Abstract
Generative AI is a fast-diffusing technology whose adoption in education outruns the ability of institutions to assure the quality of what it produces. Teaching analogies are a revealing case: language models can now generate them on demand, yet a poor analogy plants misconceptions, so the managerial problem is not generation but assurance at scale without a human rating every item. Using only small open models on one consumer GPU, with no paid interface and no human annotation, we study how to prompt for good analogies and how to evaluate them automatically. For 64 school-level concepts in four subjects, three models generate analogies under four source-domain strategies, and a four-model panel scores each on a five-dimension rubric. Constraining the source domain confers no advantage over letting the model choose; judged quality is highly predictable, but from length and subject, a verbosity bias we quantify. The panel screens out planted-wrong analogies almost perfectly yet barely agrees when ranking good ones, so automated evaluation is a valid quality screen and not a fine ranker. A plain explanation lifts weak learners while an added analogy does not, and an English safeguard does not carry over to a Hindi replication. We read the results for technology management: when automated screening becomes ownable quality-assurance infrastructure rather than a bought-in service, what to make rather than buy, and why the cross-lingual gap is a technology-transfer priority for developing economies.