AI-assisted grading: what it can and cannot do

A clear guide to automated assessment, evidence, human review and the boundaries of judgment.

Start with the rubric

The system can apply explicit criteria more consistently than vague ideas such as originality or depth.

Every suggested outcome should point to the relevant part of the submission so a reviewer can verify it.

Borderline, ambiguous and high-impact cases must move to an authorized reviewer.

Test representative work, measure agreement and appeals, then expand only after an approved result.