Benjamin Henke

The question of whether large language models (LLMs) can exhibit moral capabilities is of growing interest and urgency, as these systems are deployed in sensitive roles such as companionship and medical advising, and will increasingly be tasked with making decisions and taking actions on behalf of humans. These trends require moving beyond evaluating for mere moral performance, the ability to produce morally appropriate outputs, to evaluating for moral competence, the ability to produce morally appropriate outputs based on morally relevant considerations. Assessing moral competence is critical for predicting future model behaviour, establishing appropriate public trust and justifying moral attributions. However, both the unique architectures of LLMs and the complexity of morality itself introduce fundamental challenges. Here we identify three such challenges: the facsimile problem, whereby models may imitate reasoning without genuine understanding; moral multidimensionality, whereby moral decisions are influenced by a range of context-sensitive relevant moral and non-moral considerations; and moral pluralism, which demands a new standard for globally deployed artificial intelligence. We provide a roadmap for tackling these challenges, advocating for a suite of adversarial and confirmatory evaluations that will enable us to work towards a more scientifically grounded understanding and, in turn, a more responsible attribution of moral competence to LLMs.

I motivate the systematic study of artificial affect by arguing that its possibility has significant intellectual, practical, and ethical implications. I argue the extant consciousness-centric approaches fail to address these implications, while behavioral approaches to AI sentience fail to address the inherent gaming problem in behavioral approaches to AI mentality. In place of these approaches, I propose a functional approach focusing on the roles that affective states play within a cognitive economy. I exemplify this approach with the case of bodily pain by examining its constituent sensory, evaluative, and motivational functions. This analysis establishes the tractability of a functional approach to artificial pain while also highlighting several outstanding questions that must be answered before we can settle the question whether artificial pain is possible in near-future AI systems.

When viewing a circular coin rotated in depth, it fills an elliptical region of the distal scene. For some, this appears to generate a two-fold experience, in which one sees the coin as simultaneously circular (in light of its 3D shape) and elliptical (in light of its 2D ‘perspectival shape’ or ‘p-shape’). An energetic philosophical debate asks whether the latter p-shapes are genuinely presented in perceptual experience (as ‘perspectivalists’ argue) or if, instead, this appearance is somehow derived or inferred from experience (as ‘anti-perspectivalists’ argue). This debate, however, has largely turned on introspection. In a recent study, Morales, Bax, and Firestone (2020) aim to provide the first empirical test of this question. They asked subjects to find an elliptical coin seen face-on from a search array that also included a circular coin seen either face-on or at an angle. They found that subjects reacted more slowly when the distracting circle was seen at an angle, such that it’s p-shape matched that of the target ellipse. From this, they concluded that the similar p-shape between the ellipse and circle constituted a phenomenal similarity between the two, and thus that perspectivalism is true. We show that these results can also be explained by pre-attentive guidance by unconscious representations (in what follows, just “unconscious pre-attentive guidance”) and that this explanation is at least as plausible as one from phenomenal similarity. Thus, we conclude that the experiment does not support perspectivalism over anti-perspectivalism.

I argue that beliefs based on irresponsibly formed experiences — whose causes were not appropriately regulated by the subject — are doxastically unjustified. Only this position, I claim, accounts for the higher epistemic standard required of perceptual experts. Section I defends this standard and applies it to a pair of cases in which either an expert umpire or a complete novice judge a force play in baseball. I argue that when the latter, but not the former, fails to follow rules about perceiving force plays, their resulting belief is justified. Section II shows that this difference can be explained by the fact that the novice, but not the expert, formed her experience responsibly. Section III shows that alternative explanations of the expert’s unjustified belief — from defeat, reliability, and inference — fail. Section IV shows that the epistemic relevance of responsible experience formation has broad implications for the epistemology of perceptual beliefs.

According to what I’ll call the ‘two visual systems account’ (TWO-SYSTEMS), the visual system is divided into two independent sub- systems, a ventral system implementing ‘vision for perception’ and a dorsal system implementing ‘vision for action’ (Milner and Goodale, 2006). TWO-SYSTEMS is widely discussed in philosophy due to the counter-intuitive role that it posits for conscious experience in the control of actions. However, recent evidence undermines the model’s core tenets: it no longer appears that the ventral and dorsal streams constitute isolated processing systems, and there is now evidence for the involvement of both streams in conscious experience and online motor control. I articulate a new ‘direct dorsal control account’ (DORSAL-CONTROL), show that it is immune to three empirical challenges facing TWO-SYSTEMS, and show that it nonetheless has similarly significant implications for the perceiving mind. Section 1 articulates TWO-SYSTEMS and its philosophical implications. Section 2 describes three empirical challenges to TWO-SYSTEMS. Section 3 articulates DORSAL-CONTROL and shows that it is immune to the three challenges. Section 4 articulates DORSAL-CONTROL’s philosophical implications.