Effect of the SARA Citation-Verification Framework on Medical Students' Verification of AI-Generated Citation Hallucinations: A Randomized Controlled Trial With an Explanatory Sequential Mixed-Methods Component

AI-generated citation errors can occur at different levels. Some are source-level errors, such as fabricated references or incorrect bibliographic details. Others are more difficult to detect because the cited publication is genuine, but the evidence may be irrelevant, may not support the direction, magnitude, population, or conclusion of the claim, or may be presented without important qualifiers or limitations. In this study, these problems are examined across a five-level citation-error taxonomy ranging from non-existent sources to contextual omission or overgeneralization.

The SARA framework was developed as an educational scaffold to make citation-verification reasoning explicit. It prompts learners to ask whether the source exists, whether the citation details are accurate, whether the source is relevant to the specific claim, and whether the evidence is aligned with the claim as written. The intervention uses modelling, coached practice, feedback, articulation, reflection, and progressive removal of support so that learners move from guided verification toward independent judgment.

A key feature of the study is assessment of delayed transfer rather than immediate post-training performance alone. At 12 weeks, participants will review a realistic literature-review passage containing both valid and problematic citation-claim relationships without being told how many errors are present or being prompted to use SARA. This is intended to evaluate whether the verification strategy has become sufficiently internalized to be applied independently in a new scholarly context.

The explanatory qualitative component will explore why some learners successfully transfer verification skills while others continue to miss deeper evidence-claim problems. Think-aloud and stimulated-recall interviews will examine what triggers verification, how learners decide that evidence is sufficient, when they stop checking, and how they distinguish source existence from evidentiary support.