Investigating Sampling Impacts on an LLM-Based AI Scoring Approach: Prediction Accuracy and Fairness

AI scoring capabilities are commonly implemented in educational assessments as a supplement or replacement to human scoring, with significant interest in leveraging large language models for scoring. In order to use AI scoring capability responsibly, the AI scores should be accurate and fair. In this study, we explored one approach to potentially mitigate bias in AI scoring by using equal-allocation stratified sampling for AI model training. The data set included 13 open-ended short-response items in a K-12 state science assessment. Empirical results suggested that stratification did not improve or worsen fairness evaluations on the AI models. BERT based AI scoring models resulting from the stratified sampling method but trained on much less data performed comparably to models resulting from simple random sampling in terms of overall prediction accuracy and fairness on the subgroup level. Limitations and future research are also discussed.