Many adolescents show suicidal distress in words long before a clinician, teacher, or crisis worker sees them. Thoughts can rise and fall within days, and a short screening appointment often misses shame, hopelessness, or an indirect farewell. Specialist services remain scarce, so prevention still depends on episodic contact that does not match how risk actually moves.
A mini review published in Frontiers in Public Health asks whether large language models can turn those language signals into timely and accountable human action. The researchers argue that the tools matter for public mental health only if a person stays responsible for what follows an alert. They define adolescents mainly as people aged 10–19, and they treat evidence from adult or mixed youth samples as adjacent rather than direct proof for this group.
The strongest support so far is for supervised flagging of risk signals, especially in crisis chats and hotline conversations where the language is time stamped and produced in acute distress. Encoder-based classifiers, which sort text rather than write replies, have used features such as self-reference, negation, low self-worth, and absolutist wording to identify suicidal ideation or high risk behaviour in youth crisis text users. Related hotline studies have combined machine learning with psychological constructs during live chats. Much of this work is still retrospective, and labels often blend ideation, intent, and behaviour in ways that hide clinically important differences.
Evidence is more preliminary for summarising context or extracting suicide related information from clinical notes, school counselling records, and other unstructured text. Models can surface housing instability, family conflict, trauma, substance use, and barriers to care. That may help a counsellor organise a scattered history, but extraction is not the same as a tested suicide prevention programme. Generative systems, which produce their own replies, are on much weaker ground. Evaluations have found method related information, false reassurance, and invented services in answers to suicide prompts. For adolescents, hallucinated help and emotional dependence can delay a caregiver and meet a tendency to act quickly.
Online disclosures need extra care. Young people may write about hopelessness or burdensomeness on social platforms before they contact a service, yet slang, irony, memes, and fast changing peer norms make automated reading unreliable. Most of this detection work remains adult centred. Across settings, the review finds little proof that model outputs improve the outcomes that matter, including completed safety planning and linkage to care.
Privacy and consent are practical barriers, not side issues. These records can reveal family conflict, abuse, sexuality, or self harm thoughts, and routine de-identification is not enough. Guardian permission alone may not protect trust, while adolescent assent alone may not cover imminent risk. Cultural differences in distress language, uneven clinical notes, and alert thresholds can also widen inequity if a model fails in groups that already struggle to reach care.
No published prospective trial has yet measured how long an alert takes to reach a reviewer, whether escalation is proportionate, or whether care follows. The authors say near term use should stay under human review and should be judged by review speed, appropriate escalation, safety planning, linkage to care, and equity. Automated emergency dispatch or mandatory family notification should not rest on model output alone. Developers, schools, clinicians, and crisis services need a clear chain of responsibility, with records of inputs, uncertainty, human review, and follow up.
Until those trials exist, large language models belong as adjuncts that help staff notice fragmented risk language, not as independent counsellors for adolescents in crisis.
