Large language models may struggle to detect culturally embedded filicide-suicide risks

Area of Science:

• Artificial Intelligence
• Mental Health
• Sociocultural Studies

Background:

• Large language models (LLMs) are increasingly used in various analytical tasks.
• Assessing mental health risks requires nuanced understanding, including cultural context.
• Previous research has not extensively evaluated LLM capabilities in detecting culturally specific risks.

Purpose of the Study:

• To evaluate the performance of six advanced LLMs in identifying domestic violence, suicide, and filicide-suicide risks.
• To assess LLMs' comprehension of culturally specific cues within a Taiwanese flash fiction.
• To determine the limitations of current LLMs in culturally sensitive mental health risk detection.

Main Methods:

• Six LLMs (GPT-4o, GPT-o1, DeepSeek-R1, Claude 3.5 Sonnet, Sonar Large, Gemma-2-2b) analyzed the flash fiction "Barbecue".
• Models were prompted to interpret risks, simulating different mental health expertise levels.
• Analysis focused on the detection of domestic violence, suicide, and filicide-suicide, with attention to cultural context.

Main Results:

• All LLMs detected domestic violence risks.
• GPT-o1, Claude 3.5 Sonnet, and Sonar Large identified suicide risk using cultural cues; others missed it.
• No LLM understood the cultural context of the mother sparing her daughter, indicating a gap in recognizing non-Western sociocultural nuances.

Conclusions:

• LLMs demonstrate partial effectiveness in detecting overt risks like domestic violence.
• Current LLMs struggle with interpreting culturally embedded suicide and filicide-suicide risks, particularly those relying on non-Western contexts.
• Further development is needed to enhance LLMs' cultural competency for accurate mental health risk assessments.