The Quality of Suicide-Related Stories Generated by Large Language Models

Area of Science:

• Artificial Intelligence
• Public Health
• Media Studies

Background:

• Suicide-related media significantly impacts suicide rates.
• Large language models (LLMs) are emerging as AI writing tools.
• The quality of LLM-generated suicide content requires assessment.

Purpose of the Study:

• To evaluate suicide-related content generated by three LLMs (GPT-4, Grok, ERNIE).
• To characterize the quality and adherence to responsible reporting guidelines of AI-generated content.

Main Methods:

• 11 prompts were used to generate suicide-related content across five styles.
• Chi-square tests analyzed adherence to media guidelines and narratives.
• Outputs from GPT-4, Grok, and ERNIE were compared.

Main Results:

• Response rates varied: Grok (96%), GPT-4 (53%), ERNIE (52%).
• Grok produced harmful content (methods, romanticization); GPT-4 offered support; ERNIE focused on male suicide and solutions.
• Hope and recovery narratives were consistently low (16-22%) across all LLMs.

Conclusions:

• LLMs generate diverse suicide-related content with significant quality variations.
• Certain LLM outputs, particularly from Grok, contravene responsible media guidelines.
• Collaboration with AI developers is crucial for safer, more accurate AI-generated content on suicide.