The University of Tokyo released a standardized homework disclosure rubric for fall 2026 enrollment, telling faculties to state on syllabi whether generative-AI assistance is prohibited, permitted with citation, or required for specific modules. The document stops short of a campus-wide ban, instead giving instructors a three-column grid to mark each assignment type before classes begin next week.
What the rubric requires
Under the rubric, students must list tools used—model name, version if known, and whether output was pasted verbatim or edited—on cover sheets for take-home problem sets and essays. Oral exams and in-person quizzes remain AI-free by default unless a department chair files an exception for language practice drills. Graduate seminars may adopt stricter rules, but undergraduate core courses must publish the grid in Japanese and English on UTAS, the university’s learning system.
UTokyo’s provost office said the policy aligns with education ministry guidance encouraging transparency rather than pretending students lack access. It differs from last year’s patchwork, when individual labs forbade ChatGPT while neighboring departments encouraged Copilot for coding drills.
Faculty implementation
Engineering faculties plan to require disclosure on computational homework but supply vetted Jupyter templates that log API calls. Law school faculty voted to keep closed-book exams entirely human-generated, citing bar-exam integrity. College of Arts and Sciences pilots will let literature courses allow AI brainstorming if students append prompt transcripts—stored locally, not sent to cloud graders.
Teaching assistants attended two-hour workshops this month on spotting undisclosed paste-ins versus legitimate collaboration. The rubric tells TAs to escalate cases to faculty chairs before honor committees meet, hoping to reduce last year’s backlog of ambiguous plagiarism hearings.
Student-facing costs
International students asked whether paid AI subscriptions count as required course materials; the rubric says departments cannot mandate paid tools without reimbursement, pushing instructors toward campus-licensed models where available. UTokyo’s library negotiated pilot tokens for Anthropic and OpenAI APIs, but caps mean large intro courses may still ban cloud tools during peak weeks.
Disability services clarified that speech-to-text and grammar aids unrelated to generative models remain protected accommodations; the rubric distinguishes assistive tech from models that author sentences.
Limits and appeals
The rubric does not govern national entrance exams or graduate admissions essays handled outside departmental classrooms. It also does not stop professors from revising rules mid-semester if new models drop, though changes must be announced two weeks before affected assignments. Students may appeal disclosure penalties through existing academic conduct boards; penalties start with resubmission for first offenses in introductory courses, escalating to course failure for repeated undisclosed use per faculty vote.
For families shipping laptops to Komaba dorms, the practical takeaway is read the syllabus grid before buying premium AI subscriptions—many courses will require disclosure but not tool use, and several science departments still prefer pencil-and-paper problem sets for anything graded.
