Daria Zakharova

Recent philosophical work has proposed a “lightweight account” of introspection, on which a system introspects when it represents its own mental states in a way that makes these states accessible for guiding behavior. This approach has informed empirical proposals for detecting introspective abilities in current LLMs. I argue that this lightweight account fails to capture what is essential to genuine introspection. This paper proceeds through three increasingly concessive but individually sufficient challenges to the attribution of introspective abilities to LLMs. First, LLMs lack the persistent subject necessary for genuine introspection, as current models lack the psychological continuity relationship needed for self-knowledge. Second, LLM self-reports violate the immunity to error through misidentification that characterizes genuine introspection, because they are based on public textual information that could equally support judgments about another system’s states. Third, by centering on functional self-monitoring and behavioral control, the lightweight account fails to distinguish introspection from ubiquitous self-regulatory processes in complex systems.

This dissertation investigates conceptual and epistemological issues in the philosophy of artificial intelligence, arthropod cognition and scientific practice. It presents four individual research papers as its four main chapters, which focus on introspection in large language models (chapter one), LLM self-reports resulting from specific training as evidence about any possibly conscious internal states in them (chapter two), competing explanations of intelligent behaviour in Portia jumping spiders (chapter three), and the epistemology of AlphaFold – a case of an opaque deep learning system that produces scientific knowledge (chapter four). The overarching question across this dissertation’s individual investigations is how our existing concepts and methodologies for attributing cognitive capacities, consciousness or knowledge to artificial systems, phylogenetically distant animals, and new technology-driven scientific practice hold up when pressured by new evidence and new technological capabilities. The cases I examine are those of biological and artificial systems that are substantially opaque to our understanding, or where the evidence about their internal organisation is at present significantly limited. The central task this dissertation assumes is to investigate the limits of behavioural inference in such cases and to examine what is at stake in getting any such attributions right. To this end, I provide a philosophical analysis of the methodology, the epistemological challenges, and the conceptual foundations of the respective specific cases, each holding broader philosophical implications for its respective domain. In investigating these cases, this dissertation contributes to the philosophical literature across several subdisciplines: the philosophy of artificial intelligence, the philosophy of comparative cognition, and the epistemology of artificial intelligence in science. I examine the issues in these respective subdisciplines via a philosophy of mind lens, broadly construed.

The success of AlphaFold, an AI system that predicts protein structures, poses a challenge for traditional understanding of scientific knowledge. It operates opaquely, generating predictions without revealing the underlying principles behind its predictive success. Moreover, the predictions are largely not empirically tested but are taken at face value for further modelling purposes (e.g. in drug discovery) where experimentation takes place much further down the line. The paper presents a trilemma regarding the epistemology of AlphaFold, whereby we are forced to reject one of 3 claims: (1) AlphaFold produces scientific knowledge; (2) Predictions alone are not scientific knowledge unless derivable from established scientific principles; and (3) Scientific knowledge cannot be strongly opaque. The paper argues that AlphaFold's predictions function as scientific knowledge due to their trustworthiness and functional integration into scientific practice. The paper addresses the key challenge of strong opacity by drawing on Alexander Bird's functionalist account of scientific knowledge as irreducibly social, and advances the position against individual knowledge being necessary for the production of scientific knowledge. It argues that the implicit principles used by AlphaFold satisfy the conditions for scientific knowledge, despite their opacity. Scientific knowledge can be strongly opaque to humans, as long as it is properly functionally integrated into the collective scientific enterprise.