluating technological change. First, I address van de Poel’s concern that the general contextual integrity approach will not supply the right guidance in cases where morally problematic technological change poses no threat to contextual integrity. Second, I elaborate on how the approach supplies mechanisms for balancing caution with the need for change.

Throughout scientific history, there have been cases of mainstream science dismissing novel ideas of less prominent researchers. Nowadays, many researchers with different social and academic backgrounds, origins and gender identities work together on topics of crucial importance. Still, it is questionable whether the privileged groups consider the views of underprivileged colleagues with sufficient attention. To profit from the diversity of thoughts, the scientific community first has to be open to minority viewpoints and epistemically include them in mainstream research. Moreover, the idea of inclusive science poses stronger requirements than the paradigm of open science. We argue that the concept of integration of different opinions is insufficient because the process of integration assumes adjusting oneself to the majority view and fitting into the dominant paradigm while contributing only with smaller amendments. Epistemic inclusion, on the other hand, means dynamically changing the research paradigm during the interaction with diverse methods and hypotheses. The process of inclusion preserves marginalized views and increases epistemic justice.

Artificial intelligence is currently changing many areas of society. Especially in health, where critical decisions are made, questions of control must be renegotiated: who is in control when an automated system makes clinically relevant decisions? Increasingly, the concept of meaningful human control (MHC) is being invoked for this purpose. However, it is unclear exactly how this concept is to be understood in health. Through a systematic review, we present the current state of the concept of MHC in health. The results show that there is not yet a robust MHC concept for health. We propose a broader understanding of MHC along three strands of action: enabling, exercising and evaluating control. Taking into account these strands of action and the established rules and processes in the different health sectors, the MHC concept needs to be further developed to avoid falling into two gaps, which we have described as theoretical and labelling gaps.