How do Teaching Assistants perceive LLM-Generated Feedback?

• In Event: Innovative approaches to improving postsecondary teaching

Abstract

Feedback is vital for learning but often limited by instructor workload. This study explores whether feedback generated by a large language model (LLM) can expedite grading and enhance feedback quality in a college economics course. We developed an AI system to generate essay feedback aligned with detailed rubrics and conducted 20 one-hour think-aloud sessions with five teaching assistants (TAs) who evaluated the AI’s feedback. The TAs found that AI-generated comments provided helpful praise, explanations, and questions, and could save time and improve consistency in grading. They emphasized the need for detailed rubrics and human oversight to correct errors. The findings suggest that LLM-generated feedback, used as a support tool, has promise for improving instructional feedback practices in higher education.

Authors

• Xinyi Lu, University of Michigan

• Aditya Mahesh, University of Michigan

• Mitchell Dudley, University of Michigan

• Zeija Shen, University of Michigan

• Larissa Sano, University of Michigan

• Xu Wang, University of Michigan

• Jason Godfrey, Harvard University