Ground Truths in Suicide Research: The Current State of AI-Based Suicide Detection in Social Media

Key Findings

• — 195 studies were identified; the field is growing rapidly and is concentrated on a few platforms, with Reddit the sole source in 48% (94 studies) and present in 113, Twitter in 64; almost all studies rely on text, mostly English.
• — Only 14 studies (7.2%) established ground truth through individual-level assessment such as a clinical instrument; the majority infer risk from observable content features such as linguistic markers or community membership.
• — Because labels are inferred from posts, the predictive task shifts from identifying individuals at risk to classifying posts that contain suicidal or distress-related language, limiting the ability to detect people who do not express such content explicitly.
• — Dataset reuse is widespread; one prominent example labels 232,000 Reddit posts solely on subreddit membership, treating posts from certain communities as evidence of risk and others as its absence.
• — The authors call for ground-truth approaches more closely tied to validated clinical assessment such as the Columbia Suicide Severity Rating Scale.

Methodology Notes

Umbrella review of 22 systematic reviews (studies to 2022) plus an updated structured literature review; no meta-analysis; heterogeneity and search cut-off acknowledged. Authors at Ariel University, University of Cambridge, Reichman University, University College London and the Technion. CLPsych 2026 proceedings paper 1.19 (July 2026, month precision, matching the held CLPsych rows); Anthology landing page and 8-page PDF fetched; no DOI printed. Preprint arXiv 2606.28334 (26 May 2026).

Sources