Should AI be prohibited? Should students disclose every prompt? Should the syllabus distinguish between brainstorming, editing, and drafting?
Those questions matter, but they do not solve the central problem.
Students can ignore a prohibition. They can underreport their use. They can use AI in ways that fall between policy categories. And even when AI is allowed, they can rely on it so heavily that the intended learning never occurs.
Why Bans Are Difficult to Enforce
A rule such as "Do not use generative AI" changes the instructions around an assignment, but it does not change the assignment itself.
Students still complete the same essay, quiz, report, or project under largely unobserved conditions.
Researchers call this a discursive change: a change that operates through language and depends on student awareness, interpretation, and voluntary compliance.
The difficulty is not simply that students are dishonest. In fact, most students using AI are responding rationally to their incentives.
Even students trying to comply may struggle to distinguish between:
- AI-assisted editing and rewriting,
- brainstorming and idea generation,
- outlining and drafting,
- feedback and substantive contribution.
In one study of AI-use declarations, students described these boundaries as confusing and difficult to track across repeated interactions with AI.
A policy can clarify expectations. It cannot make an unchanged take-home assignment secure.
Ask What Students Must Still Be Able to Do
Before rewriting the AI section of your syllabus, identify the learning outcome.
What must the student be able to do without delegating the core judgment to a machine?
For an essay, it may be: construct an argument, select credible evidence, address counterarguments, or explain why one interpretation is stronger than another.
For a spreadsheet, it may be: choose the right model, interpret the result, identify an unreasonable assumption, or explain the decision the analysis supports.
For code, it may be: diagnose an error, justify an architecture, explain a tradeoff, or modify the program under a new constraint.
Once you know what must remain with the student, you can permit AI around that boundary without surrendering the assessment.
Distinguish AI Use From AI Over-Reliance
Students can use AI appropriately and still learn.
They can also follow every stated rule while allowing AI to replace the cognitive work the assignment was meant to develop.
That is over-reliance.
Using AI to think
The student uses AI to generate possibilities, question assumptions, or improve an idea they can independently evaluate.
Using AI instead of thinking
The student accepts AI-generated reasoning without developing or understanding it.
The challenge for assessment is to make that difference visible.
Five Changes You Can Make This Week
1. Grade decisions, not just outputs
Ask students to explain two or three consequential decisions they made.
For example: Why did you choose this source? Why did you reject another approach? What assumption had the greatest effect on your answer? Which part of the work required the most judgment?
These questions focus attention on the intellectual work behind the artifact — because the artifact was never the point.
2. Require a brief AI-use reflection
Avoid a simple checkbox asking whether AI was used.
Instead, ask: What did you ask AI to do? What output did you reject? What did you verify independently? Which final decisions remained yours?
Students in declaration studies expressed interest in explaining how AI fit into their workflow rather than merely selecting yes or no.
A reflection is still self-reported, so it should not be treated as proof. Its value is in making students examine their own use.
3. Add a short oral follow-up
Ask each student two or three questions based on the submitted work.
One question should test understanding. One should introduce a variation. One should ask the student to evaluate or defend a decision.
"Your recommendation depends on demand increasing next year. What happens to your conclusion if demand remains flat?"
This is much more revealing than asking the student to repeat the final answer.
4. Make students evaluate AI output
Provide an AI-generated response containing a subtle weakness.
Ask students to identify unsupported claims, questionable assumptions, missing context, or an inappropriate recommendation.
Students need practice judging AI, not merely operating it.
5. Build multiple checkpoints
Instead of assigning one large project worth 40% of the grade, create several smaller moments: proposal, preliminary reasoning, draft decision, feedback response, final defense.
The purpose is not to create more grading. It is to collect enough evidence that one polished final artifact cannot carry the entire claim of learning.
Use Policies to Support Assessment, Not Replace It
A useful AI policy should clearly state: what AI use is permitted, what students remain responsible for, what they may be asked to explain, and how their understanding will be assessed.
"You may use generative AI to brainstorm, receive feedback, and improve presentation. You remain responsible for the accuracy, reasoning, and final decisions in your submission. You may be asked to explain your process and defend any part of your work."
That policy is useful because it is connected to an assessment structure.
Without the follow-up, it remains a request for compliance.
Do Not Make Students Choose Between Honesty and Grades
Research on AI declarations found that students often fear that disclosure will draw suspicion, lower their grade, or function as an admission of wrongdoing. In some cases, non-disclosure becomes a strategic decision rather than an oversight.
If AI is permitted, students need confidence that honest disclosure will not automatically be punished.
The tone should be:
"Show us how you used the tool and what intellectual responsibility remained yours."
Not:
"Confess whether a machine touched your work."
The Practical Goal
You may not be able to stop every student from opening ChatGPT.
You can make it much harder for students to earn credit without understanding what they submit.
That is the more durable response.
An oral assessment platform can help at scale, but the underlying principle does not require software:
Preserve the assignment. Add evidence of process, judgment, and personal understanding.