Large Language Models Can Forecast Suicidal Ideation From Future Self-Narratives

Objective: Standard self-report methods for suicide risk assessment have limitations, including an individual’s reluctance to disclose acute suicidal thoughts or lack of self-insight. This is the first longitudinal study to provide proof-of-concept validation of a computational narrative-based method to predict subsequent suicidal ideation independent of explicit disclosure.Method: We recruited 164 adults from an online community cohort, oversampled for suicide risk, who completed baseline suicide risk questionnaires and interviews about their future self. We used one large language model (LLM) to generate two narratives per participant: one with a future potential suicide event, and another without a future suicide event. A second LLM estimated the perplexity (linguistic plausibility) of each narrative relative to the participant’s baseline interview. The differential plausibility scores of the narratives were assessed for convergent validity and the prediction of suicidal ideation at 18-month follow-up.Results: Participants for whom a suicidal future was more plausible at baseline had significantly higher rates of subsequent ideation at the 18-month follow-up (82.8% vs. 47.6%; χ²=12.26, p<.001; AUC = 0.70). The method demonstrated both convergent validity with baseline assessment measures and incremental validity. In addition, 75% of the individuals classified as low-risk by a standard questionnaire, but high-risk by the linguistic forecast, went on to report subsequent ideation.Conclusions: The language people use to describe their future contains quantifiable linguistic signals of future suicidal ideation. Using perplexity as a proxy for psychological plausibility, linguistic forecasting offers a scalable, implicit, and theoretically grounded approach to enhance suicide risk assessment.