Students should know whether they may use AI for brainstorming, drafting, editing, coding, analysis, or feedback. Faculty should not improvise expectations after work has already been submitted.
What AI Policies Do Well
Many institutions have adopted frameworks that classify AI use by level.
A common traffic-light model uses categories such as:
- red: AI prohibited,
- amber: limited AI use permitted,
- green: AI use encouraged.
Other frameworks use four or five levels to distinguish between planning, editing, content generation, and full AI integration.
These systems can help faculty think more carefully about the role AI should play in an assignment. They can also reduce ambiguity for students.
That is valuable.
The problem begins when institutions assume that communicating the category has secured the assessment.
What Is a Discursive Change?
Researchers distinguish between discursive and structural assessment changes.
A discursive change alters instructions, rules, policy language, declarations, or warnings.
The mechanics of the assessment remain the same.
A take-home essay is still a take-home essay. The only difference is that the instructions now say AI may be used for editing but not drafting.
The success of the rule depends entirely on students understanding and following it.
Why Traffic-Light Policies Create an Enforcement Illusion
The traffic-light metaphor feels stronger than it is.
Real traffic lights do not work because drivers have been clearly informed that red means stop. They work because they are embedded in a system of physical design, monitoring, enforcement, and consequences.
An assessment traffic light has the appearance of that structure without the enforcement mechanism.
Researchers describe this as an enforcement illusion: a policy appears to create control even though it only communicates expectations that may be difficult or impossible to verify.
The red label does not prevent a student from using AI.
The amber label does not establish that the student used it only for the permitted tasks.
The green label does not prove that the student retained the underlying knowledge or judgment.
This is a pattern we've seen before: your AI policy is losing before class starts.
More Detailed Policies Do Not Fix the Verification Gap
Institutions often respond by making policies more specific.
They distinguish between planning and drafting, correction and rewriting, assistance and substitution, acceptable and unacceptable prompts.
This may help some students, but it also reveals a deeper problem.
The more precisely an institution specifies permissible AI use, the more obvious the gap becomes between what faculty can describe and what they can verify.
The researchers call this the discursive paradox.
A professor may be able to write:
"AI may be used to improve sentence clarity but not to generate substantive arguments."
But how will the professor know whether a polished paragraph was clarified, substantially rewritten, or created from scratch?
The rule is clear. The evidence is not.
Why Declarations Do Not Solve the Problem
AI-use declarations ask students to state whether and how they used AI.
They can promote reflection and transparency. They are not a reliable verification mechanism.
At King's Business School, as many as 74% of students failed to complete a mandatory AI-use declaration, even though other sections of the same coversheet were routinely completed.
The reasons included:
- fear of academic consequences,
- uncertainty about what counted as AI use,
- inconsistent expectations across courses,
- peer norms,
- and concern that disclosure would be treated as self-incrimination.
A declaration asks students to provide information that may feel risky while offering faculty limited ability to verify the response.
What Is a Structural Assessment Change?
A structural change alters the mechanics of the assessment.
Examples include:
- requiring a live demonstration,
- adding an oral defense,
- observing part of the process,
- introducing authenticated checkpoints,
- connecting several assessment moments,
- or asking students to apply the same concept under a new condition.
The key is that the assessment itself generates evidence of student capability.
Its validity does not rest entirely on the student voluntarily following an invisible boundary.
How to Connect Policy to Structure
A good AI assessment policy should answer four questions.
1. What is permitted?
Be concrete and assignment-specific.
2. What remains the student's responsibility?
Examples include accuracy, source evaluation, model selection, argument quality, and final decisions.
3. What evidence will the student provide?
This could include a process note, draft checkpoint, oral explanation, or live application.
4. How will that evidence affect grading?
Students should know whether the oral component evaluates recall, reasoning, application, judgment, or all four.
This creates two-way transparency. Students disclose their process, while the institution explains how the information will be used and what consequences follow.
A Stronger Example Policy
Consider this language:
"Generative AI may be used for brainstorming, feedback, and presentation support. You remain responsible for all claims, evidence, analysis, and decisions in the final submission. As part of the assessment, you may be asked to explain your process, defend a decision, identify limitations, or apply your reasoning to a new scenario."
This policy does not promise to control every AI interaction.
It tells students what matters and connects the rules to a verification process.
Policy Still Matters
The argument is not that AI policies are useless.
Policies help establish fairness, consistency, shared language, and appropriate expectations.
But policy should support assessment design rather than substitute for it.
The institution's goal is not merely to prove that students obeyed a tool-use rule. It is to determine whether students developed the capability the course claims to teach.
That requires evidence.
For institutions looking to move beyond policy language, AI-resistant assessment provides a useful starting point.