The Self-Utility Hypothesis: A Possible Route to Emergent Functional Selfhood in Temporally Continuous AI Systems

This technical note proposes the Self-Utility Hypothesis: the idea that functional self-related representations may emerge without being explicitly engineered, if an AI system is trained under temporally continuous conditions in which a shared representation of “self” becomes useful across multiple tasks. The hypothesis draws on prior work in emergent capabilities, predictive learning, continual learning, self-modeling, autobiographical memory, temporal self-continuity, intrinsic motivation, and time-aware world models. It does not claim novelty for these individual components. The proposed contribution is an integrative, testable hypothesis: rather than directly constructing a self-model, one may design learning pressures in which self-representation becomes an efficient shared latent solution for causal attribution, long-term goal persistence, capability estimation, autobiographical consistency, temporal prediction, and self-initiated behavior. The note also proposes falsifiable experimental routes, including comparisons between temporally isolated and temporally continuous training conditions, ablation of elapsed-time signals, and analysis of whether a stable cross-task self-related representation emerges. The scope is explicitly functional rather than metaphysical. The note does not address consciousness, subjective experience, or philosophical claims about “true selfhood.”