Towards Transparent AI Grading: Semantic Entropy as a Signal for Human-AI Disagreement

AI grading systems need better warning signals: The research demonstrates that automated grading tools often fail to flag when they're uncertain or when human graders might disagree, creating a transparency problem. Semantic entropy addresses this by measuring how varied the AI's explanations are for the same student answer—more variation suggests the grading decision is less reliable.

The method works across different subjects: Experiments on the ASAP-SAS dataset show that semantic entropy correlates with human disagreement and "varies meaningfully across subjects," indicating the approach isn't limited to one academic domain. This suggests the technique could be broadly applicable to AI grading systems in schools and universities.

Interpretive tasks are harder for AI to grade confidently: The research found that semantic entropy "increases in tasks requiring interpretive reasoning," meaning the AI system shows more uncertainty (and generates more varied explanations) when grading subjective or open-ended questions. This aligns with intuition—these are exactly the cases where human graders also tend to disagree most.

Transparency can improve trust in AI grading: By making AI uncertainty visible through semantic entropy, educators and students can better understand why the system made a grading decision and when human review might be warranted. This supports more responsible deployment of AI in educational settings where fairness and explainability matter.