Filippos Stamatiou

Postmortem avatars (PMAs) — AI systems that simulate a deceased person by being fine-tuned on data they generated or that was generated about them — have attracted growing scholarly attention, yet their potential role in clinical settings remains largely unexplored. This paper examines the ethics of deploying PMAs as therapeutic tools in grief therapy. Drawing on the dual-process model of grief, the theory of continuing bonds, and the philosophical framework of fictionalism, we propose two potential therapeutic applications: incorporating PMAs into established imaginal exercises such as the empty chair exercise, and treating the process of PMA creation as an art-therapeutic exercise in its own right. We consider five ethical objections to these applications and argue that none constitute knock-down arguments against therapeutic use, particularly given the risk-mitigating role of the clinical context. We conclude by identifying outstanding governance challenges and calling for empirical research, without which neither the promise nor the dangers of therapeutic PMAs can be adequately assessed.

Should developers be held responsible for the predictions of their neural networks—and if not, does that introduce a responsibility gap? The claim that neural networks introduce a responsibility gap has seen significant pushback, with philosophers arguing that the gap can be bridged, or did not exist in the first place. We show how the responsibility gap turns on whether we can distinguish between foreseeable and unforeseeable neural network predictions. Empirical facts about neural networks tell us we cannot, which seems to force developers to either assume full responsibility or no responsibility at all, introducing a responsibility gap—unless, of course, the same empirical facts hold true of humans, in which case there is no gap, but the trouble is simply with the classical notion of responsibility. We revisit and revise Mele’s Zygote, as well as the famous Palsgraf case, and argue that in fact, what complicates responsibility assignment for neural networks also complicates responsibility assignment for humans, and humans seem to confront us with the same all-or-nothing dilemma. Thus, we agree there is no technology-induced responsibility gap (there was no gap in the first place), but for slightly different reasons than our predecessors.

Mechanistic interpretability (MI) aims to explain how neural networks work by uncovering their underlying mechanisms. As the field grows in influence, it is increasingly important to examine not just models themselves, but the assumptions, concepts and explanatory strategies implicit in MI research. We argue that mechanistic interpretability needs philosophy as an ongoing partner in clarifying its concepts, refining its methods, and navigating the epistemic and ethical complexities of interpreting AI systems. There is significant unrealised potential for progress in MI to be gained through deeper engagement with philosophers and philosophical frameworks. Taking three open problems from the MI literature as examples, this paper illustrates the value philosophy can add to MI research, and outlines a path toward deeper interdisciplinary dialogue.

Knowledge claims are abundant in the literature on large language models (LLMs); but can we say that GPT-4 truly "knows" the Earth is round? To address this question, we review standard definitions of knowledge in epistemology and we formalize interpretations applicable to LLMs. In doing so, we identify inconsistencies and gaps in how current NLP research conceptualizes knowledge with respect to epistemological frameworks. Additionally, we conduct a survey of 100 professional philosophers and computer scientists to compare their preferences in knowledge definitions and their views on whether LLMs can really be said to know. Finally, we suggest evaluation protocols for testing knowledge in accordance to the most relevant definitions.

The widespread use of voice-assisted applications using artificial intelligence raises questions about the dynamics of trust and reliance on these systems. While users often rely on these applications for help, instances where users face unforeseen risks and heightened challenges have sparked conversations about the importance of fostering trustworthy artificial intelligence. In this paper, we argue that the prevailing narrative of trust and trustworthiness in relation to artificial intelligence, particularly voice assistants, is misconstrued and fundamentally misplaced. Drawing on insights from philosophy and artificial intelligence literature, we contend that artificial intelligence systems do not meet the criteria for participating in a relationship of trust with human users. Instead, a narrative of reliance is more appropriate. However, we investigate the matter further, to explore why the trust/trustworthiness narrative persists, focusing on the unique social dynamics of interactions with voice assistants. We identify factors such as diverse modalities and complexity, social aspects of voice assistants, and issues of uncertainty, assertiveness, and transparency as contributors to the trust narrative. By disentangling these factors, we shed light on the complexities of human-computer interactions and offer insights into the implications for our relationship with artificial intelligence. We advocate for a nuanced understanding of trust and reliance in artificial intelligence systems and provide suggestions for addressing the challenges posed by the dominance of the trust/trustworthiness narrative.

Control is typically accepted as a necessary condition for moral responsibility. Thus, humans are morally responsible for their actions only if we can realise the right kind of control. Are there good reasons to think that humans can psychologically realise control? This paper is an attempt to address this question by establishing choice and agenthood as separate but interconnected aspects of control. I consider two challenges to the claim that humans can realise the kind of control required for moral responsibility. First, an empirical challenge from cognitive neuroscience provides a familiar way to argue against the realisation of the choice aspect by human psychology. Second, a more formidable conceptual challenge to the aspect of agenthood presents us with scepticism about the kind of explanations that psychology can provide. The second challenge suggests that, in psychological accounts of choice, the agent disappears. Drawing on recent empirical models of cognitive control and philosophical accounts of agency, I conclude that the psychological explanation of choice is consistent with the aspect of agenthood being realised by human psychology.

This work develops a philosophically credible and psychologically realisable account of control that is necessary for moral responsibility. We live, think, and act in an environment of subjective uncertainty and limited information. As a result, our decisions and actions are influenced by factors beyond our control. Our ability to act freely is restricted by uncertainty, ignorance, and luck. Through three articles, I develop a naturalistic theory of control for action as a process of error minimisation that extends over time. Thus conceived, control can serve to minimise the influence of luck on action and enable freedom in uncertainty. In Article 1, Thor Grünbaum and I argue for a psychologically plausible account of the kind of control that is necessary for moral responsibility. We begin by establishing the relationship between agentive contribution and responsibility-level control. One way to determine whether one the right kind and degree of control to be morally responsible is to track one’s degree of contribution to an action. However, a psychologically plausible account of control in terms of agentive contribution may seem to contradict the types of functional-mechanistic explanations used in cognitive science. In cognitive psychology and cognitive neuroscience, personal-level capacities are explained by a set of sub-personal mechanisms. Often, such explanations leave no room for a contribution by the agent. By integrating insights from theories of cognitive control and incorporating them into a philosophical account of intentions, we propose a way of thinking about the distribution of cognitive control resources as something the agent does. Article 2 argues that a classic argument concerning luck, originally aimed at libertarianism, generalises beyond any specific theory of free will, and regardless of whether determinism or indeterminism is true. I call this mental luck. Because we all make decisions under conditions of relative uncertainty and limited information, it is possible for an agent to make decisions that are contrary to their own motivation. In such situations, it may be a matter of luck whether the agent makes the right decision. Mentally lucky decisions are not rationally governed by attitudes within the agent's perspective, and thus, may be indistinguishable from unlucky ones. From the perspective of the agent, such decisions may resemble the outcome of a lottery. Therefore, mental luck poses a challenge to most prominent theories of free action and moral responsibility. Finally, article 3 engages with the issue of resultant luck, namely luck in how things turn out. Resultant luck raises a challenge for theories of moral responsibility because its existence suggests that one may be responsible for factors beyond one’s control. Prominent responses to resultant luck led to a choice between internalism and scepticism. I argue that familiar cases of resultant luck are based on the assumption that actions are events. Instead, I propose an alternative ontology of action as an ongoing goal-directed process with a many-shots structure. Described this way, cases of resultant luck are not representative of ordinary action. The proposal of action as a many-shots process is consistent with predictive coding, a cognitive architecture which centres around error minimisation. Under this framework, cases of resultant more luck are no longer failures of action, but rather anticipated errors to be settled within the ordinary process of action.