The grass is not always greener: Teacher vs. GPT-assisted written corrective feedback

Abstract

Written Corrective Feedback (WCF) is a crucial pedagogical practice where teachers annotate student writing to correct errors and improve language skills, albeit one that is time-consuming and laborious for large classes or under time constraints. However, the advent of advanced generative artificial intelligence and large language models, specifically ChatGPT, has introduced new possibilities for automating such educational tasks. GPT models with their transformer architecture and self-attention mechanism can perform complex natural language tasks including assisting teachers in providing WCF. This study compares the WCF produced by teachers and ChatGPT, examining their respective capabilities while identifying differences in their feedback practice. Findings reveal teacher provided WCF typically involves a consistent combination of direct correction and indirect feedback forms addressing both local and global issues, albeit with a degree of inaccuracy. ChatGPT-assisted WCF tends to be in the form of metalinguistic feedback and/or reformulation of the original text. However, GPT also frequently varies in its entire approach to WCF provision even when using the same prompt on the same text, while also providing grammatically accurate yet redundant WCF in certain cases. We discuss the implications of these findings for L2 writing practice.