Difficulty is not the defence
The first instinct is to make the task harder. More words, more sources, more formal analysis. That is the model's strongest ground. A demanding, abstract, well-structured essay is close to the easiest thing you can ask it for.
What defeats it is not difficulty. It is specificity to something it has no access to: your room, your week, your students, this town.
Four levers
1. Local and recent.
Before: "Discuss the causes of rural to urban migration."
After: "Interview one person in this neighbourhood who moved here from somewhere else. What did they tell you, and which of the three causes we studied does it fit — or does it not fit any of them?"
The model can write the essay. It cannot have the conversation, and the interesting answers come from the person whose reason did not fit the textbook.
2. Make the process the product.
Submit the first draft, the feedback you received, and a note on what you changed and why. Weight the revision, not the final text.
A student can generate a final text in a minute. Generating a plausible revision history that matches the feedback you actually wrote on their draft in class is more work than doing the assignment.
3. Anchor it to the room.
"Use the data our class collected on Tuesday." "Respond to the argument Priya made about the second experiment." "Compare your result with the one on the board."
The model was not there. Nothing it produces can reference what happened.
4. Ask them to defend it out loud.
Two minutes per student, spread across a week while others work. This changes behaviour before anything is submitted, because they know it is coming. It is also the fastest assessment you will ever do — you learn more in two minutes of questions than from four pages.
Two honest limits
This costs time you may not have. With 60 students you cannot redesign everything. So do not. Pick the two or three pieces in the term that carry real weight and redesign those. Let low-stakes practice stay low-stakes, and stop policing it.
"AI-proof" is not a stable property. A task that defeats the model this year may not next year, and anyone who sells you a permanent solution is selling something. What is stable is a task about something only this student, in this place, this week, could produce. Build on that and you are not chasing a moving target.
The other direction
Some tasks should assume the tool and assess what is left.
Ask an AI to answer this question. Print or copy its answer. Find three things wrong with it — factual errors, missing steps, or claims it cannot support. Correct them, and explain how you know.
This is harder than the original essay, not easier. You cannot find the errors without the knowledge you were going to be assessed on, and it teaches the habit you most want them to leave school with. It also works well for students without much access, since one printed transcript serves a whole class.
Before you move on