AI essay graders systematically mark grammar too harshly—and small bias tests can catch it

Researchers tested large language models on real essay-scoring datasets and found they consistently give lower scores for grammar and basic writing mechanics than human raters do, even when they score overall essay quality reasonably well. This matters because schools considering AI essay graders now have evidence of a specific, measurable bias—and a practical fix: test the AI on a small sample of human-scored essays to measure and correct the bias before deployment, without needing expensive retraining.

The interesting part here is not that AI is biased—that's expected. It's that the bias is directional, stable, and detectable with a tiny validation set. A school could literally test an AI on 50 hand-scored essays, measure exactly how much it over-penalizes grammar, and then adjust all its scores accordingly. That's cheap and practical enough to actually work. The finding also reveals something about how AI scoring works: it struggles more with rule-following (grammar) than judgment (essay quality), which is the opposite of what you might expect. Teachers considering AI grading now have a usable playbook instead of a binary choice between trust or reject.