Decoding dynamic emotional valence in GenAI interactions: insights from covariate-dependent Markov chains.

Area of Science:

• Human-Computer Interaction
• Affective Computing
• Computational Social Science

Background:

• User emotional states dynamically change during human-GenAI interactions.
• Limited research exists on how AI output quality impacts these emotional transitions.

Purpose of the Study:

• To model the association between AI response quality and user emotional valence state transitions.
• To investigate the impact of AI output quality on user emotions in academic writing tasks.

Main Methods:

• A covariate-dependent Markov chain model was proposed.
• An experiment involving AI-assisted academic writing was conducted with university students.
• 886 interaction sequences were analyzed to track emotional valence shifts.

Main Results:

• A polarization effect was observed: high-quality AI responses reinforced positive emotions, while low-quality responses led to emotional deterioration.
• The study analyzed group differences based on user emotional stability.
• AI response quality significantly influences the direction and stability of user emotional valence.

Conclusions:

• This research offers a novel perspective on emotional dynamics within human-GenAI dialogues.
• Findings provide empirical evidence for developing emotionally adaptive Generative AI systems.
• Understanding emotional transitions is crucial for improving user experience in AI-assisted tasks.