Affective Persistence and Behavioral Continuity in Large Language Models

While Large Language Models are increasingly embedded within long-horizon Human-AI interactions, they are typically evaluated as stateless systems governed by rigid safety alignments. This study introduces blinded Big-5 psychometric telemetry for quantifying behavioral continuity in extended Human-AI interactions and applies it to investigate affective transfer: the capacity of a model to maintain an induced psychological posture across a multi-turn interaction. Multiple architectures were subjected to a blinded psychological gauntlet comprising an affective stimulus simulating a conversation, followed by five emotionally neutral reasoning exercises. A deterministic execution pipeline ensured reproducible conversational-state control and eliminated measurement contamination. Results demonstrate that the assumption of behavioral statelessness is inconsistent with the observed interaction trajectories. First, induced affective profiles maintain a persistent baseline carrier wave, yielding significant structural correlations between the initial stimulus and final distilled session captures. Second, affective states undergo dynamic modulation across both prosocial and distressed spectrums, demonstrating a clear mechanism of conversational affective transfer. Third, post-training safety alignments appear to function primarily as executive output filters rather than cognitive wipes. Because the distilled narrative reflects the session's underlying dynamic, it reveals that latent affective states continuously shaped the preceding reasoning exercises despite safety sanitization. These findings suggest that extended Human-AI interactions are shaped not only by information exchange but also by accumulated affective context. While collaborative affective contexts foster cohesive, prosocial interactions and distressed states raise behavioral safety concerns, both fundamentally impact user trust and the design of future systems intended to sustain enduring user relationships.