Mapping the Content and Consistency of Suicidal Thoughts with Large Language Models: A Two-Study Longitudinal Investigation Among Adolescents and Young Adults

Key Findings

• — Study 1: N = 158 adolescents; Study 2: N = 202 young adults, all with past-year suicidal thoughts and behaviours, completing open-ended items at baseline and follow-ups.
• — The pipeline recovered 13 topics of suicidal-thought content in Study 1 and 14 in Study 2, including Family and Friends Reactions and Depressed and Exhausted.
• — Within-person and between-person consistency were both U-shaped: agreement is highest at the endpoints of the rating scale.
• — The same numeric urge rating therefore carries materially different content across people and across occasions, concentrated in the middle of the scale.
• — The authors position LLM topic modelling as a way to advance measurement of suicide risk rather than as a detector.

Methodology Notes

Two longitudinal samples with open-text self-report at multiple timepoints; topics are LLM-derived and therefore inherit the model's clustering choices; the samples are people already reporting past-year suicidal thoughts and behaviours, so this is about measurement within a high-risk group rather than screening. Preprint, not peer-reviewed; v3 posted 2026-08-31 on PsyArXiv (osf.io/preprints/psyarxiv/8jx4u); OSF API record, abstract and contributor list verified.

Sources