Explainable AI for Suicide-Risk Detection in Online Text: A Review of Methods, Evaluation Practices and the Black-Box Gap

Automated classifiers are increasingly studied as tools for identifying online posts that may indicate suicide risk, so that trained human reviewers can prioritise them. Such systems can report high accuracy, but a flag that arrives without an account of what produced it is difficult to trust in a safety-critical setting. This paper is a structured narrative review of text-based suicide-risk detection and of the role of explainability in it. It is organised around three dimensions: modelling approach, data and labelling practice, and evaluation design, including whether explanations are tested for faithfulness and whether author-level leakage and missed cases are addressed. The review is purposive rather than exhaustive: it synthesises two prior surveys and a set of five primary studies whose bibliographic details and key characteristics were verified against their published records, and it combines them with the established literature on explanation methods and evaluation pitfalls. No systematic database search, screening count or pooled estimate is claimed. Within the studies examined, accuracy and F1 are the dominant reported measures; explanation, where present, takes the form of attention weights validated against proxies rather than tested for faithfulness; and fine-grained risk prediction remains weak. The paper contributes a critical synthesis, a comparison of explanation methods, and a proposed reporting checklist, and argues that shared evaluation standards are needed before reported performance can be compared fairly or used to justify deployment