AI Question Generation, Marking and Instant Feedback: Practice and Limits in Maths, Chinese and English Writing

Short answer: the right role for AI marking is "first pass + specific feedback + data", while teachers keep "correction, follow-up, decisions". Step-by-step maths marking, rubric-based Chinese feedback and English grammar/structure suggestions each have their own method and limits.

The three pain points teachers most want solved

1. Volume: 30 students, two assignments a week — 6–8 hours of marking weekly.
2. Latency: students get work back three days later, after the error has been forgotten.
3. Vagueness: "watch your structure" gives the student nothing to act on.

AI can address all three — provided the rubric is clear.

Scenario 1: step-by-step maths marking (with handwritten photos)

Flow: teacher creates the assignment (questions + model answers + key steps) → students type or photograph answers → AI recognises and compares step by step → flags the wrong step and the reason → teacher corrects in one click.

Feedback example: "Step 2: you added the denominators of 3/4 and 1/8 directly to get 12. When denominators differ, find a common denominator first. Try rewriting 3/4 with denominator 8."

Limits: messy handwriting, skipped steps and non-standard notation reduce accuracy. Ask students to write "one step per line" and let teachers edit at any time.

Scenario 2: Chinese composition feedback against a rubric

Rubric template (P5 narrative, 20 marks):

The AI grades against this rubric and, for each dimension, gives a score and a specific suggestion quoting the original (e.g. "In paragraph 3, 'I was so happy' could be shown through action or dialogue instead").

Scenario 3: English writing grammar and structure

The AI labels error types (tense, subject–verb agreement, articles), suggests sentence-level rewrites, and checks paragraph structure (topic, supporting and concluding sentences). Key: teachers can set a "flag only, don't rewrite" mode so students fix their own work.

Three checkpoints teachers must keep

1. Correction: every AI first pass is accepted or edited in one click; the system learns the teacher's preferences.
2. Sampling: teachers fully mark 10% of each assignment and compare with the AI.
3. Follow-up: use the multi-dimensional analysis (which dimension is weakest class-wide) to decide what to teach next, rather than just returning scores.

Data is worth more than scores

After an assignment, teachers should see mastery per learning objective, a ranking of common error types, a list of students needing follow-up, and AI-suggested follow-up exercises. Accumulated, this is the evidence for grant reports and parent meetings.

Rollout suggestions

• Start with one subject and one question type (e.g. P4 maths calculation), then widen after two weeks.
• Have the panel write the rubric together first; the AI marks only as well as the rubric is written.
• Tell students: the AI is the "first reader"; the teacher is the "final marker".

Takeaway

AI marking is not handing the red pen to a machine. It turns "wait three days" into "instantly" and "watch your structure" into "paragraph 3 could be changed like this". Teachers keep correction, sampling and follow-up; students get faster, more specific feedback; teachers get time and data.