Artificial Comparative Psychology

Raphaël Millière (Macquarie University)

Artificial intelligence has made tremendous progress in recent years, bolstered by the remarkable effectiveness of deep neural networks trained on large datasets. Yet despite impressive behavioral similarities between humans and AI on a wide array of tasks, from language use to complex reasoning and problem-solving, the field remains deeply divided on whether these systems can be meaningfully said to possess human-like cognitive capacities.    To move beyond this impasse, this paper proposes a framework for artificial comparative psychology grounded in the conceptual tools of the philosophy of cognitive science. This framework has three key tenets: (1) Embracing a bidirectional relationship between performance and competence, recognizing that neither impressive successes nor failures are sufficient to establish or rule out particular cognitive capacities in AI; (2) Developing sophisticated behavioral testing paradigms that probe for specific competences in a targeted way, supported by an explicit cognitive ontology specifying the processes involved, without anthropocentric assumptions; (3) Leveraging interventional methods inspired by neuroscience to uncover the representations and computations actually used by AI systems to perform targeted tasks, identifying human-recognizable building blocks while also allowing for the possibility of unhuman-like solutions.   Central to this framework is a focus on mid-level computational components (e.g. algorithmic circuits for variable binding, rule induction, analogical mapping, etc.) that bridge low-level neural mechanisms and high-level cognitive capacities. Tracing the extent to which AI relies on such components can illuminate long-standing debates about machine-human comparisons, as well as the diverse possible paths to intelligent behavior.   Ultimately, artificial comparative psychology aims to overcome the false dichotomy between ascribing human-like capacities to AI and denying any meaningful overlap. By triangulating behavioral and mechanistic evidence from well-designed experimental paradigms to map the space of possible explanations of performance patterns, this approach can build a more nuanced understanding of cognition in all its forms—biological and artificial alike.