Before Using AI to Draft Report Card Comments, Separate Assessment from Writing

Suppose that, near the end of a term, someone at a grade-level meeting asks, “Could generative AI draft report-card comments?” Teachers have lesson reflections, student work, self-assessments, and observation notes. They want to spend less time polishing prose, but students with fewer records are also at greater risk of having gaps filled with bland, safe-sounding language.

This article’s position is clear: assign generative AI only the task of turning evaluation evidence already established by teachers into written comments. People must first decide what to assess, which records support the assessment, and how to handle insufficient evidence. The production criterion should be whether each sentence can be traced back to the original record and checked, not whether the wording sounds natural.

Separate drafting comments from making learning assessments

Version 2.0 of Japan’s Ministry of Education guidelines says generative AI in school administration should be used only within a scope where education staff can judge output appropriateness; ultimately they must review and refine the work and take responsibility for it. The guidelines also cite summarizing classroom reflections and drafting various documents as examples of use. [1]

Trials have also gone beyond writing alone. Iwakura City considered drafting report-card and student-record comments from learning histories, student self-assessments, and teacher observation records as a priority use case for generative AI in school administration. The report positions this as “comment-drafting support,” alongside classroom improvement and support for addressing school absenteeism. [2] This was a trial-stage effort; it does not mean the validity of generated comments or learning outcomes was independently proven.

The distinction to make is between assessment and drafting. The former means selecting records in light of learning objectives and judging attainment or change. The latter means shaping a confirmed judgment into language students and parents can understand. If AI is asked to do both at once, it becomes difficult to tell whether evidence is insufficient or the wording is weak.

Build an “evidence ledger” first; generate prose afterward

Before generating comments, prepare an evidence ledger for each student. At a minimum, separate the relevant learning objective, record type, date, specific facts, teacher judgment, and whether the record may be used in a comment. For example, a hypothetical entry might say, “In a September group presentation, the student checked two sources in the materials and revised the explanation.” This is easier to verify later than “participated actively in the presentation.”

Keep unused records in the ledger too. Do not turn gaps such as an extended absence, few submitted assignments, or observations made only by another teacher into a zero or a negative assessment. Treat “no record,” “assessment on hold,” and “student confirmation needed” as distinct states. Starting to collect records at year-end is too late, so align the level of detail at the start of the school year and check for gaps midway through each term.

There are three practical options. First, enter no personal information and ask a general-purpose AI only for phrasing examples. This is easy to start safely, but it is not tied to an individual student’s evidence. Second, have teachers fill in facts in fixed sentence patterns. This is easy to explain, but workload remains as combinations of records grow. Third, let AI consult an evidence ledger in a controlled environment and return both a draft and its references. This can provide more individualization, but preparing the environment alone does not make assessment judgments correct.

Therefore, use the third approach narrowly in the initial prototype. Instruct AI not to infer new assessments, not to write strengths or areas for improvement without evidence, and to attach the identifier of the referenced record to each sentence. Before editing the prose, teachers check whether the references match its meaning. After making it easier to read, they check again that the strength of the assessment has not changed.

Do not let rare records disappear in a summary

Schools have tried using AI to summarize student reflections. Fujiyamadai Junior High School in Kasugai City reported a case in which generative AI grouped and summarized reflections after social-skills training. [3] The purpose of identifying group-wide patterns must be distinguished from the purpose of deciding an individual student’s learning assessment.

A summary can highlight common opinions while diluting a rare but important change. For example, if a student who usually writes very little asks a specific question for the first time, that one original entry matters more than the class-wide summary. When collecting evidence for comments, do not return only the categories with the most records. Show separately the records that could not be classified, those whose pattern changed from before, and those teachers marked for confirmation.

Test data should include hypothetical cases not only for students with abundant records, but also for students who recently transferred, have extended absences, use multiple languages, or receive special support. Measure the rate of avoiding unsupported sentences, the rate of retaining important records, and the time teachers need to check references—not the fluency of the draft. If review time does not decrease, consider limiting generation again to a subset of standard expressions.

An environment that can handle grade data is not the same as permission to handle it

Japan’s Ministry of Education guidelines say highly important grade information and similar data must not be entered into prompts, except in individually contracted environments with appropriate safeguards. [1] A 2026 trial report describes a validation environment handling grade lists, attendance registers, student records, worksheets, and student work, as well as comment-drafting support that references student records. [4] This is a trial report by a contractor; it does not guarantee a production standard or prove effectiveness.

Boards of education and schools should not decide whether data may be entered solely by asking, “Have we removed the names?” Combining grade, class, unusual activity history, and free-text comments can let people within the school infer identity. Check who can read each class’s ledger, whether permissions remain after a staff transfer, and where generation history and output text are stored.

Japan’s Personal Information Protection Commission analyzed reports of leaks and similar incidents at schools over roughly two fiscal years, from April 2023 to April 2025, and published causes and examples of preventive measures for boards of education and school staff. [5] Comment generation should not be procured as a standalone AI feature; it should be reviewed in connection with the board’s information management, school administration systems, and device and account operations.

Pilot in a way that allows stopping midyear, not only at year-end

Align the rollout with the school calendar. Set assessment items and the evidence ledger at the start of the year, run tests with synthetic data midway through the first term, and operate in parallel in a limited number of classes before term-end. Keep comments written without AI drafts and compare missing evidence, wording bias, and review time. If use begins only just before report cards are finalized, there will be no time to return to the previous process when a problem appears.

In the 2024 school year, the Tokyo Metropolitan Board of Education designated 20 schools—nine continuing schools and 11 new ones—as generative AI research schools. Topics included AI literacy, classroom use, and staff use of AI for school administration. [6] Checking operational conditions in schools with different school types and support structures before a system-wide rollout can also inform procurement specifications.

However, do not copy a procedure that worked at a pilot school directly to other schools. Before rollout, each school should check the scope of connections from GIGA devices to administrative data, the roles of homeroom and subject teachers, who reviews within the school, and permission updates at the start of a new school year. Keep a fallback procedure that lets teachers write comments from the evidence ledger during network outages or service interruptions.

This approach is unsuitable if there is no capacity to record assessment evidence and teachers cannot check references for every sentence in an AI draft. If prose generation is introduced first under those conditions, polished wording will conceal gaps in the records. Review recordkeeping and the division of review responsibilities within the school first.

Practical implications

Recommendation 1: Give AI the writing task only after the assessment is settled. Separate learning objectives, specific facts, and teacher judgments in an evidence ledger; do not let AI infer new assessments.

Recommendation 2: Attach references to draft comments and review rare records separately. Relying only on aggregated results can miss changes in a student or clues that support is needed.

Recommendation 3: Put the school calendar and exit procedure into board procurement requirements. Test permission updates, storage locations, review time, and fallback operations during shutdowns, then decide whether to expand from limited use.

References

1. Japan Ministry of Education, “Guidelines for the Use of Generative AI in Primary and Secondary Education,” Version 2.0 (2024-12-26; checked 2026-09-24)
2. Japan Ministry of Education, “Results Report: Next-Generation School Administration Digitalization Demonstration Project (Demonstration Study on Generative AI in School Administration)” (2025-03, FY2024 project; checked 2026-09-24)
3. Leading DX School, “Generative AI Pilot School Results Conference” (2024-04-23; checked 2026-09-24)
4. Sky株式会社, “Results Report for Project III: FY2025 Demonstration Study on Generative AI for School Administration in a Secure Environment” (2026-03-17, contractor-submitted material; checked 2026-09-24)
5. Japan Personal Information Protection Commission, “Points to Note on Handling Personal Information in Light of Personal Information Leak Incidents at Schools” (2025-06-25; checked 2026-09-24)
6. Tokyo Metropolitan Board of Education, “FY2024 Generative AI Research School Project” (updated 2024-05-27; checked 2026-09-24)

This article was created with AI assistance and automatically reviewed before publication. Information checked on: 2026-09-24.