Abstract
Commercial textbooks impose structural constraints: high cost, fixed reading level, and limited local relevance. Generative AI enables instructors to produce weekly readings tailored to student interests and comprehension levels, but introduces risks around sourcing and unsupported claims. This mixed-methods classroom study documents a proof of principle for AI-generated readings in a graduate educational leadership course, examining practical acceptability and instructional boundaries. The intervention used an AI assistant to generate weekly readings through iterative student prompting, implementing dual tailoring (differentiation) along two dimensions: interest (examples, sector focus, professional role) and comprehension level (pacing, scaffolding, vocabulary density). Data include end-of-course survey responses from 24 students and systematic artifact analysis of weekly reading session logs. Students rated the readings as effective for preparation, with 75 percent agreeing they learned more than in a comparable course without AI assistance. Artifact analysis confirms substantial tailoring through sector anchoring, role framing, and comprehension adjustments triggered by clarification requests. However, readings often lacked internal traceability, citation quality was inconsistent, and over-specific institutional claims appeared without supporting evidence. Students demonstrated bounded trust, valuing the readings while requesting instructor oversight. The study contributes design principles for responsible implementation. An emergent secondary finding is that the intervention also appeared to function as implicit training in critical engagement with probabilistic text: over the course of the semester, students learned to request checkable evidence and to recognize when claims exceeded available warrants.