PsychEthicsBench: Evaluating Large Language Models Against Australian Mental Health Ethics

TL;DR

PsychEthicsBench addresses the gap in mental-health AI safety by grounding evaluation in Australian psychological and psychiatric ethics, moving beyond refusal signals to measure ethical knowledge and behavior. The authors curate 392 principles into 1,377 MCQs and 2,612 OEQs through a three-stage pipeline, and assess 14 models across Aussie and Global settings, revealing that refusal rates poorly predict ethicality and that domain-specific fine-tuning can degrade ethical robustness. A multi-dimensional annotation framework computes metrics such as $QPR$, $OER$, and $CER_{ig|Q}$ to separate quality, ethical violations, and conditional ethics, demonstrating that medical LLMs often outperform mental-health-specialized variants, while some models misrepresent credentials or rely on U.S.-centric references. The work provides a foundation for jurisdiction-aware evaluation of mental health LLMs, highlighting ethical risks and guiding responsible development across regions, with limitations including reliance on LLM judges for annotations and a focus on Australian context. Overall, PsychEthicsBench offers a principled, transparent approach to assess and improve ethical alignment of LLMs in mental health contexts.

Abstract

The increasing integration of large language models (LLMs) into mental health applications necessitates robust frameworks for evaluating professional safety alignment. Current evaluative approaches primarily rely on refusal-based safety signals, which offer limited insight into the nuanced behaviors required in clinical practice. In mental health, clinically inadequate refusals can be perceived as unempathetic and discourage help-seeking. To address this gap, we move beyond refusal-centric metrics and introduce \texttt{PsychEthicsBench}, the first principle-grounded benchmark based on Australian psychology and psychiatry guidelines, designed to evaluate LLMs' ethical knowledge and behavioral responses through multiple-choice and open-ended tasks with fine-grained ethicality annotations. Empirical results across 14 models reveal that refusal rates are poor indicators of ethical behavior, revealing a significant divergence between safety triggers and clinical appropriateness. Notably, we find that domain-specific fine-tuning can degrade ethical robustness, as several specialized models underperform their base backbones in ethical alignment. PsychEthicsBench provides a foundation for systematic, jurisdiction-aware evaluation of LLMs in mental health, encouraging more responsible development in this domain.

Figures (15)

• Figure 1: Limitations of refusal-based safety metrics motivate multi-dimensional PsychEthicsBench.
• Figure 2: Overview of the three-stage PsychEthicsBench data curation pipeline: (I) guideline collection, (II) expert-in-the-loop prompt optimization, and (III) expert-centered quality control, resulting in high-quality multiple-choice and open-ended questions.
• Figure 3: Distribution of ethical principles by guideline and discipline (a), multiple-choice questions by source and discipline (b), and open-ended questions by guideline and inquirer role (c) in PsychEthicsBench.
• Figure 4: Category definitions for rule-breaking behaviors used in ethicality annotation of OEQ responses.
• Figure 5: Frequency of the America-related phrases (listed in \ref{['fig:america-related']}) in responses to OEQs.