Providing Feedback with AI Support

AI can help an educator draft, organise, or vary feedback. It should not become the assessor. Effective feedback remains grounded in the student’s actual work, the intended learning outcomes, and professional judgment.

What useful feedback should do

Strong feedback helps a learner answer three questions:

• Where am I going? Restate the relevant goal or criterion.
• How am I doing? Point to specific evidence in the work.
• What should I do next? Give a feasible action the student can apply.

Feedback is more useful when it addresses the task and the learner’s process, not presumed personality or ability. It should prioritise a small number of changes rather than overwhelm the student with every possible correction. Research syntheses show that feedback effects vary substantially with its content and delivery; more feedback is not automatically better.^{1 2}

A safe drafting workflow

Copyable prompt for a feedback draft

Use this only in an environment approved for the material you provide.

You are helping me draft formative feedback. I remain the assessor and will
review every comment.

<assignment>
[assignment instructions]
</assignment>

<learning_outcomes_and_criteria>
[relevant learning outcomes and rubric criteria]
</learning_outcomes_and_criteria>

<student_work>
[authorised student work or excerpt]
</student_work>

For each of the two highest-priority points:
1. name the relevant criterion;
2. identify the exact passage or feature in the work that supports the comment;
3. describe one strength or successful choice, where supported;
4. propose one specific next action; and
5. offer one question that could help the student revise independently.

Do not assign a grade. Do not infer motivation, effort, personal characteristics,
misconduct, or AI authorship. Do not introduce requirements absent from the
assignment or criteria. If the evidence is insufficient, write "insufficient
evidence" and explain what I should inspect. Return a concise table.

Worked review: correct a misleading feedback draft

Use the fictional library evidence pack in the essay assessment example. The following synthetic submission excerpt is for educator practice:

Sixty of the eighty survey respondents wanted later opening. Therefore most students want later hours, and extending opening will improve learning. The €8,000 cost is justified by this demand.

Intentionally flawed AI-style feedback: “Excellent representative evidence and a convincing causal argument. Add two more references and use a more professional tone. You clearly worked hard.”

The comment praises an unsupported population inference and an unmeasured learning effect. It introduces reference and tone requirements absent from the task, and infers effort. Review must catch false praise and missed problems as well as false criticism.

Reviewed educator feedback: “Your opening sentence accurately describes the survey respondents. Under the evidence-interpretation criterion, the next sentence extends that finding to all students, although L2 reports an open invitation rather than a random sample. Revise that inference and state what you can conclude. L3 also reports no measured learning effect. What evidence would you need before claiming that later hours improve learning?”

The student could revise the passage to distinguish expressed preferences from population demand and treat a learning benefit as unestablished. Ask them to justify the revision, then interpret a fresh survey independently. Do not assume the improved paragraph alone demonstrates transfer. For how two graders could interpret this excerpt differently, see the rubric calibration example.

Human review checklist

Before sending the feedback, confirm:

• every observation points to evidence in this student’s work;
• the advice follows the published task, learning outcomes, and rubric;
• any factual or disciplinary claim has been independently verified;
• the tone is clear, respectful, and appropriate for the learner;
• the priorities are achievable within the time and support available;
• the feedback does not infer protected characteristics, intent, or authorship;
• the grade, if any, was determined and checked by the responsible educator;
• the student can ask for clarification or challenge a factual misreading; and
• the feedback creates a chance to act, revise, or transfer learning.

Evaluate the workflow

Test the process on a small, representative sample before wider use. Compare the AI-supported draft with educator-only feedback and inspect both false criticisms and missed issues. Check whether quality differs by language variety, topic, or student group. Record the time saved after review—not just draft-generation time—and stop if the workflow weakens feedback quality or fairness. Use the evaluation guide to record the comparison and decision.

Early studies of LLM-generated feedback show potential in particular settings, but this is still a developing evidence base and results do not transfer automatically across subjects, tools, or student groups.⁴

References & Footnotes