Hostname: page-component-7b7567c8c-wxvwp Total loading time: 0 Render date: 2026-11-15T08:54:39.041Z Has data issue: false hasContentIssue false

Testing Empathy in Machines: A Comparative Study of Large Language Models and Motivational Interviewing

Published online by Cambridge University Press:  31 July 2026

J. B. Rodrigues*
Affiliation:
Departamento de Psiquiatria e Saúde Mental, Unidade Local de Saúde de Lisboa Ocidental, Lisboa, Portugal
J. P. Azenha
Affiliation:
Departamento de Psiquiatria e Saúde Mental, Unidade Local de Saúde de Lisboa Ocidental, Lisboa, Portugal
P. Trindade
Affiliation:
Departamento de Psiquiatria e Saúde Mental, Unidade Local de Saúde de Lisboa Ocidental, Lisboa, Portugal
C. Adão
Affiliation:
Departamento de Psiquiatria e Saúde Mental, Unidade Local de Saúde de Lisboa Ocidental, Lisboa, Portugal
*
*Corresponding author.

Abstract

Core share and HTML view are not available for this content. However, as you have access to this content, a full PDF is available via the 'Save PDF' action button.
Introduction

Large language models (LLMs) are increasingly proposed as tools for behavioral healthcare, with the potential to augment or even automate aspects of psychotherapy. Among therapeutic frameworks, motivational interviewing (MI) is especially relevant as it is well-validated for eliciting behavior change and improving adherence. MI’s structured, empathetic style provides a rigorous benchmark for assessing whether LLMs can replicate therapeutic dialogue. While LLMs promise scalability and accessibility, their application in psychotherapy remains high-stakes, raising concerns of safety, fidelity, and ethical deployment.

Objectives

This study aims to evaluate whether current leading LLMs reproduce the principles of MI, focusing on empathy, change talk, rolling with resistance, and support for autonomy.

Methods

A standardized MI script addressing ambivalence toward alcohol reduction was applied across five leading LLMs: GPT-4, GPT-3.5, Claude, Gemini, and Llama-2. Outputs were coded against MI core processes (engaging, focusing, evoking, planning), MI principles (empathy, discrepancy, resistance management, non-confrontation, self-efficacy), and common techniques (open questions, reflections, affirmations, eliciting change talk, avoidance of directive advice).

Results

GPT-4 most consistently embodied MI: it engaged and focused effectively, elicited change talk, and demonstrated empathy through accurate reflections. It avoided confrontation and supported self-efficacy, applying all five principles with notable fidelity. Claude showed similar strengths, especially in empathy and discrepancy development, but was less reliable in rolling with resistance. Gemini balanced empathy with structure, producing clear open questions and affirmations, but defaulted to directive recommendations, limiting adherence to non-confrontation and autonomy support. GPT-3.5 generated open questions but frequently lapsed into advice-giving, undermining MI principles and the planning process. Llama-2 showed the lowest fidelity, often offering prescriptive advice rather than reflections, with minimal engagement in evoking change talk or consolidating commitment. Across models, nuanced techniques such as managing resistance and strengthening commitment were inconsistently applied, with GPT-4 and Claude demonstrating the highest overall fidelity to MI principles and techniques.

Conclusions

LLMs can reproduce surface features of MI and show promise for scaling access to supportive dialogue. However, consistent therapeutic stance and deeper MI techniques remain beyond current capacity. These findings echo concerns that while LLMs may simulate skills, they do not yet deliver the clinical sensitivity required for psychotherapy. Responsible development should prioritize domain-specific training, interdisciplinary oversight, and hybrid models where clinician supervision ensures therapeutic integrity.

Disclosure of Interest

None Declared

Information

Type
Abstract
Creative Commons
Creative Common License - CCCreative Common License - BY
This is an Open Access article, distributed under the terms of the Creative Commons Attribution licence (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted re-use, distribution, and reproduction in any medium, provided the original work is properly cited.
Copyright
© The Author(s), 2026. Published by Cambridge University Press on behalf of European Psychiatric Association
Submit a response

Comments

No Comments have been published for this article.