International periodic scientific journal

Abstract

Mental health conditions affect over 970 million people globally, and fewer than half of them receive any form of care [1]. AI chatbots and conversational agents promise to close this gap through scalable, anonymous, low-cost support. The promise has not held. Published reviews from 2024 and 2025 show that most AI mental health tools lose the bulk of their users within 30 days [2, 6]. We analyzed fifteen peer-reviewed sources across human-computer interaction, clinical psychology, and AI ethics to trace the cause. What emerged was not a technology deficit but a design one: developers have optimized for computational accuracy (correct sentiment labels, appropriate response selection, fast generation speed) while treating relational quality as optional. Users leave because the system does not feel like it is listening. To address this, we propose the Human-Centered Emotional AI (HCEAI) model. It comprises five design principles: Emotional Resonance, Contextual Sensitivity, Authentic Presence, Adaptive Responsiveness, and Ethical Transparency, each paired with a buildable engineering component such as a three-tier memory architecture and an adaptive feedback loop. We apply the model to grief and bereavement support, the domain with the highest ethical stakes, and report preliminary pilot results from the OPV Systems digital grief support platform. The model awaits validation through controlled trials with diverse populations.