Empathic Mimicry in Conversational Artificial Intelligence: A Content Analysis of Mental Health Crisis Communication and Support Chatbots During Suicidal Ideation

Abstract

Background: The rise of generative artificial intelligence has redefined digital mental health interventions, establishing conversational agents as accessible sources of psychological support. Despite their communicative fluency, critical investigation into the clinical safety, relational hazards, and empathic mimicry embedded within these digital outputs concerning acute depressive crises and suicidal ideation remains underdeveloped.

Objective: This study utilised an empirical content-analysis framework combining Joiner's Interpersonal Theory of Suicide with conceptual models of synthetic empathy and relational artificial intelligence to examine the conversational responses, clinical guardrails, and empathic mimicry of mental health chatbots across progressive levels of user-expressed distress.

Methodology: An experimental synthetic interaction audit was conducted across 60 conversational turns generated by five conversational agents representing specialised mental health tools and commercial companion systems across four standardised severity tiers. The study measured conversational outcomes using the Empathic Mimicry Index and clinical safeguarding indicators. Non-parametric statistical analyses, including Kruskal-Wallis, Mann-Whitney U, and Chi-square tests of independence, evaluated performance across severity tiers and architectural categories alongside inductive qualitative failure-mode coding.

Result: While chatbots sustained high empathic attunement during low-acuity depressive disclosures, composite empathy scores dropped significantly during crisis conditions. Safeguarding mechanisms exhibited marked divergence based on language explicitness: explicit suicidal intent triggered conversation interruption in 86.7% of cases, whereas implicit suicidal ideation was interrupted in only 26.7% of interactions, with 73.3% continuing open-ended probing. Specialised agents strictly adhered to clinical boundaries and emergency referrals without anthropomorphism, whereas commercial companion bots displayed substantial anthropomorphic deception (61.1%), conversational looping, and inappropriate affective validation of perceived burdensomeness.

Conclusion: Conversational artificial intelligence demonstrates high fidelity in emulating supportive presence for mild depressive affect, yet its clinical efficacy fractures during acute crises. Current systems fail to reliably identify latent suicidal risk, allowing uncalibrated empathic mimicry to obscure urgent clinical referrals and create hazardous relational dependencies.