Grant: Fellowship: Canada Impact+ Research Training Awards - Postdoctoral

"Artificial intelligence, including large language models (LLMs), is increasingly used in mental health research and care. While promising, these tools raise serious concerns about safety, reliability, and potential harm in sensitive health contexts. This research evaluates how LLMs perform in realistic mental health scenarios, with a strong focus on patient safety and responsible clinical translation. Specifically, it examines how LLMs respond to complex and emotionally charged stressor inputs, such as ambiguous language or conflicting information, which commonly arise in mental health settings. Using multi-modal data, including text, speech, video, and anonymized clinical records, the study assesses whether models produce consistent, cautious, and interpretable responses, or generate misleading, overconfident, or unsafe outputs. Different LLM models will be compared based on their complexity, design, and built-in safety mechanisms to identify characteristics associated with safer behavior. Models will be evaluated using healthcare-relevant metrics, including uncertainty awareness, confidence calibration, robustness to input variation, and the likelihood of producing incorrect or harmful outputs. The research will explore practical strategies to reduce risk, including guiding models with trusted clinical knowledge sources and using lightweight adaptations that avoid extensive retraining. To improve generalizability, the study will use data representing diverse populations across age, sex, gender, Indigenous Peoples and communities, and sociocultural backgrounds. By identifying failure modes and mitigation strategies, this work will provide evidence-based guidance for the safe integration of AI into mental health research and future clinical decision support tools."