Prescriptive Persistence: Quantifying the Breakdown in Human-AI Pedagogical Co-Regulation in ELL Writing Feedback

As AI-driven feedback tools become embedded in the classroom, they increasingly participate in the pedagogical regulation of student learning and evaluation. Yet unlike human teachers, who dynamically calibrate intervention based on intelligibility and learner development, automated systems apply prescriptive norms without context for the individual learner. This raises concerns about their role as collaborative partners with educators as opposed to pedagogical authorities. For English Language Learners (ELLs), a statistically significant and growing population in U.S. public schools, this misalignment poses particular risks of digital gate keeping, constraining learner agency and voice erasure for students whose written expression reflects valid semantic intent. This study introduces the Prescriptive Gap \textbf{(\(\Delta P\))}, a novel metric designed to quantify the divergence between AI corrections and human pedagogical edits. Using more than 12,000 sentences from the Cambridge Learner Corpus (CLC) First Certificate in English (FCE) dataset, we compare ChatGPT5.2's grammatical feedback to human experts across diverse first language (L1) backgrounds and proficiency levels. We reveal a breakdown in human-AI pedagogical co-regulation, where AI systems fail to respect human thresholds for intervention, disproportionately constraining learner agency for ELL writers. These findings characterize a form of prescriptive persistence where AI driven feedback systematically exceeds pedagogical intent, underscoring the need for evaluation frameworks that account for semantic preservation and the learner's voice in educational AI tools.