Towards Language Models for AI Mental Health Assistant Design

Abstract

Advances in Artificial Intelligence (AI) Language Models (LMs) and their new applications are continuously reported. LMs, respond to plain text that is readily human interpretable. Based on these human-like responses, hopes are created for achieving human-level performance for various language tasks soon. This paper discusses challenges in applying current LMs to design an AI Mental Health Assistant. The results of experiments are encouraging but show that significant research and development efforts are necessary to reach the practical usefulness of AI. We discuss that chaining multiple LMs might be needed to filter or post-process the results. Additionally, the models themselves might need to go through enhanced training with a more significant emphasis on empathy, ethics, and moral standards, especially in the very sensitive mental health area.