AI Cheating Is a Design Problem. Draw It, Then Defend It.

AI can draft the essay. It cannot stand next to a hand drawn page and answer a classmate’s question. A 2026 study built an assignment around that gap.

The study was published on September 21 in the journal Discover Artificial Intelligence by Khalid Khamis Al Saadi of Qatar University and two coauthors, and it reached a wider audience this week when the science news site Bioengineer.org covered it with a headline promising that doodles and a live defence “may outsmart AI cheats in university assessment.” The headline is about AI cheating. The study is about something more useful to anyone who grades: where the evidence of learning has to live once a fluent paragraph costs nothing.

Most of the AI cheating conversation is still about detection. Run the essay through a tool, read the score, decide what to do. The authors start somewhere else. They note that research on AI detection tools has documented their instability, their susceptibility to paraphrasing and a potential bias against multilingual authors, and they argue that integrity should be built into the design of the assessment itself. They call it integrity by design. You stop asking whether the essay is real and start asking for evidence an essay cannot supply.

What the study actually did

The setting was three cohorts of an undergraduate Educational Technology course at a public university in Oman: 76 students working in 23 groups. Each cohort went through the same four phase sequence.

First, bounded AI ideation. Students could use a generative AI tool such as ChatGPT to brainstorm, explore the topic and generate examples, with full disclosure. The boundary was stated out loud: AI could help open up the topic, and it could not substitute for the stages that came next.

Second, the hand drawn page. Each group had to turn those early ideas into one collaborative, hand drawn conceptual doodle. The authors call this an analogue bottleneck, a checkpoint where generated prose is useless and students have to select, condense and organize ideas themselves. The drawing was not judged for artistic quality. It was judged for what the group chose to include, what they left out and how they connected the pieces.

Third, the gallery walk defence. Groups displayed their pages, gave a short explanation, answered questions from their peers and were evaluated against explicit rubric criteria. This is the step that turns a drawing into evidence, because the group has to explain why the page is built the way it is.

Fourth, an individual written reflection on what each student learned and what tensions the process raised.

What AI cheating did to the polished paper

Here is the uncomfortable part for anyone with a stack of essays. A clean, well organized submission used to be decent evidence that a student had done the thinking. Now it mostly proves that something, somewhere, can write. The authors make this point in their own framework: generative AI can now simulate the outputs instructors have long treated as proof of constructive work, like essays, summaries and even digital diagrams, so an instructor can no longer tell from the artifact alone whether it reflects the student’s own effort.

Notice what the design does not do. It does not ban AI. Students used it, disclosed it, and by the third cohort, in spring 2025, the authors describe students treating AI brainstorming as an ordinary starting step. What changes is where the grade comes from. The evidence moves away from the polished output and toward three things AI cannot do on a student’s behalf in the room: choosing what goes on one page, drawing the connections by hand and answering a question about those connections in real time.

That framing matters more than the doodle. The authors say plainly that visibility does not depend on whether an artifact is handwritten or digital. It lives in a student’s ability to respond to questions, clarify a misunderstanding and adjust an explanation on the spot.

What the pages and the defenses showed

Two raters, the course instructor and a second researcher who did not teach the course, scored all 23 group pages on two dimensions, each from 1 to 4. Conceptual Density counted the distinct course concepts a page made visible. Relational Explicitness rated how clearly the page connected those concepts through arrows, spatial layout and labels. Their initial agreement was 87 percent. The authors are careful here: that is simple percentage agreement, not a chance corrected statistic, and they present it as a sign of coding consistency rather than proof of reliability.

The pattern was still clear. Of the 23 pages, 19, or 83 percent, paired high conceptual density with explicit connectors. Groups that put a lot on the page also showed how it fit together. The space limit of a single page forced choices, and the choices showed the thinking.

The thematic analysis found four themes. The constraint of drawing by hand as a way of making sense appeared strongly in 20 of 23 groups. AI as a scaffolder rather than a scribe was evident in 21. Friction as part of the learning, things like drawing anxiety and arguments over workload, surfaced in 18. And public explanation as accountability was documented in all 23 groups, the only theme that showed up everywhere.

That last one is the finding worth carrying into your own course. One student wrote, “When someone asked why we linked policy to practice with one arrow, I realised I could not defend it.” No detector produces that moment. A peer question does, and the student finds the gap before you have to point it out.

What this study cannot tell you

Read it as a strong design idea with honest limits, not as settled proof. It is a qualitative study at one institution, in one course, with 76 students. The course instructor was also a researcher and one of the coders, and the authors name that dual role as a limitation even though reflections were anonymized and coding began only after final grades were in. They did not measure long term retention or transfer to later courses. They did not compare hand drawn pages against digital drawing tools under identical defense conditions.

They also flag one observation as tentative. Groups that leaned toward digital design tools sometimes gave weaker verbal justifications during the gallery walk, while groups that wrestled with paper explained their choices more clearly. The authors call this an observation from the defense sessions, not a measured comparison. Treat it the same way.

None of that weakens the practical point. You do not need a controlled trial to see why a student who has to answer a question out loud about their own page is giving you better evidence than a student who uploads a file.

Draw and defend, in one class session

You can run a version of this sequence on an assignment you already teach, without rewriting your course. It will not end AI cheating by itself. It moves the grade toward evidence a student has to produce in the room. Here is the shape of it.

1. Write the AI boundary before anything else. One sentence students read first: where AI is welcome in this assignment, and where it stops. In the study, it was welcome for brainstorming with disclosure and not for the synthesis, the drawing or the defense.

2. Ask for one page, by hand. Name how many course concepts the page must show and require that every connection be drawn and labeled. Say out loud that artistic skill is not graded, because several students in the study felt insecure about their drawing.

3. Hold the defense. Each group or student gets a short window to explain the page, then answers questions from you or from classmates. Write three questions in advance that can only be answered by someone who made the choices on the page, such as why two ideas are connected and what was left off.

4. Grade the page on two criteria and keep a reflection. Use the study’s two dimensions, how many concepts are visible and how clearly they connect, each on a four point scale. Then collect a short individual reflection so every student, not just the group, leaves a record of what they learned.

We made a related argument when Dartmouth authorized an AI detector and its own rules said the score is not the case, and again last week in why your judgment has to come before any AI grade. The thread is the same. The more fluent the tools get, the more the evidence has to come from what a student can do in front of you.

Your next step

Take one assignment you are worried about this term through The Draw and Defend Assignment, by hand, in about 25 minutes.

When you finish you will hold a finished assignment sheet: the AI boundary sentence, the one page instructions, three defense questions only the maker of the page can answer, a two criterion rubric with what a 4 looks like on each, and a reflection prompt. Then the next thought arrives on its own. That is one assignment. Your course has more, and every one of them is now a question about where the evidence lives.

Redesigning one by hand is the right way to learn the method. Redesigning every assignment in a course that way is the slow part. The Assessment Builder builds assignments, rubrics and tagged question banks, grounded in your Course Record, the five minute record of your course, your outcomes, your policies and your voice that every tool in the toolkit uses. Bring the sheet you build today as the pattern.

If you would rather think it through with other instructors first, the Harvest Kernel Learning Community is free, and this question is the conversation there this week.

Sources

• Khalid Khamis Al Saadi, Raja Maznah Hussain and Gwo-Jen Hwang, “AI supported doodling and gallery walk defence provide visible evidence of student reasoning in higher education assessment,” Discover Artificial Intelligence, volume 6, article 1195, published September 21, 2026. doi.org
• Bioengineer.org, news coverage of the study, October 19, 2026. bioengineer.org