A high school physics teacher put it bluntly on X last week: “There is no universe in the multiverse where I’d assign a take-home assignment for a serious grade.” That’s one honest response to ChatGPT — move everything in-class, proctor everything, trust nothing. It works, and it costs you every assignment that needs more than 50 minutes of thinking.
There’s a middle path that doesn’t get written about much, because the people selling AI detectors can’t sell it and the pedagogy blogs keep it abstract: redesign the assignment so the AI shortcut stops producing a passing grade. Not AI-proof in the absolute sense — nothing is — but AI-resistant in the way that matters: a student who outsources it to ChatGPT still has to do the actual learning to survive the assignment’s structure.
The move takes one prompt and about five minutes per assignment. Here’s the prompt, why it works, and what it did to five real assignment types.
Why redesign beats detection
Quick version, because we’ve documented the long version: AI detectors flag human writing constantly — Stanford researchers measured a 61% false-positive rate on essays by non-native English speakers — and universities from Berkeley to Washington State have stopped accepting detector scores as evidence. Washington State’s own review found a third of detector-based integrity cases ended “not responsible.” The detector route produces false accusations, appeals, and parent meetings, and it still misses the students who paraphrase the output.
Redesign inverts the problem. Instead of trying to prove what a student did last night, the assignment structurally requires things an AI session can’t fake: their in-class presence, their process over time, their specific life, their voice under a follow-up question. The research on this converges on three mechanics:
None of these punish the honest student — they’re just good assessment, which is why Ohio’s model AI policy toolkit recommends the same pattern. They make the dishonest path slower than doing the work, which is the only deterrent that has ever consistently worked.
The prompt
Open ChatGPT, Claude, or Gemini and paste this with your assignment inside:
Here is an assignment I give my [8th grade history] students:
[paste the full assignment instructions]
Redesign it so a student cannot complete it by pasting it into an
AI chatbot, while keeping the same learning objective. Apply three
changes:
1. Split it into stages I can see (proposal, draft, revision,
reflection) — tell me what to collect at each stage
2. Add one short in-class or oral component tied directly to the
student's own submission (under 2 minutes per student)
3. Require at least one source an AI cannot invent: personal
experience, a local interview, something specific from our
class discussions, or data we collected together
Then show me: (a) the rewritten assignment instructions at the
same reading level as mine, (b) what the AI-shortcut version
would look like now and why it fails, (c) what I grade at each
stage so my total grading time stays roughly the same.
The last line matters most. Redesign advice usually dies on grading load — five stages sounds like five times the work. Forcing the AI to rebalance (“grade the proposal pass/fail, spot-check drafts, grade only the final + reflection deeply”) keeps the workload honest, and it’s genuinely good at this part.
Five assignments, before and after
Here’s the pattern applied to the five most-outsourced assignment types. The “after” column is condensed from real runs of the prompt above:
| Assignment | Before (one prompt away) | After (AI-resistant) |
|---|---|---|
| Persuasive essay | “Write 5 paragraphs on whether school should start later” | Position declared in an in-class paragraph Monday; outline conferenced Wednesday; final essay must cite our class survey data + rebut the argument a classmate made in discussion (named) |
| Lab write-up | “Write up the pendulum lab: hypothesis, method, results, conclusion” | Data table signed off in class; write-up explains your group’s anomaly (every group’s data differs); 90-second defense: “why did your period change when…?” |
| Book report | “Summarize the book and give your opinion” | Reading journal collected at 3 checkpoints; final piece connects one scene to something in your own life the class knows about you; oral: pick the passage you’d cut and defend it |
| Reflection paper | “Reflect on what you learned this unit” | Anchored to three specific moments from class (the AI wasn’t there); references your own checkpoint answers; ends with a question you still have — asked out loud |
| DBQ / document analysis | “Analyze these 4 documents on the New Deal” | One document swapped for a local primary source we found together; annotations photographed in class; thesis defended in a 90-second stand-up before writing begins |
Notice what changed in every row: the final product still exists, but it’s now the last visible step of a trail — and somewhere in that trail is a moment where the student stands next to their own work and talks about it. A student who had ChatGPT write the essay fails the 90 seconds. A student who used AI to polish their own thinking passes easily — which, if your classroom rule is disclosure-based, is exactly the distinction you wanted to grade.
Running it without blowing up your week
Start with one assignment, not your whole gradebook. Pick the one that came back most obviously AI-written last year and run the prompt on it today. The rewrite takes five minutes; deciding you believe in it takes one graded cycle.
Reuse the oral-defense questions. The prompt generates them, and three good ones (“what would you cut?”, “where did this idea come from?”, “what surprised you?”) work for any submission all year. Ninety seconds per student, once per major assignment, is the entire time cost — and it doubles as the comprehension check the AHRQ-style “explain it back” research says beats any scanner.
Keep the stages lightweight. Proposal = two sentences, pass/fail. Draft = collected, spot-checked, not graded. The trail’s value is that it exists — most disputes end the moment you ask to see the draft history.
Say the quiet part to students. “This assignment is built so AI can’t do it for you — here’s why that’s a compliment” reframes the design from suspicion to respect, and the integrity research consistently shows explained rules outperform enforced ones.
What this can’t fix
- A determined cheater with a tutor, a sibling, or a very patient AI session can still fake a process trail. You’re raising the cost of dishonesty above the cost of learning, not eliminating dishonesty from the species.
- In-class components need class time. Ninety seconds × 30 students is a real cost; that’s why it attaches to major assignments, not weekly homework.
- It doesn’t retroactively resolve last semester’s suspicions. This is a design fix, not an investigation tool — for the accusation you’re sitting on right now, the false-positives guide covers that conversation.
- Some assignments shouldn’t be saved. If the redesign reveals the original was really “produce five paragraphs that look like thinking,” the honest fix is a better assignment, not a fortified version of the old one. The prompt tends to expose these — consider it a feature.
The bottom line
You can’t detect your way out of AI-written homework, but you can design your way out: process over product, a live 90 seconds, one source no chatbot can invent. One prompt rewrites the assignment; one graded cycle proves it works. If you want this redesign move taught step by step — plus the classroom rule, the parent email, and the 90-second in-class check that completes the system — our new course AI for Teachers: Your First Week Back starts free and walks the prompt through five worked assignment types.
Sources
- GPT detectors are biased against non-native English writers — Stanford HAI
- WSU discontinues Turnitin AI detection (Feb 2026 memo) — WSU Provost
- AI Model Policy for Ohio Districts and Schools — Ohio Department of Education and Workforce
- Model syllabus statements for generative AI — Suffolk County Community College CTL
- How states are regulating AI in education — MultiState
- AI use in schools growing, but district policies haven’t caught up — GovTech
