Instructional features of AI-generated English-learning videos on Douyin: associations with platform-visible audience responses and creator visibility

Purpose This study examines how platform-visible audience response counts are distributed across AI-generated English-learning videos on Douyin and how those counts are associated with human-coded instructional features and creator visibility. Likes, favorites, comments, and shares are treated as digital trace outcomes rather than evidence of language learning, allowing the analysis to distinguish public reception from educational effectiveness.Design/methodology/approach The dataset comprises 1120 unique videos from 97 creators, drawn from 1327 manually pre-screened records. Two independent coders viewed every video and rated six instructional features on 0–2 implementation/salience scales; weighted kappa values ranged from .717 to .772. Primary analyses used negative-binomial (NB2) regression with creator-clustered sandwich standard errors and Benjamini–Hochberg FDR correction. Poisson GEE, log(1 + x) OLS, a strict post-hoc validation subset, and leave-dominant-creator-out analysis served as sensitivity checks.Findings Audience responses were highly concentrated. After FDR correction, contextualized scenarios were positively associated with likes (IRR = 1.78), while pronunciation/listening was positively associated with likes (IRR = 2.18) and favorites (IRR = 1.66). Follower count was most consistently associated with comments (IRR = 1.27). Two NB2 moderation effects survived FDR correction but were not FDR-significant in Poisson-GEE sensitivity models and were therefore treated as exploratory.Originality/value The study separates platform response from learning, codes instructional features directly from videos, and treats AI generation as the empirical sample condition rather than a causal factor. It shows how educationally oriented video features can be examined without using platform popularity as a proxy for pedagogical quality.