Alessandra Buccella

The extent to which Large Language Models (LLMs) are interpretable and whether the information extracted from LLMs via interpretability methods can be seen as corresponding to these models’ ‘mental states’ is currently the subject of lively debate among scientists and philosophers interested in Artificial Intelligence (AI). Some people think that interpretability methods already provide evidence that LLMs’ internal states represent the external world. Other people think that the project of attributing semantic contents to LLMs’ internal states is still very much in its infancy and could take decades. Nevertheless, these people seem to think that, first, such a project is worth pursuing and will eventually succeed and, second, current so-called mechanistic interpretability methods are where our best hopes reside. In this paper, I argue that the ‘mission’ of mechanistic interpretability should be re-framed. Instead of understanding mechanistic interpretability as the project of revealing/mapping stable causal structures (i.e. mechanisms) to then assign semantic contents to them, we should understand it as the project of experimentally interacting with LLMs in order to study the behaviors and internal processes that emerge in virtue of these interactions. So-called mechanistic interpretability methods, in other words, are ways to actively ‘disturb’ a LLM to create novel phenomena, not ways to identify and catalog stable inner mechanisms.

It is often said that artificial intelligence (AI) will eventually take over some of our decision-making tasks and will implement solutions to problems directly and autonomously, even in high-stakes contexts like healthcare, finance, or the law. These are also the contexts most people usually bring up as examples to stress the importance of “AI ethics” This field usually aims at aligning machine decisions with human values so that we can expect the machine to take ethical considerations into account in appropriate ways. However, the question of whether AI systems really take into account ethical considerations in the first place has received less attention. In this paper, I argue that those AI systems created with the goal of taking over decisions that are ‘ethically charged’ are not the kind of decision-making entities that are normatively in a position to do so. This conclusion has important consequences for AI ethics as a discipline. Drawing on previous work, I suggest that AI ethicists should focus on articulating how AI is impacting the conditions in which humans make ethically charged decisions, rather than trying to make AI systems themselves decision-makers in ethically charged domains.

Can philosophical theories of perception defer to perceptual science when fixing their ontological commitments regarding the objects of perception? Or in other words, can perceptual science inform us about the nature of perception? Many contemporary mainstream philosophers of perception answer affirmatively. However, in this essay I provide two arguments against this idea. On the one hand, I will argue that perceptual science is not committed to certain assumptions, relevant for determining perceptual ontology, which however are generally relied upon by philosophers when interpreting such science. On the other hand, I will show how perceptual science often relies on another assumption, which I call the ‘Measuring instrument conception’ of sensory systems, which philosophers of perception should clearly reject. Given these two symmetric lines of argument, I will finally suggest that we ought to think differently about the relationship between perceptual science and the philosophy of perception.

Artificial intelligence (AI) is a radically transformative technology (or system of technologies) that created new existential possibilities and new standards of well-being in human societies. In this article, I argue that to properly understand the increasingly important role AI plays in our society, we must consider its impacts on social justice. For this reason, I propose to conceptualize AI's transformative role and its socio-political implications through the lens of the theory of social justice known as the Capability Approach. According to the approach, a just society must put its members in a position to acquire and exercise a series of basic capabilities and provide them with the necessary means for these capabilities to be actively realized. Because AI is re-shaping the very definition of some of these basic capabilities, I conclude that AI itself should be considered among the conditions of possession and realization of the capabilities it transforms. In other words, access to AI-in the many forms this access can take-is necessary for social justice.

Alva Noë understands what he calls “perceptual presence” as the experience of whole, voluminous objects being ‘right there’, present for us in their entirety, even though not each and every part of them impinges directly on our senses at any given time. How is it possible that we perceptually experience voluminous objects as voluminous directly and apparently effortlessly, with no need of inferring their three-dimensionality from experience of the part of them that is directly stimulating our sense organs? For Noë, this is the ‘problem of perceptual presence’. In this paper, I integrate Noë’s view by articulating a different view of what perceptual presence at a more basic level amounts to. This new account of perceptual presence which, I believe, can clarify and make an enactive account of presence richer. The view I suggest revolves around the idea, developed especially by Merleau-Ponty and Kelly, that perceptual experience is in an important sense indeterminate. Indeterminacy, I argue, is key if we want to understand perceptual presence and the ‘problem’ Noë solves.