AI-assisted work in higher education: separate the public answer from the student record

Define the audience and the controlling source

Suppose a fictional university team is updating a public FAQ about orientation check-in. The source pack contains an approved event timetable, a current accessibility contact page and an old draft FAQ with last year's room names. The intended reader is any new student viewing the public website. The team asks AI to suggest a clearer structure and wording from the approved pack, then has a staff owner check every time, location, eligibility statement and contact detail against its controlling source. This is a synthetic example; it describes no real institution or deployment.

The distinction between a public answer and a student record is operational, not merely linguistic. A source checklist should state who approved each document, its version, whether it contains personal information, the facts it may support and what to do if sources conflict. In the example the old FAQ is useful only as a list of questions readers may ask; it cannot establish current room names. The approved timetable controls the event details. The accessibility page controls the contact route. No student names, disability requests, grades or advising notes enter the exercise.

UNESCO's guidance on generative AI in education and research emphasizes a human-centred approach, privacy and the need to validate uses in educational contexts. NIST's AI RMF Core asks organizations to understand the intended purpose and context, users and oversight roles. These are broad frameworks, not a university policy or authorization to use institutional data. The institution's current policies and responsible owners govern any real task.

Sources: [1] [2]

Make a claim-level answer sheet

For each proposed FAQ answer, record the public question, proposed sentence, controlling source, source version, checker and disposition. In the fictional pack, “Check-in opens at 08:30 in Hall B” is compared with the approved timetable. If the draft instead says Hall A because it copied the old FAQ, the owner corrects it before publication. “Students with an accessibility request should email a named coordinator” is not published if the current public page provides only a central contact route. The team uses the current public route and does not infer an individual's support needs.

The sheet also has a blank-source disposition. An AI-drafted answer may claim that every attendee receives a meal or that late arrivals can register without notice. If the approved pack does not support either statement, the team removes it and asks the responsible office for an authoritative answer. A fluent sentence is not evidence. Keep the rejected sentence and the reason for rejection in a short correction log so a reviewer can see what changed.

Verify the final page as a reader would

Run a small review using the synthetic pack. Give one reviewer the proposed FAQ and source sheet, but not the drafting conversation. Ask them to find the current check-in location, the accessibility contact and what happens when an answer is unsupported. If they cannot trace an answer to the controlling source, the sheet or the page needs repair. A second reviewer can deliberately compare the old and current timetable to see whether the version distinction remains visible. This is a usability and source-trace exercise, not a claim of educational benefit or institutional compliance.

The NIST Map playbook discusses documenting intended use, limitations and human oversight. For this example, the limitation is explicit: AI helps structure and phrase a public FAQ but does not determine policy, interpret an individual student's circumstances or send a personal response. The named staff owner signs off the public text. If the task changes to student-specific guidance, this design stops and the institution must establish a different approved process.

This page is about source authority in a public higher-education communication. It is not a generic prompt guide, a learning-outcomes study or a compliance checklist. Two Prune may use synthetic workplace materials to assess whether a participant recognizes the stale version, checks claims and communicates an evidence-backed answer. Such an assessment is bounded performance evidence for human interpretation, not a decision about a student's education or an employment outcome.

Sources: [3]