EQ2AI: Quantifying How Conversational AI Shapes Emotional Continuity of Self

Conversational AI increasingly incorporates affective capabilities, enabling systems to recognise and respond to users' emotional states across multi-turn interactions. Although prior work demonstrates that conversational and psychological framing can influence affect and behaviour, such effects are often evaluated through individual responses or aggregate pre and post outcomes, leaving less understood how affective influence unfolds across successive turns. This paper presents EQ2AI, a prompt-driven, psychotherapy-informed conversational agent that varies intervention tone (positive, neutral, or anti-therapeutic) and psychotherapy modality across turns without model retraining. A RoBERTa-based classifier fine-tuned on GoEmotions estimates user affect at each turn, while Plutchik's oppositional structure informs intervention direction. We evaluate 366 simulated conversational trajectories (2,298 turn-level observations) alongside a real-user study ($n=17$), treating the per-turn affective trajectory as the primary unit of analysis. We find that intervention tone is a stronger determinant of trajectory than psychotherapy modality, producing a distinct \textit{amplification--stability tradeoff}: positive tone produces the largest cumulative affective movement but greater volatility, whereas neutral tone produces smaller but more stable trajectories. These results show that affective influence in conversational AI can be characterised not only by whether affect changes, but by the magnitude, stability, volatility, and temporal shape of that change. EQ2AI provides a systematic research instrument for studying and comparing prompt-driven affective interventions across multi-turn dialogue, while the correspondence between simulated and real-user trajectories motivates larger-scale validation of simulation-based approaches to affective interaction research.