AI 批改：哪些能交出去，哪些必须留下
在批改工作里，把「誊写性的部分」和「判断性的部分」划开。
本文尚未翻译成中文，以下正文为英文。
Marking is two jobs wearing one name. Separating them is most of the answer.
The transcription job is recording what is there: this answer is missing a unit, this paragraph has no topic sentence, this proof skips a step. Repetitive, rule-shaped, and the same across thirty scripts.
The judgement job is deciding what it means and what to say: is this student stuck on the concept or on the notation, is this the third time, does this one need pushing or reassuring.
AI is genuinely useful at the first. It is not doing the second, however fluent the output sounds.
A workflow that respects the line
- Mark five scripts yourself, first. You are calibrating, and you are finding out what this cohort actually got wrong — which is rarely what you predicted.
- Write the rubric from those five. Now it describes the real distribution of errors, not an imagined one.
- Use the tool for the transcription pass. Feed it the rubric and let it flag against it. Ask for observations, not grades.
- You write the comment. The one thing the student actually reads. This takes a fraction of the time when the observations are already listed.
- Spot-check. Pull five marked scripts at random and check the flags. If the tool is wrong in a patterned way, you need to know before the class does.
What not to automate
The grade. Not because a model cannot produce a number, but because the number is the part with consequences, and the accountability stays with you regardless of what produced it.
Anything a student reads unedited. A confidently wrong correction from a machine is more damaging than no feedback, because it carries authority the student has no way to challenge.
The pattern across the class. "Nineteen of them made this mistake, so I taught it badly" is the single most valuable output of marking. That insight comes from you reading, not from a summary.
The data question, which comes first
All of the above assumes you have an answer about where student work goes. Pasting a class set of essays into a personal consumer account is not a workflow, it is a disclosure. Sort out the tooling question with whoever runs procurement before you sort out the marking question.