Citations Are Not Enough: Grounding AI-Generated Instructional Materials in Real Sources

Generative AI is increasingly used to produce content from assigned sources, but access to a source does not guarantee fidelity to it. A generated claim can be plausible or even factually correct while still being unsupported by the source it is supposed to represent. We study this problem in instructional generation, where generated materials are expected to remain faithful to an assigned textbook. We present a source-grounded instructional generation system that extends an existing instructional generation framework by constraining what evidence the model can use during generation. The system parses and indexes a textbook, binds each course chapter to a small set of textbook sections, restricts per-slide retrieval to those sections, and checks each generated claim against the assigned source, counting it as supported only when the judge returns supporting evidence that is mechanically verified to occur in that source. We measure Grounding Fidelity, the proportion of generated claims supported by the assigned source. On three matched chapters, our system achieves 77.5\%, compared with 64.6\% for the original generator given the complete chapter text; across the twelve course weeks with assigned chapters, fidelity rises from 47.9\% to 60.9\%. Ablations show that chapter binding matters more than grounding instructions: removing the binding drops fidelity below the original generator despite retaining retrieval. Together, these results suggest that source access alone is not enough; source-grounded generation depends on controlling what evidence guides the model.