Where there’s an error, can GenAI find a way? Exploring personalized vocabulary exercises in L2 Spanish

Abstract Lexical errors are among the most frequent and persistent issues in second language (L2) production, yet research on vocabulary exercises has largely focused on the acquisition of new words rather than on the correction of misused ones. This exploratory repeated-measures study investigates the potential of generative artificial intelligence (GenAI) to provide personalized vocabulary practice targeting learner-specific lexical errors in Spanish. Eighty-five L2 and heritage learners across proficiency levels completed 15 individualized exercises generated by a custom GPT-based system from their own writing errors. The tasks varied in goal, format, linguistic complexity, and involvement load (IL). Error recognition and correction were assessed at immediate and delayed post-tests, and mixed-effects models showed that higher exercise performance and task-induced IL predicted better outcomes. Both recognition and correction declined at the delayed post-test, although only recognition did so significantly. Among exercise features, productive, contextualized formats such as text editing, translation, and sentence reconstruction showed strong positive associations with correction, while goals emphasizing paraphrastic flexibility were less favorable. At the learner level, outcomes also varied substantially. Overall, the findings suggest that GenAI can support scalable, individualized, error-driven vocabulary practice. Still, variability across learners and exercises, together with incomplete retention, indicates that personalization alone may not ensure durable learning. Future research should thus examine whether feedback, recycling, and adaptive review improve retention.