The best answer to AI in the classroom isn't a sharper detector. It's a sharper assignment. If a prompt can be answered in full by pasting it into a chatbot, no amount of policing fixes that. You're trying to catch behavior the task itself invited. Designing AI-resistant assignments turns the problem inside out. Stop asking "how do I stop students from using AI?" Start asking "what can I require that AI alone can't produce?" That second question is more useful, more honest, and a lot less hostile than turning every grading session into an interrogation.
I wrote this for teachers and professors who want something usable by Monday morning, not a manifesto. We'll cover what makes a task hard to outsource, the design moves that hold up, where detection still earns its keep, and how to stay fair to everyone in the room.
Why "AI-Proof" Is the Wrong Goal
No assignment is truly AI-proof. Models summarize, argue, code, outline, and mimic a writing voice. Chase a task no AI can touch and you'll exhaust yourself for nothing. The realistic target is AI-resistant: work where a model can assist but can't stand in for the student, and where a fully machine-written submission ends up obviously thin, generic, or flat wrong.
That shift changes what you build toward. Aim for AI-proof and you drift into surveillance, lockdown browsers, and a default of suspicion. Aim for AI-resistant and you quietly become a better teacher. Tasks that demand specificity, process, and personal judgment happen to be the tasks that produce real learning. You're not digging a trench against your own students. You're making work that's worth their time.
Here's a quick gut check. Paste your prompt into a chatbot yourself. If the raw output would pass, the assignment is the problem. Good news: that's fixable.
Anchor the Task in Things AI Can't Access
The most reliable move is to require inputs the model doesn't have. A general language model can write about photosynthesis or the French Revolution beautifully, because that material saturates its training data. What it can't reach is your specific, local, recent, or personal context.
Build the work around things like these:
- In-class material the model never saw. A particular lecture moment, a thread from last week's discussion, a visiting speaker's argument, a dataset you handed out on paper. "Connect Tuesday's debate about X to chapter four" routes students through content no model can retrieve.
- Personal experience and reflection. Have students tie a concept to their own internship, the town they grew up in, a side project, a hiring decision they once made. Generic AI reflection reads like a horoscope. You'll spot it in the first paragraph.
- Recent or local events. Something from this week, this campus, this city. Knowledge cutoffs and a flat lack of local awareness turn currency and place into natural friction.
- Primary sources you provide. Hand out the exact document, image, or raw spreadsheet and require analysis of that artifact, not a tour of the topic in general.
None of this bans AI. A student can still brainstorm with it or run a grammar pass. But the part that earns the grade has to come from a place the model can't get to.
Make the Process Visible, Not Just the Product
Grade only the final artifact and you're grading the one thing that's easiest to fake. Assess the path that got them there. That's where shortcuts come apart.
Stage the work
Break a big assignment into checkpoints. A topic proposal. Then an annotated source list. Then a rough outline, a first draft you mark up, and finally the polished version. Each piece is small, low-stakes, and feeds the next. A student who outsources everything at the eleventh hour can't go back and manufacture a coherent paper trail. The gaps between stages become a signal all on their own.
Require process artifacts
Ask for the messy evidence of thinking. Handwritten brainstorms. A phone photo of a whiteboard. A research log. The version history from a Google Doc. A short "decisions I made and why" note clipped to the submission. This stuff is a pain to fake convincingly, and it's useful for honest students to produce.
Add an oral or in-class component
A two-minute "walk me through your argument" beats any tool. So does a brief presentation, or an in-class paragraph that extends the take-home work. A student who did the thinking can explain their choices. One who didn't usually stalls. That's fair, transparent evidence, not a dashboard guess.
Design Prompts That Reward Specificity and Judgment
The wording of the prompt carries a surprising amount of the load. AI is strong on the generic and weak on the pointed.
- Demand a stance, then a defense against the strongest objection. "Argue X" is trivial. "Argue X, then steelman the best counterargument and answer it" needs layered reasoning that thin AI output rarely holds together.
- Require synthesis across named sources. "Reconcile the disagreement between Author A in our week-three reading and Author B's framework" forces real integration of specific texts, not a topic summary.
- Ask for application to a novel scenario. Give a fresh case and have students apply the framework to it. Transfer to an unfamiliar situation is exactly where surface-level answers crack.
- Build in iteration and critique. Hand students an AI-generated draft and have them tear it apart: weaknesses, gaps, factual errors. This teaches AI literacy and turns the model into a subject of analysis instead of a ghostwriter.
That last one deserves a spotlight. Bringing AI into the assignment as something to evaluate usually works better than banning it. Students learn to catch invented facts, hollow reasoning, and missing nuance. Those are skills they'll lean on long after your course ends.
Be Transparent About AI Expectations
AI-resistant design only works if students actually know the rules. A big chunk of integrity problems come not from cheats but from honest confusion about where the line sits. "AI helped me brainstorm" and "AI wrote this" feel like different acts to a student, and plenty of them genuinely can't tell which side of your policy each one lands on.
Spell it out on every assignment, not just once in the syllabus:
- What's allowed (brainstorming, grammar checks, outlining help) versus what isn't (generating substantive content you then claim as your own).
- How to disclose AI use. A single line naming the tools and how they were used does the job.
- Why the assignment is built the way it is, so students read the task as being about their learning rather than a hoop to jump through.
Clarity like this reduces misconduct better than any detector, and it earns trust on top. For a deeper framework on AI policy, classroom norms, and moving from detection toward dialogue, our resources for educators pull together approaches you can adapt to your department.
Where AI Detection Still Fits With AI-Resistant Assignments
Stronger assignment design cuts your dependence on detection. It doesn't erase the need to verify. Detection works best as a backstop, never the frontline weapon. When something feels off, a voice that doesn't match a student's earlier work, a sudden jump in polish, sources that smell fabricated, a check points you toward where to look harder.
Two principles keep this fair:
- Treat detection as triage, not a verdict. An AI detector returns a probability, not proof. A high score is a reason to read more carefully and open a conversation, never an automatic accusation. Detectors can flag false positives, and some students get swept up more often than others, non-native English speakers and very formulaic writers among them.
- Pair every signal with human judgment. What you know about the student, their staged drafts, their process artifacts, and a direct conversation all outweigh any single number. Detection narrows where to focus. It doesn't decide anything for you.
Run it that way and detection supports good design instead of replacing it. The assignments carry the weight. The tool just confirms your instinct when something doesn't add up.
A Quick Self-Audit for Any Assignment
Before you send out your next prompt, run it through these:
- Could a chatbot pass on the prompt alone? If so, add a required input it can't access.
- Does the task need specific sources, data, or experiences rather than a general topic?
- Am I grading the process, or only the final product?
- Is there an in-person or oral component where the student shows ownership?
- Have I stated, on the assignment itself, what AI use is and isn't allowed?
- Does the work reward judgment and synthesis over recall and summary?
Answer "yes" to most of these and you've built something genuinely AI-resistant. You've probably also built something more worth doing.
Frequently Asked Questions
Can any assignment be completely AI-proof?
No, and chasing that will only wear you down. The realistic target is AI-resistant: tasks where AI can assist but can't replace the student's thinking. Anchor the work in inputs the model can't access, in-class material, personal experience, recent local events, primary sources you hand out, and require visible process. Against those constraints, AI alone produces something noticeably thin, generic, or wrong.
Should I ban AI entirely in my classroom?
A blanket ban is hard to enforce and skips a teaching opportunity. A cleaner approach: define what's allowed (brainstorming, grammar help) versus what isn't (generating content students claim as their own), require disclosure, and design tasks that reward original thinking. Lots of instructors get better results bringing AI into the assignment, having students critique AI-generated drafts, than trying to keep it out.
How does AI detection fit with AI-resistant assignment design?
Detection is a backstop, not a frontline tool. Strong design cuts how often you need it, but when a submission feels off, an AI detector helps you decide where to look more closely. Treat the score as probabilistic triage, paired with your own judgment, the student's drafts, and a conversation. Never as standalone proof of misconduct.
What's the single most effective change I can make?
Stage the work and grade the process, not just the final product. Requiring a proposal, source list, outline, and draft, each one building on the last, makes last-minute outsourcing obvious and rewards genuine effort. It's the change that most reliably turns a prompt a chatbot could answer into one that needs a real student behind it.
Want a ready-to-use framework for evaluating and redesigning your prompts? Download our assignment-design rubric and start building assignments that reward real thinking, then keep an AI detector on hand as your backstop when a submission doesn't add up.
Try it on your own writing