Your Professor Uses AI. You Can’t. Where Should Colleges Draw the Line?

Ella Stapleton wanted her money back. The Northeastern University senior had found signs of AI use in materials for Rick Arrowood’s organizational behavior course, whose syllabus prohibited students from using AI. As Kashmir Hill reported in The New York Times in May 2025, Stapleton sought a refund of roughly $8,000. Northeastern denied it. Arrowood acknowledged using AI and said he wished he had reviewed the materials more closely.

Students’ irritation is understandable. Their chatbot use can become an academic integrity case; an instructor’s becomes an efficiency measure. From the student’s seat, the distinction looks convenient, especially when the course materials contain mistakes.

AI crosses the line when it takes over the intellectual responsibility a person is supposed to carry. Students must do the thinking an assignment exists to develop. Faculty must exercise the expertise and judgment their teaching promises. Different responsibilities justify different permissions; both sides owe an honest account of the assistance they receive.

Colleges can make that principle usable through three tiers:

1. Formatting: AI changes presentation without supplying ideas or judgments. Establish a clear baseline policy; routine formatting needs little individual disclosure.
2. Substance: AI contributes an argument, explanation, lesson, or interpretation. Disclose the contribution and identify the thinking the person still performed.
3. Evaluation: AI influences feedback, grades, or misconduct decisions. Require advance notice, a defensible process, and a qualified human who can explain and review the outcome.

These tiers describe consequences, not product names. One chatbot can perform all three kinds of work. Obligations rise with the responsibility delegated.

The asymmetry that isn’t hypocrisy

An introductory writing assignment might ask students to choose evidence and build an argument. Letting AI supply the thesis and paragraphs can remove the exercise itself. A professor using AI to suggest discussion questions, then selecting and rewriting them after reading the text, is performing a different task.

But faculty status supplies no blanket exemption. A professor who substitutes an AI summary for the reading and cannot discuss the text has also handed away essential work. Checking facts does not restore the missing engagement on either side.

The strongest case for faculty assistance is practical. Imagine an adjunct teaching five sections or a teaching assistant facing 150 papers. Reducing routine preparation could create more one-on-one time and better feedback. That is a serious educational benefit, provided the time saved reaches students. Institutions also need to address the workload that makes shortcuts attractive.

Campuses are already drawing these distinctions. An August 1, 2026, memo from Jenny Trinitapoli, master of the University of Chicago’s Social Sciences Collegiate Division, sets expectations for its Social Sciences Core. Alongside independent student work, it promises “syllabi designed by human scholars at the University of Chicago.” It generally excludes AI-assisted grading unless instructors are carefully assessing its validity against human grading, with experimentation disclosed in the syllabus. The memo, published by the Chicago Maroon, makes faculty responsibility part of the bargain.

The policy covers the Social Sciences Core, not the university. UChicago’s law school separately announced restrictions on laptops and phones in first-year classes in July 2026. Meanwhile, the university’s June 2 announcement offered Claude Enterprise across campus, with classroom rules left to vary. Access to a tool and permission to use it on every task are distinct decisions.

What students owe: instructions, not slogans

The three tiers help only if instructors connect them to the learning objective. “Use AI responsibly” leaves students guessing. Even “brainstorming is allowed” is slippery on assignments where deciding what to argue is the hardest part.

Each assignment needs to identify what students must produce independently, what assistance is permitted, and what to disclose. In introductory programming, generating a loop may bypass the skill being assessed. In an advanced project, the same assistance may be acceptable when assessment centers on architecture, testing, and explaining design decisions.

Tier one also needs context. A grammar correction may be incidental in a history paper but central to a language exercise. Reorganizing an essay can change its argument. Calling a feature “editing” does not settle which tier it belongs in.

There is evidence that well-designed tutors help: a 2025 Scientific Reports study led by Greg Kestin, involving 194 Harvard physics students and a custom tutor built on GPT-4, found stronger immediate learning outcomes than the comparison classroom lessons. It covered two lessons, so it supports testing carefully designed tutoring rather than assuming general chatbots produce lasting learning gains.

For permitted substantive assistance, a short account can identify the tool, its task, and what the student accepted or changed. Disclosure cannot authorize a prohibited use: acknowledging that AI wrote an independent essay does not complete the assignment as intended. Nor does accountability automatically require surrendering unrelated chat histories.

What faculty owe: disclosure before detection

Faculty need to apply the same tiers to their own work. Reformatting an announcement warrants little attention. Generating a case study or a lecture’s explanation warrants a clear account of what AI supplied and how the instructor reviewed it.

Cornell’s teaching guidance recommends acknowledging AI assistance in course preparation. A useful disclosure goes beyond naming the software: students need to understand where the instructor’s knowledge and decisions enter the work.

Here is model syllabus language for a writing course that permits limited assistance:

You may use AI for practice questions and copyediting, but your submitted arguments, evidence selection, and drafts must be your own; attach a brief account of any AI assistance. I may use AI to help prepare teaching materials, and I will identify substantive contributions and review their accuracy and suitability. I will read and assess your work myself, will not use AI to assign grades, and will not upload your submissions to an AI service without an institution-approved process explained to you in advance.

That statement makes commitments students can understand and question. A course permitting AI-generated drafts would need different language, including what students must demonstrate themselves. An instructor using AI in feedback would need to describe that workflow explicitly.

Disclosure is also a better starting point than mutual surveillance. In August 2023, Vanderbilt disabled Turnitin’s AI detector, citing reliability and transparency concerns. A detector result cannot carry the burden of an academic misconduct finding by itself. Students accusing faculty need evidence, too: polished prose, repetitive phrasing, or an odd slide alone cannot establish how something was created.

Transparency does not excuse poor teaching. An instructor who announces that a lecture was generated but cannot answer questions about it has disclosed the problem without meeting the obligation.

The hardest line: AI in grading

Tier three demands more because an error can affect a student’s academic record. There is a meaningful difference between turning an instructor’s own observations into clearer comments and asking a model to read, judge, and score an essay.

The first use still requires checking that the comments express the instructor’s judgment. The second delegates judgment itself. A cursory glance at the output does not make that delegation accountable.

For substantive written work, human review needs to include reading the submission, applying the published criteria, and checking the rationale. An AI summary alone is a weak basis for judging an unconventional argument. The instructor must be able to explain why a passage lost credit and reconsider the decision when challenged.

Where institutional policy permits AI-assisted evaluation, students need advance notice of its role and a route to substantive human review. Colleges need evidence that the method works for the assessment in question. The acceptability of machine-scored quizzes does not establish the validity of automated literary criticism. UChicago’s restrictions on AI-assisted grading put that evidentiary burden on faculty experimenting with it.

The data question begins before an upload. At covered U.S. institutions, FERPA generally requires consent before disclosure of personally identifiable information from education records unless an exception applies. A vendor qualifying under the school-official exception must meet conditions including institutional control over use and maintenance of records and restrictions on reuse and redisclosure. Faculty cannot assume a personal chatbot account meets those requirements. Institutions need to approve the specific workflow, including access, retention, and permitted uses of student work.

What the institution owes both

Colleges need to make these obligations possible: clear assignment policies, approved tools, training, realistic workloads, and fair access when a course requires AI. An enterprise subscription alone does not provide any of those guarantees.

They also need to make learning visible without turning every student into a suspect. The Associated Press reported in March 2026 on professors reviving oral defenses to assess understanding. Draft discussions and explanations of source choices can serve a similar purpose. These assessments need advance notice and appropriate accommodations; hesitation can reflect anxiety, language differences, or disability.

Students need a practical route when faculty fall short. Preserve the syllabus and disputed materials, ask what AI contributed, and request an explanation or human review. If that fails, ask the department chair or dean which teaching-complaint or grade-appeal process applies and what deadlines govern it. An ombuds office, where available, can help identify options: the University of Michigan’s student ombuds office explicitly helps students assess campus disputes and plan next steps. The applicable process will depend on the institution and the complaint; none guarantees the refund Stapleton sought.

Faculty value needs a better measure than a count of the words a professor personally typed. Selection, explanation, attention, and judgment matter. If AI saves time, students ought to see the benefit in better teaching and access to their instructors. Institutions must be able to show where that benefit goes.

The obligation running through all of this is reciprocal. A student owes the instructor an honest demonstration of learning. The instructor owes the student informed teaching and accountable evaluation. AI can assist either side, but permission to use it depends on whether those obligations are still being met.

A university can reasonably ask a student to write an essay without AI. It should be just as comfortable telling that student who will read it, who will judge it, and how that person will stand behind the grade.