Leonard, D., Gole, R. & Dacon, J. (2026) Prescriptive Persistence: Quantifying the Breakdown in Human-AI Pedagogical Co-Regulation in ELL Writing Feedback. In Proceedings of the 2026 ACM Learning @ Scale Conference

Abstract:

As AI-driven feedback systems are deployed at scale, they increasingly function as pedagogical regulators of learner writing. While human educators calibrate interventions based on communicative intelligibility, LLMs often apply prescriptive linguistic norms that risk over-correcting learner intent. We introduce the Prescriptive Gap (Δ𝑃), a metric quantifying the divergence between AI generated edits and expert human pedagogical revisions using normalized Levenshtein distance. Applying this to 12,000+ sentences from the Cambridge Learner Corpus (CLC), we compare ChatGPT 5.2’s feedback to expert edits across 14 first-language (L1) backgrounds. Our results reveal Prescriptive Persistence, a phenomenon where AI systematically exceeds human intervention thresholds. Crucially, we demonstrate that this gap is highly sensitive to instruction; “frugal” prompting can align AI intervention more closely with human pedagogical standards. These findings suggest that Δ𝑃 can serve as a design-time metric to develop more pedagogically aligned, voice-preserving feedback systems for English Language Learners (ELLs).