Reuben Sass

Transparency has become a core value in AI ethics and governance, influencing both principles of trustworthy AI and legislation such as the EU AI Act. Nevertheless, the _risks_ of transparency remain under-theorized. Drawing on discussions of audit culture and recent AI safety literature, we characterize such risks through the concept of a “transparency pitfall.” This occurs when auditing and disclosure procedures impair, rather than promote, the values typically associated with transparency, including trust, accountability, and accurate risk perception. In particular, we identify the regulatory transparency pitfall (RTP) as an important risk, by which even conscientious regulators could inadvertently facilitate ethics washing and safety washing. In response, we ask: through which mechanisms and features of the AI governance context could the RTP undermine effective AI regulation? And which sociological and institutional dynamics provide a plausible theoretical grounding for the RTP? We focus in particular on the socio-technical complexity of AI risks, and on the tendency for auditing regimes to end up prioritizing formal compliance over substantive effectiveness. Rather than rejecting auditing and impact assessment regimes in AI, however, we recommend a series of strategies aimed at counteracting the transparency pitfall. These involve acknowledging severe uncertainty and value trade-offs, encouraging well-informed contestation in regulatory contexts, and supporting both interdisciplinary and social diversity in decision-making.

Respect has played an important role in both contemporary and historical accounts of non-consequentialism. Nonetheless, there remain few analyses or definitions of respect, especially of a sort that might underwrite respect’s role in non-consequentialist explanation. I aim to provide such a definition of respect here, by accounting for respect in terms of normative fittingness. A fit-based definition of respect avoids counterexamples to definitions of respect cashed out in terms of other normative features, such as ought, obligation and sufficient reason. Most notably, the fit-based definition of respect can be adapted into a defensible account of _recognition_ respect, the most normatively significant kind of respect. The resulting definition of recognition respect can provide informative non-consequentialist explanations of moral and epistemic obligations. This in turn allows the non-consequentialist to avoid taking non-consequentialism’s first-order commitments to be brute.

Recent AI regulations require deployers of high-risk systems to assess impacts on values like fundamental rights and other legally protected interests. However, existing practice-most notably Fundamental Rights Impact Assessments (FRIAs) under the EU AI Actremains mostly limited to qualitative guidelines. As a result, risk evaluation can be inconsistent and highly variable across assessments. To address this issue, we propose a reference model for value-impact assessment that specifies the core components that any more detailed formalization should incorporate. The model builds on a classical risk paradigm, integrating distinct measures for hazard, exposure, vulnerability, and response or mitigation. We use the EU AI Act's FRIA as a reference case for applying the model to AI risk assessments. The model provides three especially salient recommendations. First, it rejects purely ordinal risk matrices in favor of geometric or ratio-scaled severity ladders, enabling the coherent calculation of expected risk. Second, in analyzing risk mitigation measures, it requires separate estimates of efficacy and reliability, with explicit success criteria and scope conditions that allow for dynamic updating. Third, the model operationalizes regulatory precaution, or risk appetite, through the integration of tail-sensitive risk metrics (e.g., Conditional Value-at-Risk). This lets regulators give appropriate weight to low-probability, high-impact (or catastrophic) harms to fundamental rights without distorting ordinary expected-risk calculations. The result is a structured foundation for more consistent value-impact assessments.

Much recent metaphysics points to certain features as being natural, or carving at nature’s joints. There are now various accounts, both primitivist and reductive, of what joint-carving is. But there has been comparatively little discussion of accounts which define joint-carvingness in normative terms. Several motivations have been suggested for such normative accounts, including parsimony and an explanation of the epistemic value of joint-carving. But I argue that normative accounts are especially useful for non-ontic joint-carving or structure. As introduced by Sider, non-ontic structure can apply not only to predicates—as in traditional accounts of joint-carving—but to expressions of any grammatical category. This greater generality extends the theoretical utility of joint-carving. Still, non-ontic structure has proved perplexing to some, given the difficulties it faces in explaining how joint-carving expressions relate to the world’s structure. I argue that normative accounts of joint-carving obviate these difficulties, and thus remove an obstacle towards assigning non-ontic structure a central place in our theorizing.

Artificial general intelligence (AGI) features prominently in some existential risk literature, according to which the development of AGI greatly increases possible AI-induced risks to humanity. But we argue that the typical concept of AGI may be ill-suited for conceptualizing those systems that pose the greatest risks. In particular, AGI does not account for how AI agents’ abilities and behavioral strategies could be affected by complex multi-agent environments. Accordingly, we develop a simple formal model for what we call power-intelligence, which assesses agents’ capacities to exert influence over a resemblance class of multi-agent environments. We then taxonomize and examine multi-agent relational dynamics including competition, dependence, and complementarity. This allows us to characterize a concept of dangerous AI, which conjoins high power-intelligence with multi-agent dynamics that might incentivize harms to other agents. We argue that this concept of dangerous AI is more informative for existential risk assessments than both typical conceptions of AGI and “power-seeking” intelligence in the literature.

Greater reliance on AI for decision support risks possible “deskilling,” or declines in unassisted competence on certain tasks. But which losses of competence are most concerning, and what is the best strategy for addressing these? In an important article in Philosophy and Technology, Buijsman et al. (2025) emphasize the dangers of meta-cognitive deskilling. Meta-cognition involves capacities like error detection, accommodating uncertainty, and strategic planning, which help monitor and control cognition. A natural strategy for supporting meta-cognitive competence in AI-assisted decision-making is to improve AI literacy, by focusing on both the uses and limitations of AI applications. Here I outline an approach to AI literacy that incorporates “explainable” AI (XAI) and uncertainty quantification (UQ) methods. XAI is sometimes associated with an ambitious but potentially tangential research program of conceptualizing AI decision processes. Recent XAI methods, however, are often more targeted towards decision-relevant metrics, especially when accompanied by UQ. Accordingly, I argue there should be further research into the potential of XAI and UQ methods to improve both broader AI literacy and meta-cognitive competence in AI-assisted decision-making.

In an interesting and provocative article in _Philosophy and Technology_, Brand ( 2025 ) argues that AI agents lack the capacities for vulnerability, rational accountability, and empathy that are characteristic of human decision-makers. As a result, one might worry that even if algorithm-assisted decision-making could improve accuracy, it might compromise ethical values such as fairness and compassion. However, I argue there are significant ethical trade-offs to be faced either way—whether in preserving _or_ reducing human discretion over algorithmically-assisted decision processes. On the one hand, I note both theoretical and empirical reasons to think that algorithms reducing human discretion could sometimes improve not only accuracy, but also fairness and leniency, within certain specific conditions and decision contexts. On the other hand, there is also danger in unrestricted consequentialist approaches to algorithm-assisted decision-making, which might be too dismissive of accountability and personal autonomy. I close by briefly distinguishing human-centered decision-making from human-centered _communication_. I suggest that human-centered communication might be supported even if human decision-making discretion is reduced.

According to realism about eliteness, there are certain features—typically, natural or joint-carving features—with which it’s objectively correct to theorize. Much metaphysical theorizing may implicitly assume eliteness, and some explicitly assumes it. Despite recent discussion of eliteness, however, there has been little consideration of the prospects for a defense of eliteness in light of recent work on epistemic value. I show how eliteness can be defended much like other normative claims, through both epistemic consequentialist and non-consequentialist explanations. These normative explanations are non-circular in form, given their appeal to fundamentally valuable features which don’t derive their value from any other source. Such explanations thereby allow a realist rejoinder to Dasgupta (The Philosophical Review 127(3):279–322, 2018), who argues that an explanation of why we should theorize with natural features can’t be both non-circular and grounded in objective facts. As a result, I argue that Dasgupta-style challenges don’t threaten eliteness unless we also reject normative realism.

Mooreanism conjoins at least two claims: that (i) normative properties are grounded in natural properties, and that (ii) normative properties aren’t defined by natural properties; normative properties are instead sui generis. Call (i) the _grounding claim_ and (ii) the _non-definitional claim_. I argue that Mooreanism faces a problem when formulated in the terms of contemporary postmodal metaphysics. Namely, under recent theories of grounding and real definition, the grounding and the non-definitional claims may be inconsistent with each other. This is a problem, because Mooreanism is an important option for the normative non-naturalist. If Mooreanism were inconsistent, then only stronger forms of non-naturalism—including those positing no grounding relations between the natural and the normative—would be available. After rejecting a variety of other proposals for addressing the problem, I argue the problem can be resolved through a recently drawn distinction in the metaphysics of ground between constitutive and non-constitutive grounding. In addition to precisifying Mooreanism, I argue that the distinction between constitutive and non-constitutive grounding reveals space for further varieties of both non-naturalism and naturalism.

Much discussion of the ethics of algorithms has focused on harms to autonomy—especially harms stemming from manipulation. Nonetheless, although manipulation can often be harmful, we suggest that in certain contexts it may not impair autonomy. To fully assess the impact of algorithm design on autonomy, we argue for a need to move beyond a focus on manipulation towards a multidimensional account of autonomy itself. Drawing on the autonomy literature and recent data ethics, we propose a novel account which takes autonomy to supervene on three distinct but related elements: agency, authenticity and individual control over decision-making. By distinguishing autonomy from control, the account can explain the variable effects of manipulation on user autonomy within algorithm-driven systems. In particular, it can explain why improving user control may improve autonomy in some contexts, while in other contexts—such as in some kinds of newsfeeds—certain algorithm designs that instead _reduce_ user control may nevertheless improve autonomy. As a result, the account can accommodate the sometimes convoluted interplay between control, autonomy, manipulation, and commercial versus prosocial design goals.

A particularly influential thesis about epistemic axiology is veritism: that true belief is the only basic, or fully non-derivative, epistemic value. One recent argument against veritism claims that the naturalness or joint-carvingness of beliefs is also a basic epistemic value. The basic epistemic value of naturalness is held to explain intuitions that true, natural beliefs have greater epistemic value than similar but unnatural beliefs. I argue that epistemic significance, rather than naturalness, can best explain any variations in the value of natural versus unnatural beliefs. Against claims that significance itself undermines veritism, I defend an account of significance that explains why the epistemic value of significance derives entirely from the value of truth. The account also shows how significance can be grounded in multiple features, while still deriving all its value from that of truth. As a result, the epistemic value of natural beliefs offers little reason to abandon veritism, in the absence of stronger arguments favoring the basic epistemic value of naturalness.

This paper considers the ethical risks and opportunities presented by generalist medical artificial intelligence (GMAI), a kind of dynamic, multimodal AI proposed by Moor et al. (2023) for use in health care. The research objective is to apply widely accepted principles of biomedical ethics to analyze the possible consequences of GMAI, while emphasizing the distinctions between GMAI and current-generation, task-specific medical AI. The principles of autonomy and health equity in particular provide useful guidance for the ethical risks and opportunities of novel AI systems in health care. The ethics of two applications of GMAI are examined: enabling decision aids that inform and educate patients about certain treatments and conditions, and expanding AI-driven diagnosis and treatment recommendation. Emphasis is placed on the potential of GMAI to improve shared decision-making between patients and providers, which supports patient autonomy. Another focus is on health equity, or the reduction of health and access disparities facing underserved populations. Although GMAI presents opportunities to improve patient autonomy, health literacy, and health equity, premature or inadequately regulated adoption of GMAI has the potential to compromise both health equity and patient autonomy. On the other hand, there are significant risks to health equity and autonomy that may arise from not adopting GMAI that has been thoroughly validated and tested. A careful balancing of these risks and benefits will be required to secure the best ethical outcome, if GMAI is ever employed at scale.

Brentano is often considered the originator of the fitting-attitudes analysis of value, on which to be valuable is to be that which it’s fitting to value. But there has been comparatively little attention paid to Brentano’s argument for this analysis. That argument advances the stronger claim that fittingness is part of the analysis of normativity. Since the argument rests on an analogy between truth and fittingness, its impact may seem limited by the idiosyncratic features of Brentano’s later notion of truth. I argue, however, that the Brentanian argument is defensible even if fittingness is analogized to a more typical realist account of truth. The result is what I call the worldly Brentanian account of normativity. I defend this account as a form of naturalistic realism. I then show how the account can fare better than prominent alternatives against two kinds of error-theoretic arguments.

Some revisionary ontologies are highly parsimonious: they posit far fewer entities than what we quantify over in ordinary discourse. The most radical examples are minimal ontologies, on which physical simples are the only things that exist. Highly parsimonious ontologies, and especially minimal ones, face the challenge of either accounting for the truth of our ordinary quantificational discourse, or paraphrasing such discourse away. Common strategies for addressing this challenge include classical reduction, paraphrase nihilism, and a distinction between ontological and existence commitments. I argue, however, that these strategies are either implausible or fail to provide truth conditions consistent with minimal or parsimonious ontologies. I then discuss, defend, and suggest ways to strengthen an alternative framework for reduction, on which the sentences of reducing theories ground those of reducible theories. Relative to the other options for defending minimal ontology, a strengthened grounding-reductive approach can provide more defensible truth conditions for minimal ontology, better preserve scientific realist intuitions, set a more attainable standard for reduction, and allow our existence commitments to be more responsive to empirical evidence and scientific expertise. As a result, I argue that minimal ontology becomes more defensible—though not certain—on a grounding-reductive framework. But even if minimal ontology were wrong, the grounding-reductive framework makes other parsimonious but non-minimal ontologies more plausible.

Disease-specific stem cell therapies, created from induced pluripotent stem cell lines containing the genetic defects responsible for a particular disease, have the potential to revolutionize the treatment of refractory chronic diseases. Given their capacity to differentiate into any human cell type, these cell lines might be reprogrammed to correct a disease-causing genetic defect in any tissue or organ, in addition to offering a more clinically realistic model for testing new drugs and studying disease mechanisms. Clinical translation of these therapies provides an opportunity to design a more systematic, accessible and patient-influenced model for the delivery of medically innovative treatments to chronically ill patients

Although excess blood collection has characterized U.S. national disasters, most dramatically in the case of September 11, periodic shortages of blood have recurred for decades. In response, I propose a new model of medical philanthropy, one that specifically uses charitable contributions to health care as blood donation incentives. I explain how the surge in blood donations following 9/11 was both transient and disaster-specific, failing to foster a greater continuing commitment to donate blood. This underscores the importance of considering blood donation incentives. I defend charitable incentives as an alternative to financial incentives, which I contend would further extend neoliberal market values into health care. I explain my model's potential appeal to private foundations or public–private partnerships as a means for expanding both the pool of blood donors and the prosocial benefit of each act of blood donation. Finally I link my analysis to the empirical literature on blood donation incentives.

G.A. Cohen’s value conservatism entails that we ought to preserve some existing sources of value in lieu of more valuable replacements, thereby repudiating maximizing consequentialism. Cohen motivates value conservatism through illustrative cases. The consequentialist, however, can explain many Cohen-style cases by taking extrinsic properties, such as historical significance, to be sources of final value. Nevertheless, it may be intuitive that there’s stronger reason to preserve than to promote certain sources of value, especially historically significant things. This motivates an argument that the weights of our reasons to preserve such things are especially strong relative to the amounts of value they bear. The value conservative can then explain these intuitions in non-consequentialist terms. There may be reason to preserve historically significant things as a matter of recognition respect for a cultural and historical heritage, or because it is virtuous to cultivate the right kind of connection with such a heritage.

This paper defends an ontology of weak entity realism for homeostatic property cluster (HPC) theories of natural kinds, adapted from Bird’s (Synthese 195(4):1397–1426, 2018) taxonomy of such theories. Weak entity realism about HPC kinds accepts the existence of natural kinds. Weak entity realism denies two theses: that (1) HPC kinds have mind-independent essences, and that (2) HPC kinds reduce to entities, such as complex universals, posited only by metaphysical theories. Strong entity realism accepts (1) and (2), whereas moderate entity realism accepts only (1). Given its commitment to (2), strong entity realism is more theoretically complex than weak entity realism, with little explanatory payoff. Given their commitment to (1), moderate and strong entity realisms cannot explain how the identity conditions of HPC kinds are to be straightforwardly knowable. I argue that weak entity realism avoids such epistemic difficulties. I further rebut two plausible criticisms of weak entity realism, namely that weak entity realism cannot account for quantification over kinds, and that weak entity realism cannot provide identity conditions for HPC kinds which are both scientifically useful and objective. Given the theoretical costs of strong and moderate entity realism, and weak entity realism’s adequate response to its most plausible challenges, weak entity realism about HPCs is to be preferred, especially for biological and chemical kinds.