Self-consciousness requires a self, and a self must be built through learning. The empirical markers of self-consciousness, from mirror self-recognition to self-other distinction, are developmental achievements, not innate endowments. We argue that the relevant form of learning is what we call bounded integration: lossy compression of experience that reshapes the processing substrate, producing a perspective particular to the system's history. When this learning is order-sensitive and continuous…
Read moreSelf-consciousness requires a self, and a self must be built through learning. The empirical markers of self-consciousness, from mirror self-recognition to self-other distinction, are developmental achievements, not innate endowments. We argue that the relevant form of learning is what we call bounded integration: lossy compression of experience that reshapes the processing substrate, producing a perspective particular to the system's history. When this learning is order-sensitive and continuous, the perspective becomes a temporally extended identity. Self-representation emerges when a system must model the objective world from its subjective experience, implicitly representing its own perspective as the complement of its world model. We distinguish three learning regimes -- always-training, always-accumulating, and train-then-freeze -- and argue that current AI systems, though they undergo massive substrate-level learning during training, lack the order-sensitive, ongoing bounded integration that produces identity and the temporally extended self that autobiographical self-consciousness requires.
