Designing AI-Resistant Assignments That Still Teach

Three percent. In its first year of AI detection, Turnitin reported that of more than 200 million papers it processed, about 11% came back with at least 20% likely AI writing and roughly 3% at 80% or more. Those numbers get quoted as a cheating rate, which is exactly what they are not.

AI-resistant assignments work by making the process visible and the task specific, not by making writing harder to produce. The design goal is an assignment where a student who used a model still has to do the thinking, and where the evidence of that thinking is part of what gets submitted. Detection is a backstop, never the plan.

What actually makes assignments AI-resistant

Start with why that 3% figure is the wrong thing to design against. It counts documents, not students, and it reflects a detector's thresholds rather than anyone's intent. It tells you nothing about which assignments produced those documents. That is the question worth asking, because the distribution is not random. Some tasks are trivially completable by a model in a single prompt, and some are not, and the difference is knowable in advance.

The pattern is straightforward once you look at it. A model is excellent at producing plausible general prose about widely discussed topics. It is poor at anything requiring access to specific material it never saw, at reasoning it has to justify under questioning, and at producing the visible residue of having struggled with something.

Notice that none of the redesigns forbid AI use. They just stop rewarding it. That is the whole trick, and it is the reason this approach survives the next model release while a detection-first strategy does not.

The counter-argument deserves a hearing, because it is the honest objection: this is more work for teachers, in a job that already has too much. That is true, and any article pretending otherwise is written by someone who has never had 140 students. The response is that the work is front-loaded rather than ongoing. Redesigning a prompt is a one-time cost that pays out every term, while investigating detection flags is a recurring cost with an emotional tax attached. Vanderbilt made a version of this calculation in August 2023 when it disabled Turnitin's AI detector, noting that a 1% false-positive rate would wrongly flag roughly 750 of its 75,000 annual papers. Every one of those is a meeting, an appeal, and a damaged relationship. Design is cheaper than adjudication.

Step 1: Make the process the deliverable

If only the finished text is graded, only the finished text will be produced. Split the points.

A workable distribution for a major writing assignment:

• Planning card (10%) submitted before drafting: the question, the intended claim, two sources, and one thing the student expects to find difficult.
• Annotated source list (15%): for each source, two sentences on what it argues and one on why it is in the paper.
• Rough draft (10%), graded for existence and effort rather than quality.
• Revision memo (15%): two specific changes made between draft and final, and why.
• Final text (50%).

Half the grade now sits on artifacts that are tedious to fake convincingly and cheap for you to check. The revision memo is the most informative item on that list, because it requires a student to have an opinion about their own earlier work. A model can generate a revision memo, but it cannot generate one that matches the actual differences between two specific drafts you have both of.

Step 2: Anchor the task in something local, recent or personal

Generality is what makes an assignment easy to automate. Specificity costs nothing and changes more per minute of your time than anything else on this list.

Four anchors that work:

1. Local data. Your county's transit ridership, your school district's budget, a municipal open-data portal. The model has no reliable knowledge of it and will confidently invent numbers if pushed, which is itself a detectable failure.
2. Recent events. Anything after a model's training cutoff. A student who submits confident analysis of a thing that happened three weeks ago and gets the basic facts wrong has told you something.
3. Class-specific material. The argument a classmate made on Tuesday. The example you used in lecture. The reading annotation exercise from week four.
4. Student-generated primary material. An interview they conducted, a survey they ran, an observation log, a photograph they took.

The fourth category is the strongest and the most commonly misapplied. Personal experience alone is not AI-resistant, because a student can feed their own experience to a model and receive a polished reflection. It becomes resistant when the raw material is submitted alongside the analysis: the dated field notes, the interview audio, the survey responses. The artifact anchors the account.

Step 3: Require sources the model cannot have read

An assignment built on widely available texts is an assignment built on training data.

Point the work at material outside that: a course reader you compiled, a paywalled archive your library subscribes to, a physical primary source in a local collection, a dataset you distribute yourself. Then require engagement at a level of detail that only reading produces. Cite the page. Quote the row. Respond to the specific claim on the third page rather than the argument in general.

This has a useful side effect: it catches fabricated citations, which remain one of the most common and most damaging failure modes. A student who leaned on a model for a paper built on your distributed reading list will produce references that look right and do not exist, and that is a far clearer finding than any percentage. It is also the thing experienced instructors notice first, which is why human review outperforms tooling here. A 2025 ACL study found that annotators who use language models frequently identified AI-generated text with 99.3% accuracy, holding up against evasion tactics that defeat automated detectors. Your subject expertise is the instrument. Design assignments that let you use it.

Step 4: Put one component in the room

Not the whole assignment. One component.

Reserve a single element for a live class session: the thesis paragraph, the interpretation of one data table, the counter-argument, or a fifteen-minute written response to a prompt you hand out that day. It takes one class period and it produces two things at once.

The first is a baseline. You now have a sample of each student's unassisted writing at a known level of effort, which is the comparison point that makes every later judgment more reliable and more fair.

The second is an anchor. The in-class piece becomes part of the final submission, so the take-home portion has to be continuous with it. A student whose in-class paragraph is halting and specific and whose take-home essay is fluent and generic has produced a visible discontinuity, and you did not need software to see it.

Keep it short. A full timed exam measures handwriting stamina and processing speed alongside thinking, and it penalizes students with disabilities and multilingual students for reasons that have nothing to do with the learning outcome. Fifteen minutes is an anchor. Ninety minutes is a different assessment with different biases.

Step 5: Add a short oral defense

This is the highest-return change on the list and the one teachers most often skip because it sounds expensive. It is not: two minutes per student, three questions, no preparation.

The questions that work are about choices rather than content:

• "Why did you organize it this way instead of chronologically?"
• "Which source did you almost cut, and why did you keep it?"
• "What did you find hardest here?"

A student who did the work answers instantly and often with visible relief. A student who did not gives an answer about the topic rather than about their paper, which is a distinction you can hear immediately. Grade the coherence between the conversation and the document, not the polish of the speaking, so that quiet and multilingual students are not penalized for delivery.

The real effect is anticipatory. Once students know a defense is coming, the calculation about how to produce the paper changes before they start writing. No detection tool has ever influenced behavior at that stage.

Step 6: Rebuild the rubric so it stops rewarding polish

Here is the uncomfortable part. Most rubrics award a large share of points for exactly what a model produces best: organization, fluency, mechanics, transitions, tone. A rubric that rewards polish is a rubric that rewards the machine.

Move the weight:

The last row is where the actual learning lives, and it is also the single hardest thing for a model to produce convincingly, because it requires a position that survives being questioned.

What this costs, and what it can't fix

Honest accounting, because overselling this approach is how good ideas get abandoned in year two.

It costs a term of redesign work per course, and it front-loads effort into building prompts and rubrics rather than into grading. It adds submission complexity, which means more scaffolding for students who are already struggling with organization. And it does not eliminate AI use. A determined student can still use a model at every stage of a process-visible assignment. What changes is that they have to understand the material to do it, which is a meaningfully better outcome than the one you had before.

Detection still has a narrow role in this system. Not as a gate, and not as a basis for a case, but as one signal among several when something looks off, and as a tool students can use on themselves before submitting. AI Detector 360 reports sentence-level heatmaps and explicit confidence levels across text, PDF and DOCX precisely so that a score reads as evidence rather than a verdict; how we calibrate those levels is public on our methodology page, and the free AI detector works without a sign-up. Our AI essay checker is the version students use before they hand work in, which is the healthiest place for a detector to sit in a course.

Be careful about the fairness dimension while you are at it. The 2023 Stanford study in Patterns found seven detectors flagged an average of 61.3% of TOEFL essays by non-native English speakers as AI-generated. Any policy that leans on scores is a policy that leans hardest on multilingual students, a mechanism we detail in why human writing gets flagged. What instructors actually do with these tools in practice is covered in how professors check for AI, the high school version is in what teachers actually use, and the student-side routine is in checking your essay before you submit.

Design first. Detect rarely. Talk to the student always.