How Well Does AI-Generated Feedback Work? Intrinsic and Extrinsic Evaluation across more than 20,000 EFL Essay Drafts

arXiv cs.CL · 6 hours ago

Researchers evaluated AI-generated written corrective feedback for English language learners by analyzing over 20,000 essay drafts from nearly 2,000 students at a university. Expert teachers and students assessed the feedback using different evaluation methods, with experienced instructors rating the feedback quality via rubric while students reported their own perceptions and engagement. The study found low correlation between teacher ratings and student perspectives, indicating that expert evaluation alone does not adequately measure whether AI feedback is actually useful to learners, and suggesting language education applications need learner-centered evaluation approaches.