A Corpus-Based Analysis of Epistemic Stance in AI-Generated Instructional Content

Abstract

AI systems generate educational content, yet their rhetorical characteristics remain insufficiently understood. This study examines epistemic stance construction through hedging and boosting devices in AI-generated discourse based on one hundred fifty AI-generated texts sampled from educational, professional, and conversational domains. Using a systematic corpus-based mixed-methods analysis that combined quantitative frequency measurements with qualitative functional interpretation, this investigation reveals significant asymmetry in rhetorical positioning. Educational texts displayed a mean hedging frequency substantially higher than booster deployment, creating a hedge-to-booster ratio that far exceeds patterns documented in human pedagogical discourse. Qualitative analysis identified numerous instances of inappropriate hedging in foundational content where confident presentation would better support learning, alongside relatively sparse and structurally formulaic use of boosters. These findings suggest that AI systems overgeneralize cautious patterns learned from training data to contexts that call for assertive instructional guidance. They also show critical limitations in current AI-based tools and establish frameworks for developing AI-mediated learning that supports rather than undermines knowledge construction in digital communities.