Generate assessments and feedback

Aligned assessments and faster grading

Creating meaningful assessments and grading student work are two essential but extremely time-consuming tasks, and the workload always peaks at the worst times of the year. ChatGPT supports both ends of this process: it helps you build rigorous assessments and speeds up grading, without ever replacing you. This distinction matters because delegating grading would mean giving up the one moment when you see exactly what each student has understood. It is also a rule: the CNIL (France’s data protection authority) reminds teachers that they cannot delegate their evaluative authority to an AI tool, and OpenAI’s usage policies prohibit any evaluative decision made by the model without a human in the loop.

Design multiple-choice questions that teach

A good multiple-choice question (MCQ) is not just one correct answer and three wrong ones. Its diagnostic value lies entirely in its distractors: if they are absurd, students find the answer by elimination and you learn nothing about their reasoning; if they reflect realistic errors, each wrong answer tells you which mistaken path the student took. ChatGPT easily generates this second type of distractor, provided you name the expected errors instead of hoping it guesses them.

“Create a 15-question MCQ on linear functions for a 9th-grade class. For each question, the distractors must correspond to frequent student errors (confusion with affine functions, sign errors, inverted slope). Provide an answer key with an explanation for each distractor.”

The explanation requested for each distractor is what turns the MCQ into a remediation tool. When you review answers in class, you already have the comment ready for students who chose B instead of C, and the review session stops being a reading of the answer key and becomes work on the errors students actually made.

Assess by skill and vary formats

For formative or summative skills-based assessment, four requirements must appear in your prompt. Specify the target skills using the exact terms from the program’s framework so the rubric vocabulary matches the report card. Ask that each exercise target only one skill at a time: as soon as an exercise mobilizes two, a weak student’s work becomes impossible to interpret. Require an associated skills rubric with observable indicators rather than value judgments. Finally, request exercises of increasing complexity for each skill, allowing a struggling student to show what they master before hitting a wall.

Written prompts follow the same logic. ChatGPT generates argumentation prompts with accompanying documents, literary commentary prompts with suggested reading angles, contextualized case studies for economics and social sciences, or open-ended problems for science and math. The suggested reading angles deserve careful review: they are generally accurate but school-like, and it is better to keep them for your answer key than to distribute them with the prompt, or you risk getting thirty identical papers.

The criteria-based rubric

A useful rubric describes each level of mastery with an observable descriptor, so that two different graders place the same paper in the same spot. Ask ChatGPT to produce this rubric for your criteria, then verify that the descriptors describe what you actually see on the paper, not the impression it leaves.

Observe the progression in the “Argumentation” row: it moves from absence to presence, then from presence to justification, then from justification to nuance. Each step adds a verifiable requirement. This is the gradation you must check in any generated rubric, because the model sometimes produces four formulations that say the same thing while only changing the adverb—“correctly,” “well,” “very well,” “perfectly”—and such a rubric distinguishes nothing.

Speed up grading

The clearest time savings come from a model answer key produced before you open the first student paper. Have it generate the expected answer for each question, frequent errors to anticipate, detailed scoring points, and sample comments for each response level. You then grade by recognizing cases you have already named rather than reinventing your standards for every paper, which improves both speed and fairness between students at the top and bottom of the stack.

For individual feedback, ChatGPT writes a constructive comment as soon as you provide the observed error. “The student made this error in their physics exercise: they confused mass and weight. Write an encouraging 3–4 line feedback that explains the difference and suggests a mnemonic to avoid confusing them again.” The division of labor is right: you identify the error, since only you have read the paper, and the tool handles the wording—the part that costs you time without requiring your subject-matter expertise. The prompt contains no names or copies: never paste a named student paper into a public tool.

These comments are worth saving rather than rewriting every year. Gradually build a bank covering each type of recurring error in your subject, with positive phrasing, concrete improvement suggestions, and encouragement adapted to different profiles. By the third year, grading a stack largely consists of assembling proven blocks that you personalize with a single sentence.

Your turn

Your first reviewed MCQ.

1. Have ChatGPT create a ten-question MCQ on a theme of your choice, naming the expected errors
2. Check each item: is the correct answer accurate, are the distractors plausible, is the level appropriate?
3. Fix what needs fixing and count your corrections
4. Have it produce the criteria-based rubric and check the gradation of descriptors
5. Note your number of corrections: it indicates how much confidence to place in an unreviewed prompt

Official sources: Getting started with ChatGPT as an educator (OpenAI Help Center) · Teacher: How to use an AI system in the course of your duties? (CNIL)

Key takeaways

• MCQ distractors must reflect realistic errors that you name in the prompt
• Create criteria-based rubrics with precise descriptors for each mastery level
• A detailed model answer key before grading significantly speeds up the process
• Identifying errors and assigning grades remain your responsibility; the tool can handle the wording
• Build a bank of reusable comments rather than writing every feedback from scratch