Can AI Express Empathy? Computational Insights from Mental Health Dialogues

Abstract The growing demand for mental health support has accelerated the adoption of AI-powered chatbots intended to alleviate gaps in access to care. The global AI mental health market is projected to reach 17.8 billion USD by 2030, yet it remains unclear whether AI-mediated therapeutic conversations differ fundamentally from human-delivered therapy in emotional expression and linguistic style. This uncertainty raises critical questions for clinical integration, user trust, and ethical deployment. This study investigated whether emotional expression patterns differ between AI mental health conversations and human therapy sessions. A total of 1,000 AI chatbot conversations and 1,000 human therapy dialogues were extracted from open-access datasets and analyzed using computational text analysis and machine learning. Emotional intensity was measured across five domains (anxiety, depression, loneliness, positive affect, anger), alongside sentiment polarity, self-disclosure markers, conversational structure, and linguistic features. Statistical tests used Benjamini–Hochberg FDR correction and effect size estimation. A Random Forest classifier with 5-fold cross-validation evaluated the distinguishability of AI and human conversations. Results showed that emotional expression was statistically indistinguishable across most domains. Machine learning classification yielded only 48.3% accuracy, equivalent to chance performance. Key distinguishing features were structural rather than emotional word count and sentiment accounted for 44% of predictive value. These findings suggest that users express emotions in comparable ways to AI and human therapists. AI chatbots should be positioned as accessibility tools that supplement but never replace clinical care.