← Muhammad Hasheem Iftikhar
ACM FAccT 2027 · Under review

When AI Work Becomes Official

Locating Authority in AI-Assisted Coursework

M. U. Saleem, W. Naveed, M. H. Iftikhar, M. A. Ijaz, H. Munir

Percevio Labs, with co-authors from Amazon, Duke University, Northwestern Polytechnical University and Harbin Institute of Technology

Advisors Dr. Muhammad Usama Saleem, Research Scientist, Amazon Dr. Ayman Ali, Assistant Professor, Riyadh College of Technology for Information and Telecommunications
Paper Live system Homepage
The Coursework Authorization Model as instantiated in TRACE
The Coursework Authorization Model in TRACE. Available course information becomes authorized evidence through source authorization; provisional AI work becomes an adopted course artifact through artifact authorization; and consequence authorization governs when that work may affect a student. Authorization provenance links the steps so decisions can be inspected and reconsidered.
67instructors surveyed
94students surveyed
13in-depth interviews
3countries: Saudi Arabia, US, Pakistan

Abstract

Human oversight is often proposed as a safeguard in AI-supported decisions, but having a person involved does not always clarify what they are responsible for. In higher education, generative AI can draft course materials, suggest grades and provide feedback, raising questions about when these outputs should become part of a course or affect students.

We report a formative mixed-methods study combining surveys of 67 instructors and 94 students with interviews with six instructors and seven students. Willingness to delegate work to AI did not mean willingness to give up control over academic decisions. Among instructors, 43 of 67 wanted the ability to review and edit AI outputs, and 41 wanted final academic decisions to remain with instructors. Students rated instructor-reviewed grading more favorably than AI-only grading, yet 41 of the 64 who considered reviewed grading fair or partly fair still wanted additional safeguards. Participants' concerns went beyond correctness to whether the evidence was appropriate, who was responsible, and whether a decision could be reconsidered.

We synthesize these findings into the Coursework Authorization Model, which distinguishes what information AI may use as evidence, which AI-produced outputs become part of a course, and when those outputs may affect students, and illustrate it through TRACE, a course-grounded prototype that makes authorization decisions explicit and inspectable.

The Coursework Authorization Model

Three decisions, not one "human in the loop"

01 · Source

What counts as evidence

Which available course materials the AI may rely on for a given task.

02 · Artifact

What becomes official

Which AI-produced version is adopted as part of the course.

03 · Consequence

When it affects students

When adopted work may be released or applied in a way that affects a student.

Meaningful oversight depends not simply on reviewing AI outputs, but on clarifying what is being authorized, who remains responsible, and how consequential decisions can be reconsidered.

Findings

People delegate work to AI, not authority over it

Instructor delegation comfort and requirements for using AI work
Delegating work without delegating authority. Instructors' comfort with delegation varied across six teaching activities, but requirements for using AI-produced work stayed present in every group, including instructors comfortable delegating all six.
Student acceptance of AI roles and fairness of AI grading
From drafting to judging. Students accepted AI drafting quiz questions and feedback far more readily than AI selecting readings or suggesting scores, and rated instructor-reviewed grading above AI-only grading.
Where each interview participant drew the line on AI involvement
Where each participant drew the line. Students separated production, source selection and grading; in a probe where a grade looked correct but rested on inappropriate material, five of six students still cared about the evidence behind it.
What students require beyond instructor review
What students need beyond the grade. Most students who accepted reviewed grading still asked for more safeguards, and they expected recourse through the instructor rather than formal appeals.

The prototype

TRACE makes every authorization explicit

TRACE is a course-grounded assistant in which each step from course material to student consequence is a separate, recorded decision. It is now publicly deployed as Fazah.

Course corpus
1. Course corpus. Materials are added to the workspace. Being available does not authorize a file for every use.
Source authorization
2. Source authorization. Before generation, the instructor chooses which materials may serve as evidence, and can record why a source is excluded.
Provisional AI output
3. Provisional output. Generated questions stay provisional until reviewed; publishing to students is unavailable.
Artifact authorization
4. Artifact authorization. Accepted items become the adopted version. Publishing stays a separate action.
Per-item provenance
5. Per-item provenance. Every question can be checked against the authorized source and the passage that supports it.
Authorization history
6. Authorization history. The record of how AI-supported work changed status. Generation is an event, not an authorization.

My role

What I contributed

  1. Built TRACE

    Developed the course-grounded prototype at the center of the paper, now deployed publicly as Fazah.

  2. Designed the study

    Designed the interview protocol and led student-side data collection with bilingual instruments across three countries.

  3. Co-developed the model and paper

    Co-developed the Coursework Authorization Model and co-wrote the manuscript.

More research

ACL submission · in preparationEduMem Thesis · MICCAI 2027 plannedEnhancing mAIstro The Bridge of HopesAssistive AI for Children with ADHD