Making assignments harder to outsource to AI: eight design changes that ask for thinking, not just a product

When a teacher says an assignment is "AI-proof," what they usually mean is that they've stopped worrying about it for the moment. No assignment is truly AI-proof, and chasing that goal leads to worse tasks. A more useful aim is to design work where the student's own thinking is visible, where copying a generated answer is less attractive than doing the task, and where you'd notice the difference if they did. That's a design problem, and unlike detection software, which research finds misclassifies non-native English writing, it's one a teacher can solve.

Start by understanding why students outsource work in the first place. Researchers at Stanford's Challenge Success project, who have surveyed students on cheating for years, found that the share of students reporting at least one cheating behaviour in the previous month, 60 to 70 percent, stayed about the same or dipped slightly after ChatGPT arrived. Their interpretation is that cheating is usually a symptom of a deeper problem: students struggling without support, drowning in homework, seeing assignments as meaningless busywork, or fearing they'll disappoint their families. Note that those surveys were in US high schools, so treat them as a pointer, not a measurement of your school. But the lesson travels: if the task feels pointless or overwhelming, a shortcut looks reasonable.

That reframes the goal. Before adding restrictions, ask whether the assignment is worth doing. Design change one: cut the busywork. If a task exists only to fill a homework slot, replace it with one that has a clear purpose the student can see. Fewer, better assignments reduce both pressure and temptation, and they're easier to mark.

Change two: start in class. Give the first ten minutes of a written task in the room, on paper, so you see how the student begins and how they write. Change three: stage the work. Ask for a plan, then a rough draft, then the final version, each with a date. A student who did the thinking can show its history; one who pasted a finished text usually can't. Change four: add a short oral follow-up. Two minutes of "walk me through your second paragraph" tells you more than a stylistic hunch, and it's kinder than an accusation.

Change five: ask for specifics an AI can't know. Instead of "Write about the causes of urbanisation," ask students to describe a change they've noticed in their own neighbourhood and connect it to two ideas from the chapter. Local, personal, and class-specific details, such as a discussion you held on Tuesday, a field visit, or a data set the class collected, make a generic answer obviously off-target. Change six: require the working. In maths and science, marks for method and for explaining why a step is valid; in languages, marks for a short note on why the student chose a particular word or structure.

Change seven: use comparison and critique. Give students an AI-generated answer, or a flawed paragraph you wrote, and ask them to find what's wrong, what's missing, and what they'd change, with reasons. This treats AI as an object of analysis rather than a tool to hide behind, and it builds the checking habit students will need. Change eight: make reflection count. A brief note on "what I found hardest and how I handled it" is difficult to fake convincingly and gives you something to teach from.

None of this requires a screen. Nor does it require you to know which students used which tool. It does require a little more planning per task, which is where AI can honestly help the teacher. It can draft several localised prompt variants for the same learning outcome, suggest a set of staged checkpoints, or produce a flawed sample answer for a critique task. You then edit and decide. Used this way, AI reduces the time cost of better assignment design, instead of increasing the risk of poor student work.

Where does Muallim fit? Muallim is built by DIGIT Pakistan for teachers and school admins; students don't use it, and it doesn't detect AI use, a limitation we discuss in our post on preventing AI misuse. What Quiz & Assessment and the Worksheet Builder can do is generate varied question formats, including short-answer and fill-in-the-blank items mapped to learning outcomes, with the answer key or marks calculated for you, in English, Urdu, or Roman Urdu. Use them for the parts of an assessment that are genuinely routine, and spend your saved time on the local, personal, and oral elements that make work hard to outsource. Everything opens in an editor, so you can add the specific question only you can write.

To begin, choose one assignment you're planning this term and apply three of the eight changes: perhaps a first ten minutes in class, a dated draft, and a two-minute oral follow-up. Keep a note of what you noticed. If the work you receive is more varied and more personal, and your marking is no slower, keep the changes and add another next term. If not, adjust. It's a slow, humane approach, and it's the one that scales, because you're improving the tasks rather than policing the students.