AI Classifies Online Suicide Content Into Levels

In a 2026 JMIR Medical Informatics study, researchers from SoftlyAI, Kyung Hee University, Sungkyunkwan University and partners collected 43,244 user posts and created a multimodal benchmark with 452 manually labeled entries (Korean and English) to classify suicide-related content into five harm levels. They used GPT-4 preannotation and evaluated zero-/few-shot approaches, reporting GPT-4 F1-scores of 66.46 (illegal) and 77.09 (harmful); few-shot learning and image captions improved moderation accuracy.