Lin-Zucker M. auteur aut IRD Bellassen J. auteur aut IRD Zucker Jean-Daniel auteur aut IRD text journalArticle text/pdf born digital access Large Language Models (LLMs) are increasingly used in Chinese as a Second Language (L2) learning, yet their ability to comply with pedagogical constraints specific to the Chinese writing system remains underexplored. This study examines whether system prompts aligned with the CEFR-EBCL framework enable LLMs to generate learner-facing Chinese texts that respect sinographic thresholds at the A1, A1+ and A2 levels. We conducted controlled experiments using two models (GPT-4.1 and GPT-4.1-mini) across ten EBCL-related written tasks. Prompt conditions with and without explicit character lists were compared. Model outputs were automatically analyzed to quantify instruction deviation, defined as the proportion of characters outside the target EBCL set. Results indicate that including explicit character lists significantly reduces out-of-list character production at the A1 and A1+ levels, particularly with GPT-4.1. At the A2 level, this effect becomes marginal. These findings provide empirical evidence on the pedagogical value and limits of prompt-based control of ChatGPT outputs for CEFR-EBCL-aligned Chinese L2 learning. specialized Chinese as a Second Language Large Language Models Prompt Engineering CEFR-EBCL Alignment Sinographic Constraints CHINE 106 020 17 1 1-29 2026 1949-260X https://www.documentation.ird.fr/hor/fdi:010097586 1949-260X [F B010097586] https://www.documentation.ird.fr/hor/fdi:010097586 https://www.documentation.ird.fr/intranet/publi/2026-08/010097586.pdf Accès réservé (Intranet de l'IRD) IRD - Base Horizon / Pleins textes 2026-08-06 2026-08-06 fdi:010097586 fre