Michigan State University will require teaching assistants to log every generative AI interaction used while grading, drafting feedback, or hosting office hours, under a pilot program that stores prompt metadata for faculty review while stopping short of banning models outright.
What the logs capture
The system records timestamps, course identifiers, and hashed prompt text—not full student essays—when TAs use approved campus tools or personal accounts linked through a single sign-on gateway. Administrators said the goal is auditability, not surveillance of keystrokes in unrelated apps.
Faculty leads in large introductory courses argued that unchecked AI drafts could homogenize feedback across sections, masking which TAs understand the material and which rely on paraphrase bots.
Why TAs are the first target
Graduate assistants grade thousands of papers in economics, biology, and writing programs where turnaround windows are tight. Provost staff said incidents last spring—identical AI-generated comments appearing in multiple sections—forced a policy response before fall enrollment peaks.
Undergraduate students rarely hold grading keys; TAs do. Logging their model use is the fastest way to trace questionable feedback without scanning every LMS submission with detector tools that false-positive on multilingual writers.
Privacy and labor tensions
The graduate employee union asked whether logs could appear in tenure dossiers or discipline files. Michigan State counsel said data retention would mirror existing email archives: available for misconduct investigations, not published to students by default.
Disability services offices requested clarity on whether TAs may use AI to simplify rubric language for accommodations; the pilot allows that use case if prompts are tagged as accessibility assistance.
How this differs from classroom bans
Some peer institutions blocked consumer chatbots entirely in STEM labs. Michigan State’s approach treats AI like calculator policies in the 1990s: permitted with disclosure. Faculty senate votes this month will decide whether logs become mandatory beyond the pilot departments.
Vendor contracts still prohibit uploading identifiable student records to public models; the logging layer adds accountability when TAs ignore those rules.
What happens next
IT teams will publish dashboards showing aggregate prompt counts by course, not individual TA names, until the pilot ends in December. If log volume spikes during midterms, deans may cap AI-assisted grading hours or require human re-reads on random samples.
For students, the near-term effect is subtle: feedback may arrive slower if TAs retype comments instead of accepting bot drafts. For administrators, the effect is measurable: a paper trail when parents complain that a teaching assistant never read their child’s essay.
Research universities across the Big Ten are watching East Lansing’s experiment; several chief information officers said they would copy the logging schema if it survives union grievance hearings this fall.
