Electronic Theses and Dissertations

Alina Zaman

Date

2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Computer Science

Committee Chair

Amy Cook

Committee Member

Allistar Windsor

Committee Member

Deepak Venugopal

Committee Member

Vinhthuy Phan

Abstract

Computer science relies heavily on graduate teaching assistants (GTAs), yet most training available to them is generic, not tailored to CS pedagogy, and rarely responsive to the needs of international GTAs. At the same time, rising enrollments and limited instructional capacity make timely, high-quality feedback on code both essential and difficult to deliver—especially in low-resource departments. This dissertation addresses that problem by pairing programmatic GTA preparation with assistive, human-in-the-loop AI and by empirically mapping where, how, and under what conditions AI can responsibly augment GTA feedback practice. First, I worked with the GTA training team to design and deploy a CS-specific GTA training program that teaches effective strategies for giving feedback on code and addresses intercultural classroom dynamics. This included Canvas modules and a GTA Orientation Week combining online and in-person sessions suited to a low-resource, high-diversity setting. Second, I analyzed GTAs’ perceptions of the training and its influence on grading practices, feedback habits, and their mindset toward teaching. Third, I worked with the research team to conduct a comparative study of prompting strategies for LLM-based feedback classification using instructor feedback from an introductory CS course. Across multiple GPT models and prompt designs (including few-shot), models underperformed on classification accuracy for three binary dimensions—prescriptive, identify-the-gap, and actionable—and their explanations were not consistently reliable. These findings argue for assistive rather than fully automated use. Finally, I studied TA–AI collaboration in practice with 13 GTAs using an AI-augmented feedback system during in-class programming exercises and post-task interviews. GTAs preferred grouped feedback when categorization appeared accurate and often edited AI-drafted feedback for clarity, tone, and context. Trust dropped sharply after noticing a grouping error, prompting manual checks or a shift toward individual feedback or hybrid workflows. The study highlights design conditions for deployable AI assistance and pinpoints that classroom-ready AI requires sufficient accuracy and transparent grouping. It also clarifies the distinctive human roles that remain pedagogical judgment, equity and tone, contextual alignment, and accountability. Collectively, these contributions offer a practical plan for training and supporting and augmenting GTAs with AI to improve feedback at scale in computer science education.

Comments

Data is provided by the student.”

Library Comment

Dissertation or thesis originally submitted to ProQuest.

Notes

Open Access

Download

Share

COinS
 

Archival Statement

This item was created or digitized prior to April 24, 2026, or is a reproduction of legacy media created before that date. It is preserved in its original, unmodified state specifically for research, reference, or historical recordkeeping. This material is part of a digital archival collection and is not utilized for current University instruction, programs, or active public communication. In accordance with the ADA Title II Final Rule, the University Libraries provides accessible versions of archival materials upon request. To request an accommodation for this item, please submit an accessibility request form.