Officials Hedge, Public Doubts: The Certainty Gap in China's AI-Education Discourse (2020–2026)

Chinese universities have launched a wave of “AI+X” programs that train students in professional domains—architecture, journalism, accounting, teaching—to use generative AI tools, often before any consensus exists that AI-assisted learning improves educational outcomes. Empirical research to date finds no reliable improvement in grades, documents cognitive offloading, and leaves long-term learning effects unmeasured. This paper asks whether that disconnect between institutional acceleration and empirical uncertainty surfaces in discourse: do official and public articulations of AI’s role in education differ in epistemic certainty, and is any such gap innate or emergent? We construct two corpora spanning 2020–2026—official discourse (policy documents and official-media posts, n = 691) and public discourse (Weibo, n = 11,345)—and measure four dimensions of epistemic stance (certainty markers, presupposition, hedging, and tolerance of uncertainty) using a hybrid design that pairs a transparent dictionary with LLM zero-shot coding, cross-validated at a mean Cohen's κ of .926. Three findings stand out. First, before ChatGPT the two corpora were epistemically indistinguishable, replicating Zeng, Chan, and Schäfer’s (2020) “striking similarity”; the gap emerged only after 2023. Second, the gap is not unidirectional: official discourse presupposes AI’s value more (D2 = 82.8% vs. 77.2%; policy documents 90.7%) yet hedges and acknowledges uncertainty significantly more than the public does (D3 and D4 reversed)—a pattern we term institutional hedging. Third, public epistemic stance is bimodal (absolute assertion or absolute doubt), whereas official discourse is layered across premise and expression. Time-series and difference-in-differences analyses locate two shocks—ChatGPT (2023) broke the stance layer (public D2 fell from 82% to 52%), and DeepSeek-R1 (2025) deepened the expression layer (D3 reached its series maximum)—suggesting the public underwent an epistemic learning process, first learning to doubt and then to qualify. We contribute a measurable, multidimensional discourse variable for the certainty gap and a hybrid coding template for large-scale epistemic-stance annotation.