Designing AI-Resistant Assignments

An AI-resistant assignment is not one that AI cannot touch; almost any take-home text task can be attempted by a chatbot. The goal is to make outsourcing the work pointless or visibly incomplete, so the easiest path to a good grade runs through actual learning. Four design moves do most of the work. First, personal and local connection. Prompts that require students to draw on a specific class discussion, their own interview, data they collected, or a local place give generic model output nothing to grab. A chatbot can invent an interview, but it cannot produce the recording or the notes a teacher asks to see. Second, in-class stages. Splitting a project into a proposal, an annotated source list, a paragraph drafted in class and a revision with a reflection lets the teacher watch the thinking develop. A final draft that looks nothing like the earlier stages becomes a conversation starter rather than an accusation. Third, oral defense. A short conversation where the student explains a choice, answers a follow-up or extends an argument is hard to fake. Oral examination is a long-standing practice, from doctoral vivas to school systems that use spoken exams. Fourth, authentic tasks. Writing for a real audience, such as a letter to a city council, a guide for younger students or a critique of a real product, demands specific judgment that generic output lacks. Common misconceptions: that obscure topics defeat AI (models handle many niche subjects and fabricate the rest convincingly), that hidden trap text in prompts is reliable (students often notice, and it erodes trust), and that detectors can replace design. OpenAI withdrew its own AI text classifier in 2023, citing low accuracy. Some teachers go further and allow AI openly for part of a task, then assess how well students critique or improve its output.