PhilPapers preprint · 2026-04-08
The empirical markers of selfhood, from mirror self-recognition to self-other distinction, are developmental achievements rather than innate endowments. Selfhood requires learning of two kinds. First, bounded integration, the lossy and order-sensitive learning that reshapes how future experience is processed, produces a perspective particular to the system's history. When this integration is continual, the perspective becomes a temporally extended identity. Second, learning a model of the world produces self-representation, where a system modelling the objective world from its subjective experience implicitly represents its own perspective as the complement of the model. We distinguish three forms of learning: always-learning, always-accumulating, and train-then-freeze. Current AI systems, despite massive learning during training, are predominantly always-accumulating during deployment. Their weights are frozen, leaving them with at most a weak, static self.
