World Models and the Emergence of a “First-Person” Perspective in an AI

Lenore Blum discusses how a first-person perspective can emerge in artificial systems using the Conscious Turing Machine formal model.

Lenore Blum (Carnegie Mellon University) simons.berkeley.edu/talks/lenore-blum-carnegie-mellon-univer… Topics in Intelligence: World Models and Social Reasoning What would it mean for an artificial system to possess a “first-person” perspective? Consider an embodied AI architecture that interacts with its environment through sensors and actuators, develops world models shaped by its history of interaction, and integrates multimodal information through global workspace–like dynamics. In such a system, internal models evolve through continual cycles of prediction, feedback, and learning. These models are therefore not static representations but arise from temporally extended coupling between agent and environment. Building on joint work with Manuel and Avrim Blum on the Conscious Turing Machine (Blum, Blum and Blum, forthcoming, 2026), a formal machine model that incorporates the above features, I propose that a “first-person” perspective can be understood as an emergent structural feature of agent-centered world models. In this view, the agent’s world model constitutes an operational “first-person” perspective: an organized point of view grounded in the system’s own interactional history. This proposal suggests that several structural features emphasized in neurophenomenology (Varela 1996)—including enaction, temporality, and integrated perspective—may be realizable in non-biological architectures. Examining how such features could arise in artificial systems may help clarify how world models support coherent agent-centered representations of their environments.