School dropout is a multifactorial and persistent problem, and artificial intelligence has been proposed as a tool for the early prediction and monitoring of students at risk. This paper gathers and confronts a deliberately heterogeneous body of evidence, comprising systematic reviews, technical studies, randomized field experiments, independent impact evaluations and investigative reporting on real deployments, to examine not only whether predictive models work, but whether that capability translates into an effective reduction of dropout, and under which conditions. The synthesis shows strong consensus on the technical feasibility of prediction, with ensemble methods reaching high reported performance, alongside a critical gap between predictive accuracy and intervention effectiveness. The experimental record is mixed rather than uniformly discouraging: alerts that merely inform advisers produced null results, whereas structured, resourced and sustained interventions produced measurable gains. Documented harms include racial bias in production, stigmatization and the capture of the tool for the manipulation of institutional indicators. Building on this analysis, the feasibility of such a project in Brazil is discussed, considering data infrastructure, institutional capacity for intervention, equity and a legal framework that changed substantially between 2025 and 2026. The paper closes with a short, conditionally favorable opinion, whose conditions are detailed.