# Take-Home Exams: Designing Assessments Bots Cannot Game

Source: https://examfence.com/blog/2026-11-08-take-home-exams-assessments-bots-cannot-game.html

[Home](../index.md) / [Blog](index.md)


**Short answer:** Design the exam around judgment, not recall. Give students a messy case, a tight window of about 24 hours, and two to three hours of work, and ask them to apply a framework and defend the choice. A bot can recite a definition. It cannot argue a close call in a specific case without giving itself away.

## Recall is the part the bot does best

Every take-home exam has a quiet assumption built into it: that the student will do the work alone, with their notes, in the time given. The bot breaks that assumption in the most boring way possible. It does not cheat cleverly. It answers the recall questions faster and more accurately than the student, because recall is what it is for.

That means the fix is not a better lock on the exam. It is a different exam. A question that asks for the definition of a term has one right answer, and the bot has it. A question that asks a student to apply that term to a case, choose between two defensible readings, and say why, has no answer to copy. The work is the thinking, and the thinking is visible in the answer.

## The window matters more than the lockdown

A 24 hour window with two to three hours of expected work is the shape that works. It is long enough that a student with a job or a bad connection is not punished for it, and short enough that the exam cannot be passed around, solved by committee, and returned. The window is the constraint, and the constraint is the assessment.

Long windows invite the other failure: the exam becomes a group project. A week to answer ten questions is a week for the answers to circulate, and the circulation is invisible to the marker. Shorten the window, and the circulation has to happen in hours, which is exactly when it gets sloppy and visible. Dawson's 2020 work on time limits made the same point from the other side: time pressure changes what students do, and the exam should be designed for the pressure it actually applies.

## Honor pledges work, and the research says why

The cheapest integrity control is also the most studied: the honor pledge. Research out of Texas A and M found that students who believed in the pledge were about a third as likely to cheat as those who did not, a 0.36 times likelihood. The pledge does not work because students fear it. It works because it asks them to decide, before the exam, what kind of student they are.

That is why the pledge belongs at the start of the exam, not buried in the course handbook. A student who signs before they see the questions has made a commitment. A student who never sees it has made nothing. The pledge is not a legal document. It is a moment of decision, and moments of decision change behavior in a way that rules do not.

## Mark the reasoning, not the conclusion

The last design choice is in the marking. If the mark goes to the conclusion, the bot wins, because the conclusion is the part it can generate. If the mark goes to the reasoning, the chain from evidence to judgment, the student wins, because the reasoning is the part that has to be theirs. A right answer with no reasoning earns little. A wrong answer with honest reasoning earns more, and the students figure that out fast.

Bengtsson's 2019 systematic review of cheating research lands in the same place: the assessments that resist cheating are the ones where the work is the answer, not the route to it. Design the exam so the reasoning is the product, and the bot has nothing to sell.

Keep reading: [IP Bans Don't Stop Exam Bots: Why Rotation Wins](https://examfence.com/blog/2026-11-06-ip-rotation-exam-bots.md), [How Exam Cheating Rings Automate Answers in Real Time](https://examfence.com/blog/2026-11-06-exam-cheating-rings-automation.md), [How should testing programs run a fair appeals process for flagged exam results?](https://examfence.com/blog/2026-10-19-appeals-process-flagged-exam-results.md).

## Frequently asked questions

### Can you stop students using AI on a take-home exam?

You cannot stop it reliably, and trying to detect it is a losing game. The better move is to design an exam where the AI does not help much: a messy case, a judgment call between two defensible readings, and a mark that goes to the reasoning rather than the conclusion. A bot can recite a definition, but it cannot argue a close call without giving itself away.

### How long should a take-home exam window be?

About 24 hours, with two to three hours of expected work inside it. That is long enough to be fair to students with jobs or bad connections, and short enough that the exam cannot be passed around and solved by committee. A week-long window invites the answers to circulate, and the circulation is invisible to the marker.

### Do honor pledges actually reduce cheating?

Yes, when students take them seriously. Research from Texas A and M found that students who believed in the pledge were about 0.36 times as likely to cheat as those who did not. The pledge works because it asks the student to decide before the exam what kind of student they are, and that decision carries into the work.

### What should I mark if the conclusion is wrong?

Mark the reasoning. A wrong answer with an honest chain from evidence to judgment shows the student did the work, and it earns more than a right answer with no reasoning at all. When the mark goes to the reasoning, the exam rewards the part the student actually has to do themselves.

[Get a free bot-traffic audit](mailto:hello@examfence.com?subject=Free%20free%20bot-traffic%20audit)