Assignments that make AI use visible.

About: AI-resistant prompts

The phrase "AI-proof assignment" has stopped being useful. Any prompt a student can type into a chatbot can produce a draft; the question is how visible the shortcut becomes once it's produced. The patterns below are not AI-proof. They are AI-resistant in two specific ways: they make the shortcut harder to take, and they make the shortcut easier to detect when taken.

Eight prompt patterns

1. Ground in a specific, dated artifact. Instead of "Analyze the causes of the French Revolution," assign "Using the three primary sources handed out in class on March 12, analyze…" Students who outsource drafting must either fabricate the sources or miss the prompt.
2. Require a dated draft history. Submissions include a link to a Google Doc or Word document with version history enabled. The history is assessed as part of the grade, not as a detection tool.
3. Add an in-class component. Split the assignment into an out-of-class draft and an in-class revision or defense. The in-class portion anchors the voice baseline.
4. Demand a local or personal referent. "Connect the argument to a specific decision made in our town council in the last six months" or "…to a decision you personally made last year." Generic AI draft text struggles to stay specific and consistent.
5. Layer a reading quiz. Include a short untimed quiz on the reading, at the sentence level. Students who read will pass; students who skimmed a summary will not.
6. Require a process memo. The submission includes a 200-word memo describing how the draft came together: when you wrote which section, what you cut, why. AI-drafted essays produce generic memos.
7. Use unusual formats. A letter to a skeptical friend, a courtroom opening statement, a museum wall-text. AI defaults to five-paragraph essay structure. Unusual formats either pull the AI out of its default or produce unmistakable stylistic drift.
8. Grade revision, not draft. Submit Draft 1, get feedback, submit Draft 2, and grade the delta. Revision is hard for AI to do invisibly.

How this pairs with detection

These patterns do not replace detection. They make detection cheaper to act on. When a prompt is specific and a draft history exists, a detector flag becomes one line in a three-part evidence chain (detector + draft history + prompt-specificity), and the conversation with the student is short. When a prompt is generic and no draft history exists, a detector flag alone carries too much weight.

What doesn't work

"Turn off spell-check" does not work. "Write in blue pen" does not work for remote classes. Anti-plagiarism scare language on the syllabus does not measurably reduce AI use in any study we've reviewed. What works is making the shortcut visible and expensive compared to doing the work.