Review learner work with AI: build a rubric and keep the final decision human

Separate observations, interpretation and decisions

An observation points to a passage in the work: “the calculation uses an unweighted mean.” Interpretation explains its effect: “the result does not answer the question about average price per unit.” The educator then decides the feedback or assessment consequence according to the published rubric.

Ask Claude to keep these stages distinct. A claim about effort, intelligence or motivation is not justified by a short submission. Keep feedback about the observable work and the next learning action.

Write a small rubric with concrete criteria

Choose three or four criteria tied to the learning objective. Each level needs a description that another reviewer can apply. Avoid vague labels such as “excellent” without explaining what evidence supports them.

Provide the task instructions with the rubric. Otherwise an assistant may penalize a feature that was never required or praise an irrelevant addition.

Calibrate with fictional examples

Prepare a strong example, a partially correct example and one with a deliberately ambiguous passage. Write your own expected observations first. Ask Claude to apply the rubric, then compare its evidence with yours.

If it interprets a criterion inconsistently, revise the wording and rerun all examples. Agreement on a few examples does not prove the system is unbiased or reliable for all learners. Calibration helps expose problems; it does not eliminate the need for review.

Handle real submissions under institutional rules

Before using real work, check your institution’s approved tools and data-sharing process. Remove unnecessary identifiers when permitted, but remember that a distinctive text can still contain personal information. Do not upload confidential material simply because the task is educational.

UNESCO’s generative-AI guidance offers a broad reference for human-centered use. Your institution’s rules and the particular assessment context still determine the appropriate workflow.

Request feedback grounded in passages

Ask for a brief strength, a priority improvement and a concrete next step. Each observation should quote or locate a passage from the learner’s work. If evidence is absent, the output should say so rather than infer an intention.

Review every comment before returning it. Remove unsupported judgments, check that the feedback matches the task and make sure it does not reveal an answer when the learner needs to complete another attempt. Keep the final grade or decision outside the automated draft.

Apply the supplied rubric to the work. For each criterion, give an observation supported by a passage, explain its relevance and propose a formative next step. Separate evidence from interpretation. Do not infer effort, personality or AI use. Do not make a final grading decision. Flag ambiguous cases for the educator.

Monitor the workflow over time

Record the kinds of comments you reject or rewrite. Repeated unsupported observations are a reason to narrow or stop the use case. Recalibrate when the rubric, assignment or product changes.

Download the human-reviewed rubric template. Use the training-session guide to align activities with the same objectives and criteria.

Sources

• →UNESCO: guidance for generative AI in education and research.
• →Claude Academy.
• →Anthropic: create and manage Projects.