An RCT across four political science courses at Houston and partner universities found GPT-4-assisted grading of short-answer questions, using instructor rubrics, matched human grading quality without raising regrade requests, easing pressure toward multiple-choice-only tests.
271
Students across four courses
~3,080 (88%)
Short-answer responses AI-graded
Details
Maturity
Pilot
Promoter
University of Houston
Period
2023-2024 data collection; published August 2025
Keywords
higher education, generative AI, assessment, political science
Context
A team from the University of Houston, University of South Carolina, Academia Sinica, University of Michigan and Kangwon National University ran a randomized trial testing whether GPT-4-assisted grading could match instructor grading quality across four large political science courses.
Results
AI-assisted grading approximated what an instructor grading a small class would produce, without increasing regrade requests, while freeing instructor time; human review remained part of every grading pipeline tested, and the authors frame it as a productivity/consistency tool, not a replacement for instructor judgment.
Implementation
Indicative cost
Low (< €50k)
Time to results
Short (< 1 year)
Staffing & skills
Instructors across four political science courses (multi-university team)
Conditions for success
Prompt templates built from instructor rubrics and worked examples
Human review retained in every grading pipeline
Do you run this practice?
Claim it —
verified implementers get a public contact pathway and can propose corrections.
Data sources
Where this practice's information was retrieved from, and when.
TU/e researchers analysed 1,450 annotated student–chatbot interactions from 70 graded STEM essays and found a strong positive correlation between interaction …
A peer-reviewed RCT across four universities randomly assigned 3,080 short-answer responses to GPT-4 or human grading. Scores were statistically indistinguishable, …