Large language models in adolescent suicide prevention: from language signals to accountable action

Adolescent suicide risk is often expressed through language before it becomes visible in clinical encounters, yet current prevention still relies heavily on episodic screening and limited specialist resources. Large language models (LLMs) may help interpret crisis chats, clinical notes, school-counseling records, and online disclosures, but their relevance to public mental health depends on a practical question: can language signals lead to timely and accountable human action? This mini review examines LLMs in adolescent suicide prevention through this Evidence-to-Action perspective. Current evidence is strongest for supervised risk-signal flagging and more preliminary for context summarization and extraction of suicide-related information from unstructured text. Evidence is much weaker for autonomous generative crisis intervention, especially for adolescents. The main barriers are not only technical accuracy, but also adolescent-specific validation, privacy and consent, cultural variation in distress expression, unsafe generated responses, and unclear responsibility after an algorithmic alert. We argue that near-term deployment should remain under human review and should be judged by whether it improves review timeliness, proportional escalation, safety planning, linkage to care, and equity. LLMs may contribute to adolescent suicide prevention only when embedded in clinically accountable and ethically governed systems. As a focused mini review, this paper maps conceptual boundaries and evidence gaps rather than proposing a validated deployment protocol.