Competence narrowing in AI-assisted oral English instruction: A quasi-experimental study of the spiral synergy framework

The rapid integration of generative AI tools into spoken English instruction has produced measurable fluency and accuracy gains, but whether these are evenly distributed across the dimensions of communicative competence remains underexplored. This quasi-experimental study compares the Spiral Synergy Framework (SSF), a three-layer pedagogical sequence (AI Consolidation, Bridging Tasks, Interpersonal Activation), against conventional AI-integrated instruction in two intact Business English classes at a Chinese higher vocational college (experimental n = 36, control n = 34), using oral recordings at three time points scored on five competence dimensions, 15 weeks of classroom observation, and an intra-rater reliability check. The two classes differed significantly at baseline on the oral composite score and three of five dimensions (control higher), though not on CET-4 proficiency or questionnaire items. Because treatment is confounded with class membership and both classes received AI-assisted practice, findings are reported as associations rather than demonstrated causal effects. Linear mixed-effects models testing Group × Time interactions across five outcomes (four dimensions and the composite), corrected for multiple comparisons, identified one outcome — real-time responsiveness — on which the classes diverged: the experimental class maintained responsiveness across the semester while the control class declined despite its baseline advantage (Benjamini–Hochberg-adjusted p = .029), a result that held under a cluster-robust sensitivity check. No corrected between-group difference was found for improvisation, coherence, naturalness, or the composite score; improvisation declined descriptively in both classes over the semester. Classroom observation data (experimental class only) document a recurring, non-monotonic cycle of AI-dependency reduction and rebound across 15 weeks. Intra-rater test–retest reliability of the oral-competence ratings was moderate to good (ICC = .67–.91), though this reflects a single rater familiar with participants rather than independent inter-rater agreement. Implications for multidimensional assessment, task sequencing, and cautious interpretation of AI-integrated curriculum research are discussed.