Designing assignments AI can't finish for you.

The academic integrity conversation has been stuck for a while on detection: can we tell if a student used AI? Mostly, no — not reliably, and not without false accusations that damage trust. The better question is the one assessment designers have always asked: what does this task actually make the student do? If the answer is "produce five paragraphs on a familiar topic", the task was weak before AI arrived. AI just made it obvious.

Five patterns that hold up

• Process over product. Assess the plan, the draft, the revision and the reflection — not only the final piece. A student can generate an essay; they cannot generate having written it.
• Local and specific. Tie the task to something the model can't see: this week's class discussion, a field trip, a dataset the class collected, a local issue.
• Make the AI use visible. Require the prompt log. "Show me how you used the tool" is a fairer and more instructive question than "did you?"
• Talk it through. A three-minute conversation about their own work tells you more than any detector. Build short oral checks into the rhythm of the term.
• Ask for judgement, not summary. "Here is an AI-generated answer. Where is it wrong, and how would you fix it?" is a task the tool cannot do for them, and it teaches exactly the skill we want.

What this doesn't mean

It doesn't mean banning the tools from homework, which is unenforceable and teaches nothing. Nor does it mean every task must be AI-proof — some drills are fine done by hand, in class, with the laptops shut. It means being honest about what each task is for and designing accordingly. This is the substance of the ethics-and-assessment module in our AI professional development, and it is the module educators tell us changed their practice most.