Redundancy-as-masking: formalizing the Artificial Age Score (AAS) to model memory aging in generative AI.

Introduction: Artificial intelligence can exhibit aging-like patterns not as a function of chronological time, but through systematic asymmetries in output-level observable memory performance under different context-persistence conditions. In large language models, semantic cues, such as the name of the day, may remain stable across sessions, whereas episodic details, such as the sequential progression of experiment numbers, may collapse when conversational context is reset. To capture this phenomenon, this study introduces the Artificial Age Score (AAS), a log-scaled, entropy-informed metric of memory age derived from observable recall behavior and defined purely at the output level, without access to internal latent-state representations. Methods: The AAS was formulated and formally analyzed under mild, model-agnostic assumptions. Its properties were established in terms of well-definedness, boundedness, and monotonicity. In the present study, the framework was evaluated in a 25-day bilingual recall protocol using ChatGPT-5.0 across stateless and persistent interaction phases. Although the broader AAS framework includes a Redundancy-as-Masking formulation, redundancy was not explicitly estimated here; all reported values were therefore computed under a redundancy-neutral setting (R = 0), yielding conservative upper bounds. Results: During persistent sessions, the model consistently recalled both semantic and episodic details, driving the AAS toward its theoretical minimum and indicating behavioral youth in recall. In contrast, when sessions were reset, the model preserved semantic consistency but failed to maintain episodic continuity, resulting in a sharp increase in the AAS and signaling an aging-like behavioral signature of continuity loss in recall behavior. Discussion: These findings support the utility of AAS as a theoretically grounded, task-independent diagnostic tool for evaluating memory degradation in artificial systems. The results are interpreted behaviorally and do not constitute evidence about internal memory mechanisms or latent-state dynamics. The empirical validation reported here is limited to this specific model version and protocol, and applications to other architectures or training regimes require revalidation. The study builds on foundational concepts from von Neumann's work on automata, Shannon's theories of information and redundancy, and Turing's behavioral approach to intelligence.