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

Key Findings

• — 164 adults from an online community cohort, oversampled for suicide risk, completed baseline suicide-risk questionnaires and future-self interviews with 18-month follow-up.
• — Participants for whom a suicidal future was more linguistically plausible at baseline had higher rates of subsequent ideation at 18 months: 82.8% versus 47.6% (chi-square 12.26, p < .001).
• — Discrimination was AUC 0.70 for 18-month ideation.
• — 75% of individuals classified low-risk by a standard questionnaire but high-risk by the linguistic forecast went on to report subsequent ideation.
• — The method demonstrated convergent validity with baseline assessment measures and incremental validity over them.
• — The mechanism uses perplexity as a proxy for psychological plausibility, so the signal is carried by ordinary future-oriented language rather than by disclosure.

Methodology Notes

Longitudinal online community cohort oversampled for suicide risk (n = 164), baseline structured interview about the future self, two LLM-generated counterfactual narratives per participant, perplexity scored by a second LLM against the baseline interview, outcome self-reported ideation at 18 months. Proof-of-concept scale, single cohort, self-reported outcome, no external replication, and neither model is named in the abstract. Preprint, not peer-reviewed; posted 2026-01-07, DOI 10.31234/osf.io/fhzum_v1. No journal version found in Crossref as of 2026-09-14.

Sources