How can teachers use AI for feedback without losing the human touch?

A new school year means stacks of first assignments, and every teacher knows the trade-off: detailed feedback helps students most, but it takes the longest to write. AI chat assistants can draft comments quickly, and that is tempting in September. The risk is just as clear. Feedback that sounds generic, or that a student suspects came from a machine, loses the very thing that makes it work: the sense that a person read their work and cares how they improve.

The approach that keeps both is a division of labor. The AI handles the mechanical part, matching a piece of writing to the criteria you already use. You handle the human part, and you always have the last word.

Start from your rubric, not from the AI

A rubric is the scoring guide that lists what you are looking for and what each level of quality looks like. It is the best possible input for an AI, because it tells the model exactly what to comment on. UNESCO's quick-start guide for higher education already lists rubrics among the things teachers can ask ChatGPT to help design, and the same rubric can then steer feedback.

Before you paste anything, remove the student's name, ID number and any personal details in the text itself, such as a family story in a personal essay. Use "Student 7" instead. More on why below.

Worked example: a rubric-based first pass

Here is a prompt for a Grade 9 persuasive paragraph. Adapt the criteria to your own rubric.

You are helping a Grade 9 English teacher draft feedback. The teacher will edit
everything you write before the student sees it.

Rubric (each criterion scored Beginning / Developing / Proficient):
1. Clear claim in the first sentence.
2. At least two reasons, each supported by an example.
3. Addresses one counter-argument.
4. Sentences are complete and correctly punctuated.

For the paragraph below:
- For each criterion, quote the sentence that best shows the current level,
  and suggest a level with a one-line reason.
- Then write 2 strengths and 1 next step, addressed to the student,
  in plain, encouraging language, under 80 words in total.
- Do not assign an overall grade.

Student 7's paragraph:
[paste anonymized text here]

Three details matter. Asking the model to quote the sentence behind each judgment lets you check it in seconds. Limiting the student-facing comment to 80 words keeps it readable for a 14-year-old. And "do not assign an overall grade" keeps the grading decision where it belongs.

The teacher's edit

Suppose the draft comment reads (illustrative):

"Your claim is clear and your first reason is well supported. Your second reason needs an example. Next step: add a sentence that responds to someone who disagrees."

It is accurate and fine. Now the teacher edits it (again an illustration):

"Maya, your opening sentence makes your position impossible to miss, a big improvement from last week. Your point about the bus schedule is convincing because you used your own experience. For your second reason, find one example as strong as that one. Then try this: imagine your cousin who disagrees, and answer her in one sentence."

The edits are small, but each one does something an AI cannot. The name and the reference to last week show the student that you remember them. Praising the bus example connects to what you know about their life. The cousin suggestion fits this particular student. OpenAI's own guide for teachers, published last year, reminds educators that the model is not always right and that its prompts are only a starting point; your edit is where that check happens.

Privacy rules for student work

Student writing is personal information, and often more personal than it looks. Canada's federal and provincial privacy commissioners, in principles published in December 2023, ask organizations using generative AI to use anonymized or de-identified information in prompts where possible, to enter sensitive personal information only when authorized, and to take particular care with children and in high-impact settings such as education.

For a classroom, that becomes a short checklist:

• Use only tools your school or board has approved for student work, and follow its policy.
• Remove names and identifying details before pasting, every time.
• Leave out sensitive content entirely. A disclosure about health, family or safety goes to the appropriate person at school, never into a chatbot.
• Don't let the AI decide grades or anything else that affects a student's record.
• Tell students and families how you use AI in feedback. Transparency builds trust, and it models the honesty you ask of students.

What should a teacher always write personally?

Anything that depends on knowing the student. That includes encouragement tied to their progress over time ("you've come a long way since September"), comments that connect the work to their interests or experience, and any response to something personal they shared. A model has never met your students; you have.

Write the hard messages yourself, too. When work shows a serious problem, a possible integrity issue, or signs that a student is struggling outside class, the words need care, context and often a conversation rather than a comment. Final grades and report card comments also belong to you, because they carry your professional judgment and you must be able to explain them to a student or parent.

A simple test helps: if the student read the comment aloud to you, would you be comfortable saying "I wrote that"? If yes, the AI helped you write faster. If not, rewrite it until you would.