Trust but Verify: Testing a Framework for AI Interrogation in Nursing Education

Abstract

Generative AI produces fluent, confident output that is sometimes fabricated. Nursing students increasingly submit work built on sources that do not exist, or that exist and do not support the claims drawn from them. This session reports two semesters of classroom observation testing whether prompt design can address that problem. The hypothesis was that structuring student prompts to require the AI to validate its own sources would improve detection of hallucinated citations. It did not. Across two sections, students who used the structured format caught fabricated citations at the same rate as students given an open prompt. Structure did change something: structured students were substantially more likely to notice that something was wrong, but no more likely than the open-prompt group to finish the check. The session presents both rounds, including the confound identified in the first and removed in the second, and the limits of the observation, among them students who did not use the assigned structure. It also reports an unplanned finding: the AI tool used to review the submissions produced the same category of error it was reviewing for, misattributing a real paper to a fabricated author across multiple students before the error was caught on human review. Attendees will leave with a way to check sources that starts cheap and gets expensive, a clearer view of why you cannot ask the model to check its own work, and a specific account of where student checking breaks down.