instructor oversight

AI-assisted short-answer grading broadly resembled instructor grading, without testing learning gains

Tobias Heinrich, Spencer Baily, Kuan-Wu Chen, Jack DeOliveira, Sanghoon Park, Navida Chun-Han Wang

PLOS ONE

Randomized grading comparisons across political-science courses analyzed 3,080 short-answer responses from 271 students. Instructor-guided GPT-4 prompts broadly resembled human grading on selected evaluation measures, with small and context-dependent differences. Human feedback was slightly more likely to be rated helpful overall. The study evaluates grading and feedback workflows, not student learning or critical-thinking gains, and low regrade-request rates cannot establish that all errors or biases were absent.