The short answer
- You can’t detect your way out of this — detectors are unreliable. Design is the durable answer, not surveillance.
- The core principle: ask for work that references things AI can’t access — your specific class, the student’s own process, live discussion, personal reasoning.
- Cheapest high-impact move: require process artifacts (outline, draft, reflection), not just the final product.
- “AI-proof” is really “AI-resistant.” Nothing is immune; the goal is assignments where using AI to skip the thinking is harder than doing the thinking.
- Some of these cost real prep time — the honest ranking below tells you which are free and which aren’t.
The instinct after ChatGPT arrived was to find the tool that catches it. Three years of evidence say that tool doesn’t exist — detectors are wrong too often, and the students they wrongly flag are the ones least able to defend themselves. The approach that actually holds up is older and less satisfying: change what you ask for. Here are nine moves, ranked by what they cost you.
The principle: ask for what AI can’t reach
A general AI has no access to your classroom, your student’s process, or this week’s discussion — so design around those. Every durable move below is a variation on one idea: require something specific, personal, or live that a model trained on the open internet simply doesn’t have. The more an assignment leans on those, the harder it is to outsource the thinking.
The nine moves, cheapest first
Free or near-free (do these first):
- Reference the specific. Require students to connect the task to a named class discussion, a lab you actually ran, or a text you read together. “Using Tuesday’s debate on X…” is invisible to a chatbot.
- Ask for the process, not just the product. Collect the outline and a rough draft alongside the final. The thinking lives in the messy middle, which is tedious to fake.
- Add a short reflection. “What was hardest here, and how did you get unstuck?” Two sentences that AI can produce generically but a student who did the work answers specifically.
- Make it current or local. Ask about something recent or specific to your town — the fresher and more particular, the thinner the model’s coverage.
Moderate prep:
- Stage the assignment. Grade the topic, then the outline, then the draft, then the final — as separate checkpoints. Outsourcing four linked stages is more work than doing one paper.
- Design for a real audience or format the model handles poorly — a hand-annotated diagram, a recorded explanation, a physical artifact.
- Build in an in-class component. A short handwritten reflection or a paragraph written in the room establishes each student’s voice, which makes off-voice submissions visible.
Higher cost, highest integrity:
- Oral defense. Two minutes per student explaining their reasoning. Nothing surfaces understanding — or its absence — faster, and nothing is harder to fake. Expensive in time; unbeatable in signal.
- Rework it live. Give a short in-class extension of a take-home task (“now apply it to this new case”). A student who understood theirs can; one who didn’t, can’t.
What “AI-proof” honestly means
It means AI-resistant, not immune — and that’s the right goal. A determined student can still get AI to fake a reflection or stage a draft. The point isn’t a locked door; it’s raising the effort of cheating above the effort of learning. When outsourcing the thinking is more work than just doing it, most students do it — which is exactly the outcome you want, achieved without accusing anyone.
It also reframes the relationship. Detection makes you an adversary scanning for guilt; design makes you a teacher asking for real work. Students feel the difference, and so do you.
Pair it with a clear policy
Assignment design works best alongside expectations students actually understand. Set those out plainly — what AI use is allowed, what isn’t, and what disclosure you expect — with a classroom AI policy, and grade with rubrics built for the AI era so you’re rewarding thinking rather than fluent prose a prompt can produce.
Next: detecting AI work honestly (when you still need to), and using AI for feedback to buy back some of the time these moves cost.
Frequently asked questions
How do you make an assignment AI can't do?
Design it around things a general AI can’t access: your specific class discussions, the student’s own drafting process, personal reflection, live in-class work, and current or local specifics. No assignment is fully immune, but ones that require process artifacts and personal connection make outsourcing the thinking harder than doing it. Detection isn’t the answer — design is.
Is it worth using AI detectors instead?
Not as your main strategy. AI detectors are unreliable, produce false positives that harm honest students — especially non-native English speakers — and are easily defeated by editing. They can be a weak signal that prompts a conversation, but building your integrity approach on detection means building it on a tool that’s wrong too often to trust.
Do AI-proof assignments create more grading work?
Some do, and it’s worth being honest about that. Grading process artifacts or oral check-ins takes more time than grading a final essay. The moves in this guide are ranked by prep and grading cost so you can pick what fits your workload — several add almost no time, and a few trade time for genuinely better integrity.
Should I just ban AI completely?
A blanket ban is hard to enforce and ignores that students will use AI in college and work. A more durable approach is to design assignments where AI can help with understanding but can’t do the graded thinking, and to be explicit about when it’s allowed and when it isn’t. Clear expectations plus resistant design beats prohibition.