Abstract
<title>Abstract</title> <p>The integration of generative AI into educational practice has created new demands on teachers' capacity to critically evaluate and adapt AI-generated content for diverse learners. This exploratory mixed-methods study investigated the "linguistic gap"—the difference in lexical, syntactic, and readability complexity between simplified and enriched content versions generated by the same teacher—among 89 Iranian EFL teachers (38 novice, 51 experienced) who produced 178 content samples during an AI-integrated professional development workshop. Automated linguistic analysis using Textstat and SpaCy, combined with reflexive thematic analysis of extreme cases, revealed that novice teachers produced significantly larger overall linguistic gaps (LG_total: novice M = 1.13, experienced M = 0.73; Hedges' g = 0.68, p = 0.002), with particularly pronounced differences in lexical (g = 0.82) and syntactic (g = 0.71) dimensions. MANOVA indicated no significant interaction between group and dimension (p = 0.160), suggesting consistent novice-experienced differences across linguistic domains. Distribution overlap analysis (OVL = 0.36) rejected the hypothesis that AI tools standardise adaptation practices. Qualitative analysis identified six patterns distinguishing polarised novice adaptation from scaffolded experienced adaptation. The findings contribute to understanding how teacher expertise shapes AI-enhanced content differentiation, with implications for teacher education, prompt engineering, and the design of AI tools that genuinely support pedagogical reasoning rather than amplifying existing competence gaps.</p>