እንዴት በ AI ጋር የውጤት አሰጣጥ ደንቦችን እና ጥያቄዎችን መፍጠር እንደሚቻል

Assessment is where AI errors cost the most, because a wrong answer key marks correct students wrong. Used carefully, though, AI can speed up two tedious jobs: writing rubric descriptions and building question banks. **Rubrics** come in two main forms. A holistic rubric gives one overall score. An analytic rubric scores separate criteria, such as thesis, evidence and organisation, at several levels, and gives students more useful feedback. The test of a good rubric is whether two graders would give the same score. 'Good use of evidence' fails that test; 'supports each claim with at least one relevant, cited source' passes. Ask the AI for observable descriptions with parallel wording, so each level differs from the next in one clear way. **Multiple-choice questions** stand or fall on their distractors, the wrong options. A good distractor tempts students who hold a real misconception but not students who understand the material. Established item-writing guidance advises you to: - avoid 'all of the above' - avoid grammatical clues that give away the answer - avoid, or use sparingly, negative wording such as 'Which is NOT' - keep the correct option from being noticeably longer than the rest Tell the AI to follow these rules and to base each distractor on a specific misconception. The main risks are wrong answer keys, especially in multi-step maths and science, and ambiguous questions with two defensible answers. Models can also show position bias, putting the correct answer in the same slot more often than chance would, so shuffle the options. A useful check is to open a fresh chat without the key and have the model answer every question. Where its answers disagree with the key, review the question by hand. A common misconception is that AI difficulty labels are measurements. They are guesses. Real difficulty comes from how students actually perform, which you only learn by trying the questions and analysing the results.