Hallucination in Large Language Models: Epistemic Status, Detection, and Validation Criteria in Higher Education

False statements produced by a language model arrive with the same confident form as true ones, with no mark to guide verification, and this lack of distinction directly affects the university classroom. This narrative review, based on documented searches and purposive selection, examines the peer-reviewed literature and retains thirty-two works and one continuously updated resource drawn from technical research on detection, epistemology, cognitive psychology, critical thinking, and educational research. The argument starts from an epistemic-functional characterization of the model, which neither takes ownership of its output as an assertion it could answer for nor has a criterion of truth of its own; it follows that every correction comes from a reference external to the generative process, which we call external anchoring. On that basis, we classify divergences according to whether they are measured against facts about the world or against material supplied by the user, review detection procedures together with their declared limits, and formulate four criteria that can be applied through a public interface, grouped into two families and mapped against the typology in a table. The closing sections examine two effects on the reader, cognitive satisfaction and the externalization of epistemic vigilance, and the institutional conditions needed to counter them.