Professors asking how to AI-proof a take-home exam are usually confronting an uncomfortable fact: the conditions that once made take-home work useful now make it easy to outsource.
Students have time, internet access, and powerful generative tools. A carefully written prohibition does not change those conditions.
Begin With the Learning Outcome
Before changing the exam, decide what you are trying to measure.
Is it factual recall, extended analysis, disciplinary judgment, written communication, quantitative reasoning, research ability, or the ability to integrate several concepts?
A take-home format may still be appropriate for complex thinking that requires time, sources, and revision.
The problem is not necessarily the format.
The problem is treating the final product as sufficient evidence of capability. As we've argued before, the finished artifact was only ever a proxy for thinking — and it stopped working as one.
Why Prohibiting AI Is Not Enough
Writing "AI is not permitted" is a discursive change. It changes what students are told, not what they are required to demonstrate.
Without reliable detection, the rule remains difficult to enforce.
This can create an uneven playing field.
Students who follow the rule complete the full cognitive workload. Students who ignore it can receive substantial assistance with little chance of verification. The assessment may therefore reward non-compliance.
Preserve the Take-Home Work, but Add Structure
A strong redesign separates the exam into multiple forms of evidence.
Part 1: The take-home response
Students complete the analysis, essay, case, model, or problem set.
AI may be prohibited, limited, or permitted depending on the course objective.
Part 2: A process explanation
Students identify:
- the approach they selected,
- their most important assumption,
- one alternative they rejected,
- and one part of the response they were least certain about.
This makes their decision-making available for examination.
Part 3: An oral verification
Ask the student a small number of individualized questions.
For example:
- Why did you choose this framework?
- Which evidence had the greatest influence on your conclusion?
- What would change if this assumption were false?
- Which part of your response is most vulnerable to criticism?
- How would you apply the same reasoning to a different case?
The purpose is not to repeat the exam orally.
It is to test whether the student can navigate the reasoning behind it. This is exactly what medicine and law never stopped doing.
Use Transfer Questions
AI can help generate an answer to a known prompt.
Understanding is more visible when the student must transfer the same concept to a variation.
If the take-home exam asks students to analyze one company, the follow-up might introduce a competitor with different economics.
If the exam asks students to solve a programming task, the follow-up might change a requirement or present an error.
If the exam asks for a policy recommendation, the follow-up might introduce a conflicting stakeholder interest.
A student who understands the concept can adapt. A student who memorized or outsourced the response may struggle to do so.
Assess the Process, Not Only the Product
Researchers argue that structural assessment often requires moving from output to process.
A final essay may be AI-generated or heavily AI-assisted. A sequence of checkpoints can reveal how the student's understanding developed.
Possible checkpoints include: thesis approval, preliminary model selection, annotated evidence, draft critique, revision explanation, and final oral defense.
You do not need all of them.
Choose the moments most closely connected to the learning outcome.
Avoid One High-Stakes Verification Event
Replacing a take-home exam with one in-person final does not necessarily solve the educational problem. It may simply replace one high-stakes moment with another.
Research-informed teaching favors scaffolding and repeated opportunities to see student thinking before the final evaluation.
A better design might include:
- a low-stakes practice defense,
- a mid-course checkpoint,
- and a final take-home exam with oral verification.
This creates a record of development instead of relying on one performance. Because sessions can run asynchronously, repeated verification no longer costs you weeks of scheduling.
Decide What Students May Offload
AI use should be connected to the cognitive work you want students to retain.
"Students may use AI to improve formatting and clarity, but must independently select the analytical method and defend all assumptions."
"Students may use AI to generate possible arguments, but must evaluate the alternatives, select the final position, and support it with course evidence."
Cognitive offloading is not inherently harmful. The key is deciding which tasks can be delegated and which capacities remain worth developing.
A Simple Redesigned Take-Home Exam
Here is a practical format:
Written component: 70%
Students complete the original take-home task.
Process note: 10%
Students submit 300 words answering: What approach did you choose? What was your most consequential decision? What uncertainty remained?
Oral verification: 20%
Students answer three questions: one about their reasoning, one variation or transfer question, and one critique or limitation question.
This structure does not require proving whether AI was used.
It requires the student to demonstrate the capability represented by the submission.
What "AI-Proof" Should Really Mean
An AI-proof exam should not mean an exam on which AI is physically or practically impossible to use.
That standard is unrealistic for most take-home work.
A better definition is:
"An assessment whose validity does not collapse merely because students have access to AI."
The final artifact may be AI-assisted. The student must still demonstrate understanding, application, and judgment.
An oral-defense model is one way to add that evidence while preserving the benefits of take-home work.