In this paper, I develop a normative framework for evaluating the use of new ICTs in the realm of religious faith. To this end, in Sect. 2, I argue for a partial account of religious faith, according to which typical religious faith requires certain skills and entails religious responsibility. Section 3 examines two specific challenges arising from the use of advanced ICTs for religious purposes, focusing on the recent debate regarding the authenticity of Islamic prayer apps and the challenges associated with so-called religious robots. Finally, in Sect. 4, based on the suggested account of religious faith in Sect. 2, I introduce the problems of ICT-driven religious deskilling and religious irresponsibility, which help explain important aspects of our intuitions regarding the considered challenges. The upshot is the general recommendation that policymakers, designers, researchers, and users of religious ICTs must remain sensitive to the risks of religious deskilling and religious irresponsibility.
Predictive processing casts perception, action, and learning as probabilistic inference. This paper asks when a Bayesian description identifies a mechanism. I distinguish computational description, algorithm, and physical implementation. Mechanism requires a causal mapping from physical transitions to inferential roles; behavioural fit cannot establish it. Priors range over model-supplied alternatives; observation does not reveal the organism’s individuation. Bayesian algorithms may be high-dimensional, continuous, nonparametric, and representation-growing. Every specified model has a fixed reachable closure; neither closure nor size decides mechanism. The question is which differences a proposed controller state treats as irrelevant. If histories assigned to that state respond differently to intervention because of timing, phase, contact, or field structure, the mapping omits part of the mechanism. A prospectively richer Bayesian or hybrid account may include those differences. High-dimensional systems and flexible model families often resist binary falsification when an experiment’s projection loses distinctions required by the claim. I therefore propose convergent tests of specified internal, coupled, and hybrid mappings. Results reject only a mapping within a declared task, scale, horizon, and margin; post-hoc revision creates a new hypothesis. The resource-bounded coupling hypothesis predicts that human-level transfer on contact-rich tasks depends on distinctions retained in organism–environment relations. An adaptive internal controller counts against it when transfer has no preregistered resource disadvantage, relational-history effects vanish in state-collision tests, the relational block adds no meaningful intervention-predictive value, and internal-realiser interventions have distinctive predicted effects. Existing evidence has not established a Bayesian mechanism for unconstrained embodied behaviour.
This paper examines the implications of Bedau’s notion of weak emergence when combined with both a hierarchical ontological model and Wolfram’s principle of computational equivalence. I argue that if a system exhibits weak emergence, the very fact of weak emergence of some of its states must itself be a weakly emergent state. Interpreted through Wolfram’s principle of computational equivalence, this implies that a system with weakly emergent states would need to ‘decide’ whether its own dynamical trajectories are incompressible—which contradicts with the formal undecidability of the incompressibility predicate. I term this aporetic conclusion the paradox of weak emergence. After evaluating possible responses, I contend that Bedau’s approach exemplifies what I call the derivational account of metaphysical relations, a framework that I will show to be highly problematic insofar as it implies a mismatch between the intrinsic limitations of formal derivations and the explanatory ambitions of metaphysical notions.
The interaction between Artificial Intelligence (AI) technologies and Human Rights Law (HRL) is gaining increasing attention in the literature. Scholars analyse both positive and negative interactions, with risks and harms bearing a stronger emphasis. Amid growing contributions to the field, a methodical synthesis of the literature is missing. This article fills this gap. We begin by mapping the interaction between AI and HRL, focusing on the types of AI applications, the impacted human rights, the relevant jurisdictions, and the involved actors. Then, we develop an interpretative (re)construction of the impact of AI on HRL and discuss how this impact affects the HRL mechanisms. Our mapping analysis reveals common themes that dominate academic discussions, including accountability, fair process, information provision and protection, and effective enforcement. In the final part, we explore the scope and nature of the directions of future AI HRL literature based on this review’s identification of gaps, blind spots and newly emerging or persistent ambiguities within the literature. In that respect, four key areas emerge as priorities for consideration: the role of private actors, the use of HR impact assessment, the adoption of relational theories, and the potential redefinition of the concept of harm, including through refocusing on vulnerability as a legal concept. Addressing these areas is crucial to strengthening the framework for mitigating AI-related risks to human rights.
With recent advances in generative AI models and their adoption by many commercial users, there has been wider discussion of the possibility that artificial general intelligence (AGI) will emerge from these systems in the near future. This discussion is muddled somewhat by conflicting conceptions of AGI and what successful trials imply. I argue for the necessity of certain social features in any system that can be ascribed general intelligence. Much of the discussion thus turns to theory of mind and attempts to model this with generative AI. These features are particularly important for a type of epistemic responsibility that has proven especially thorny for generative AI models, and I argue that the current generation of systems will not scale towards success on these fronts. While this is not an a priori dismissal of the very possibility of artificial intelligence, it implies a need to radically rethink future approaches.
Biological and artificial agents operating in complex environments have to leverage environmental structures to accomplish vital tasks. Recent research across a variety of domains—from the study of animal and human behaviour in different developmental periods and for different tasks, to computational studies of learning—has unveiled many ways in which structures are processed. This gave rise to a burgeoning field of study—structure learning. However, the diversity of phenomena studied, and the different aims and focuses of the researchers, have led to ambiguity and limited consensus on the nature of structure learning and its underlying mechanisms. In this paper we provide a synopsis of illustrative examples of structure learning, introduce the Active Inference Framework (AIF) with a focus on Structure Learning, and discuss points of contact between the two. The Active Inference Framework provides a mechanistic theory which distinguishes three levels of learning: Active Inference, Parametric Learning, and Bayesian Model Selection (a.k.a., Structure Learning), a method for the comparison and selection of models based on model evidence. We argue that when formalised under the Active Inference Framework, Structure Learning provides not only an underlying computational mechanism with aims of ecological validity, but also provides features relevant to computational accounts of structure learning more generally. The unifying aspect of the AIF in terms of having a single objective function for optimising behaviour should not be confounded with the exclusivity of this framework. The integration with other computational accounts is advised.
This article investigates the prospect of aligning AI agents with human autonomy. We show that doing so is a challenge: while there is general agreement on the concept’s broad contours, its components can be specified in a number of conflicting ways, making it difficult to decide how an autonomy-enhancing agent should act. In response, we introduce three strategies for dealing with disagreements about interpretations of autonomy. Each offers a coherent approach that captures a different set of considerations, with its own strength and trade-offs. The liberal approach is committed to user integrity and promotes human autonomy by acting strictly on what a person says they want, without trying to change their mind. The capability-boosting approach is built around the idea of flourishing and promotes human autonomy by empowering the user to act on their goals. The meta-autonomy approach prioritizes choice by giving the user the kind of autonomy they want in relation to their interactions with an AI agent. By highlighting the multifaceted nature of human autonomy, this article contributes to recent efforts seeking to identify and operationalize targets of AI value alignment. It also illustrates the challenges and trade-offs that can arise not just between—but even within—the values selected as targets for alignment and offers strategies to navigate them.
Common-sense learning and reasoning is a landmark of human-like intelligence. While classical-AI expert systems could convince at it in only narrow domains, contemporary deep-neural network models surprise with rapid performance improvements across many domains. At the same time, Bayesian Intuitive Theories have been influential in cognitive science as formal accounts of rational learning and reasoning. This paper targets Bayesian Intuitive Theories insofar as they rely on inferential-role semantics for conceptual content, to critically evaluate the promises and limits of this influential approach at mediating a theory of human-like common sense. I argue that both insights from deep-learning models and from Bayesian intuitive theories (insofar as they rely on inferential-role semantics for conceptual content) are insufficient to capture what seems to be centrally important in human-like common-sense learning and reasoning: Not just its internal consistency, but also its inherent relationship to the outside world. To address this challenge, I propose a situated revision of Bayesian Intuitive Theories that preserves the epistemic standards of rational reasoning while grounding domain-structured semantic content in ecological perception. By focusing on the role of semantic structure in embodied intervention, this proposal shows why common-sense reasoning inherently relates to the outside world. Consequently, machine-learning models mediate a theory of human-like common-sense only if the inferences that they implement satisfy both semantic and (inter)active constraints.
Synthetic data are increasingly used across the sciences and industry and are widely regarded as a solution to data problems such as privacy, bias mitigation, and data insufficiency. At the same time, the “promises” of synthetic data have faced substantial scholarly critique. We argue that much of the disagreement surrounding the usefulness of synthetic data stems from a lack of clarity about what synthetic data actually are and what they can do. In this paper, we examine several conceptualizations of synthetic data and propose a refined concept of synthesized data that more accurately captures the phenomenon commonly referred to as synthetic data. We argue that synthetic data cannot be conceptualized in contrast with other types of data. Moreover, there is no meaningful difference between synthetic data and data. We further prove this argument through an Iceberg model of the process of data synthetization and utilization, which addresses the “submerged” data risks such as bias, privacy, and validity, thus ensuring the usefulness of synthesized data. Our approach is grounded in a relational view of data, which links their intended purpose with their method of generation. We further elaborate this connection through concrete examples. In conclusion, we demonstrate the complexity of synthesized data in relation to their utility, which often remains hidden in current discourse on synthetic data.
Retrieval-Augmented Generation (RAG) systems increasingly operate not only as tools for retrieving and synthesizing information, but also as agents that can invoke external functions, modify digital environments, and execute tasks across software systems. This development raises a specific normative problem: the point at which a model’s output ceases to be merely informational and becomes an executable intervention in the world. Building on existing work in Responsible AI, accountability, and human oversight, this paper argues that tool-calling architectures place particular pressure on these frameworks because they can fuse retrieval, reasoning, and action within a single operational pipeline. To clarify this transition, the paper develops a threshold account of performative action based on four criteria: causal efficacy, autonomy, irreversibility, and moral salience. It then examines the normative consequences of crossing this boundary, showing how executable outputs can fragment responsibility, weaken effective oversight, and produce miscalibrated trust in hybrid human-AI systems. In response, the paper proposes three governance-by-design heuristics for tool-calling environments: epistemic traceability, operational reversibility, and normative containment. Together, these mechanisms aim to make actionable systems more answerable by preserving visibility into how actions arise, by creating limited opportunities for interruption or correction, and by bounding the scope of permissible delegation. The paper concludes that reclaiming the boundary between knowing and doing is essential not to restrict intelligence, but to preserve the conditions under which increasingly capable AI systems remain governable within human practices of authorization, accountability, and repair.
As artificial intelligence systems acquire increasing autonomy, the question of how humans can maintain meaningful control over their behavior becomes urgent. This question is especially pressing for agentic AI: systems built on foundation models that coordinate networks of autonomous agents to pursue complex goals with minimal human oversight. While the engineering literature has developed several mechanisms for constraining such systems, it has largely treated controllability as a purely causal property, the ability to physically intervene in a system’s operation, without engaging the normative dimensions that determine whether such intervention is meaningful. Drawing on the distinction between causal and normative control in the philosophy of action, this paper presents a structured conceptual survey of controllability mechanisms in agentic AI. We organize the literature around four paradigms: constraints and guardrails, adaptive control via reinforcement learning, agent-in-the-loop oversight, and human-in-the-loop oversight. For each, we analyze its technical foundations, its philosophical assumptions about agency and authority, and its persistent limitations. Our analysis reveals unresolved tensions across these paradigms, particularly between the scalability of automated oversight and the legitimacy of human judgment, between design-time constraint specification and runtime adaptability, and between individual agent control and system-level governance. We argue that meaningful controllability requires integrating causal intervention mechanisms with normative frameworks that address value alignment, epistemic transparency, and distributed responsibility. We conclude by proposing directions for research that bridges engineering controllability and philosophically grounded human oversight.
Concerns about the toxicity of the internet and social media are widespread. Many reasons have been offered for why social media increases online toxicity, from bad algorithms to the rise of online trolling, but perhaps the most prominent explanation has been the negative effects of social media on empathy. In this paper, we offer a framework of digital empathy by first outlining a descriptive account of digital empathy, encompassing both affective and cognitive components. By “digital empathy”, we mean empathy mediated through digital environments. We also explore the normative impacts of platforms on digital empathy and offer four main categories of moral assessment: (1) morally problematic one-off failures of digital empathy; (2) aggregate harms of failures of digital empathy; (3) impacts of recommendation systems on digital empathy by amplifying, obscuring, framing, and rushing targets of empathy; and (4) moral deskilling of empathy due to widespread use of social media. We also critically explore harm-based justifications for moderating content to explicitly govern empathetic relationships between users of a platform. We demonstrate how discussions of digital empathy can inform and enhance broader debates about the ethics of online content moderation, thereby further illuminating the moral, social, and political implications of social media’s role in facilitating user interactions.
While it is well-known that AI systems might bring about unfair social impacts by influencing social schemas, much attention has been paid to instances where the content presented by AI systems explicitly demeans marginalized groups or reinforces problematic stereotypes. This paper urges critical scrutiny to be paid to instances that shape social schemas through subtler manners. Drawing from recent philosophical discussions on the politics of artifacts, we argue that many existing AI systems should be identified as what Liao and Huebner called oppressive things when they function to manifest oppressive normality. We first categorize three different ways that AI systems could function to manifest oppressive normality and argue that those seemingly innocuous or even beneficial for the oppressed group might still be oppressive. Even though oppressiveness is a matter of degree, we further identify three features of AI systems that make their oppressive impacts more concerning. We end by discussing potential responses to oppressive AI systems and urge remedies that go beyond fixing the unjust outcomes but also challenge the unjust power hierarchies of oppression.
This paper introduces and defends the concept of “deliberative injustice” for hermeneutic technology assessment. Attempts to ethically assess emerging technologies—for example, quantum technologies—often centre on the imagined futures associated with these technologies. Competing sociotechnical “visions” shape public debate and guide decisions about funding, research, and regulation. The paper argues that these visions are not idle speculation or mere prediction, but powerful tools that shape which (and whose) futures are seen as possible, desirable, or legitimate. However, not everyone has equal power to wield these tools. Using David Lewis’s idea of the “conversational score,” the paper shows how imbalances of social power can lead to some voices being ignored or dismissed in the context of societal deliberation about emerging technologies. Those whose voices are ignored or dismissed are wronged not only because they may suffer consequences in the future, but because they are both degraded and damaged in their capacity as deliberative agents in the present. By defining deliberative injustice in terms of whose contributions to public debate receive “uptake” and on what basis, the paper shifts attention away from future outcomes and toward present-day structures of power, offering a new, deontological approach to the ethics of emerging technologies. It illustrates the manifestation of deliberative injustice in current practices of societal deliberation about emerging technologies, with a particular emphasis on quantum technologies. It concludes with a call for ethicists of technology to pay greater attention to deliberative injustice in the future.
The meta-inductive justification of induction is usually regarded as a social learning strategy. But, pre-theoretically, induction can be justified even for an isolated thinker, incapable of, or unwilling to engage in, any social interactions. This paper presents reinforcement learning as a natural meta-inductive strategy for such isolated thinkers. The meta-inductive reinforcement learner learns to choose an optimal prediction method among the available methods by applying trial-and-error learning. We use computer simulations to show that, under certain plausible assumptions, trial-and-error learning indeed leads the thinker to favor induction over non-inductive predictive methods. Under further plausible assumptions, the meta-inductive thinker is even seen to outperform all object-level methods. The results are shown to hold in both stationary and dynamic environments.
Machine learning models of misinformation detection are increasingly being used and yet their risks have been under analyzed, requiring insights from philosophy of science and political philosophy. A taxonomy of types of risks and simple models estimating them is provided that is sensitive to the value-laden features of judgments of misinformation and weighted by a potential item of misinformation’s impact on respective stakeholders. Failing to do so is incompatible with a variety of civil virtues that are desired by not only liberal democratic societies but authoritarian ones as well, suggesting the general applicability of this taxonomy of risks to contemporary and future societies.
This paper critically explores the assumption that fairness, here broadly understood as in automated decision support systems, as inherently ethical equates to ethical algorithms. While fairness has become a key benchmark for ethical assessment in AI applications, this work argues that fairness does not always guarantee ethical results. By drawing on consequentialist, deontological, and care ethics perspectives, the paper highlights that uncritical reliance on fairness may overlook individual uniqueness and reinforce existing power structures, as algorithms tend to reduce complex human identities to fixed categories. Additionally, the impartiality often valued in the debate over AI systems may be ethically unsuitable in contexts where recognizing individual singularity and situational nuances is essential. Finally, the paper advocates for a transparent evaluation of the ethical frameworks that shape AI systems, emphasizing the risk of a surreptitious introduction of ethical visions and purposes. The concept of “fair- washing” is adopted, a practice that risks portraying fairness in automated systems as inherently ethical because of fair, sidestepping responsibility to define and scrutinize the goals of these technologies.
This paper discusses the impact of generative AI (genAI) on the legal profession and the importance of virtue ethics in addressing the epistemic and ethical challenges for legal practice posed by genAI technology. It traces the development of legal genAI, and argues that its functionality currently leaves it ethically lacking (relative to the dream of AI as a substitutive technology) in terms of its competence, integrity, and compliance with ethical standards. The paper further argues that genAI also fails to contribute to the more normative, non-principlist, ethical expectations that we associate with excellent practice. While it can perform well on certain well-defined tasks, it struggles with the nuances of real-world legal problem-solving, particularly complex relational problems where practical wisdom and contextual understanding play a significant role, over and above technical domain expertise. This is because the tools currently lack the capacity for moral discernment and the ability to autonomously reason either from first principles, or from virtue-based reasoning, both of which are significant for professional decision-making. It suggests that, without more, the deep deployment of genAI in legal practice may thus lead to a diminution of practical wisdom, rather than its augmentation or enhancement. The paper concludes with some recommendations on how the legal ecosystem might manage genAI adoption in ways that better enable AI tools and the lawyers using them to augment rather than diminish virtue.
LLMs have been shown to induce representations akin to real-world information structures. According to the standard account, this is because LLMs model the human mind, and minds model the world. Thus, LLMs and minds share properties that ultimately converge upon similar structures. We contribute to this discourse in two ways: a) We argue for an alternative account: LLMs model minds, and minds model worlds, but minds also model LLMs, which in turn extend the capacities of minds. In sum, the world-mind-LLM system constitutes a three-player game, which b) we show does not necessarily converge. We explore the epistemological consequences of non-convergence for world-mind-model systems, introducing the concept of epistemic drift. We offer a new explanatory account of epistemic drift as a social phenomenon, and explore the potential safety concerns of epistemic drift with the introduction of LLMs. Finally, we contribute to existing discourse surrounding LLM agency, arguing that agency is not found in the content of LLM output, but in the way they functionally interact with minds.
Do different artificial neural networks (ANNs) “think” alike? And can their shared internal representations be cataloged? This paper explores the latter questions by assessing the feasibility of creating an Artificial Neural Repository (ANR)—a library of “universal” neural representations that are reusable across different ANN architectures. To assess the potential of such a repository, we examine three hypotheses: H1 (Uniqueness), suggesting representations are architecture-specific; H2 (Identity), proposing representations are identical across architectures; and H3 (Similarity), asserting partial representational overlap. Reviewing existing theoretical arguments and empirical findings—including studies employing Representational Similarity Analysis (RSA), Canonical Correlation Analysis (CCA), and Centered Kernel Alignment (CKA)—we find the strongest support for H3. This indicates substantial but not total representational overlap among diverse ANNs. We conclude that while a fully universal repository remains challenging, an ANR is viable if it includes mechanisms to translate between architecture-specific representations.
