Virtual Agents for Alcohol Use Counseling: Exploring LLM-Powered Motivational Interviewing

Ian Steenstra,Farnaz Nouraei,Mehdi Arjmand,Timothy W. Bickmore

DOI: https://doi.org/10.1145/3652988.3673932

2024-07-11

Abstract:We introduce a novel application of large language models (LLMs) in developing a virtual counselor capable of conducting motivational interviewing (MI) for alcohol use counseling. Access to effective counseling remains limited, particularly for substance abuse, and virtual agents offer a promising solution by leveraging LLM capabilities to simulate nuanced communication techniques inherent in MI. Our approach combines prompt engineering and integration into a user-friendly virtual platform to facilitate realistic, empathetic interactions. We evaluate the effectiveness of our virtual agent through a series of studies focusing on replicating MI techniques and human counselor dialog. Initial findings suggest that our LLM-powered virtual agent matches human counselors' empathetic and adaptive conversational skills, presenting a significant step forward in virtual health counseling and providing insights into the design and implementation of LLM-based therapeutic interactions.

Human-Computer Interaction,Computation and Language

What problem does this paper attempt to address?

The problem this paper attempts to address is:

In alcohol abuse counseling, effective counseling services remain difficult to access, especially in the field of substance abuse. Virtual agents, by leveraging the capabilities of large language models (LLMs) to simulate Motivational Interviewing (MI) techniques, can offer a promising solution. This paper explores how to develop a virtual counselor capable of conducting Motivational Interviewing to improve the effectiveness of alcohol abuse counseling.

Specifically, the paper focuses on the following three research questions:
1. **What are the differences between human and LLM-generated counseling responses in terms of linguistic accuracy, safety, and adherence to MI principles?**
2. **To what extent can LLM-driven virtual agents effectively use MI elements to promote behavior change?**
3. **From the user's perspective, what are the advantages and limitations of LLM-driven virtual agents compared to human counselors?**

To answer these questions, the researchers developed a GPT-4-based virtual counseling scenario and conducted two non-inferiority studies to evaluate the performance of LLM-generated responses compared to human experts in actual alcohol abuse counseling simulations. Additionally, they demonstrated that LLM-driven virtual agents could conduct complete counseling sessions, maintaining coherent and professionally standard therapeutic interactions.