Designing Courses That Don't Fall Apart When Students Have AI

Most people talk about this like it's a problem to solve. It's not. It's just a different set of constraints on assessment design. I spent the last two years redesigning three upper-level courses because of what happened when my students tried asking AI to do their work. It was messier than I expected and the fixes are mostly boring. The core mistake everyone makes is trying to make assignments AI-proof by making them harder. That doesn't work. Students will spend more time on prompt engineering than they save. The real shift is toward process-based assessment where the artifacts students submit are the ones that prove they did the work, not the final output.

Practical Steps Toward Robot Proof Higher Education In The Age Of Artificial Intelligence

Start by mapping every assignment in your course to a specific cognitive skill. Bloom's taxonomy is the standard reference but you don't need to cite it. Just list whether each task asks students to remember, apply, analyze, evaluate, or create. AI handles the top two easily. It can synthesize and summarize at a decent level now. The skills that actually survive are analysis under specific constraints and original synthesis tied to lived experience in the class. I redesigned a mid-level statistics course around this. The old midterm had students run regression analyses on provided datasets and interpret the output. Every single student in my spring 2025 section ran the analysis through an AI tool and submitted clean writeups. The grades looked fine. The learning was not there. I rebuilt the exam around field data collection. Students had to go out, collect a small dataset themselves, define their variables, justify their sampling strategy in writing before they touched any numbers, and then do the analysis. The AI could help with the mechanics but it couldn't replace the decisions that happened in the first two weeks of the assignment. The average quality of statistical reasoning went up and the plagiarism rate dropped to near zero. Here is what actually works in practice:

Oral assessments still matter more than most people want to admit. A five-minute conversation about what a student wrote takes less time than you think if you batch it efficiently. I grade three students per office hour slot and cover a whole section in about 30 minutes total. The student who actually did the work can explain their reasoning. The student who used AI will either stall or repeat generic phrasing. Version your assignments annually. Reusing the exact same prompt for three years in a row guarantees that the top of the class will have optimized prompts by the second semester. Change the dataset, shift the case study location, alter the constraints. A 15-minute tweak to an assignment prompt now saves you from spending three hours auditing submissions later. Require process documentation. Have students submit a brief log of their workflow. Not a detailed diary. A half-page note about what they tried first, what didn't work, and what they changed. This takes 10 minutes to grade and it separates people who engaged with the material from people who generated a clean final product without the journey.

Use in-class writing for low-stakes work. Anything worth less than 15 percent of the grade can be done in person under supervised conditions. Reading responses, short reflections, problem sets. This removes the incentive to outsource the easy points and frees you to design heavier assignments around skills AI struggles with.

There is a specific edge case that burned me and I should document it because nobody talks about this. I had a strong student who was genuinely struggling with English as a second language. She submitted excellent written work that was clearly AI-assisted. When I pulled her in for an oral check, she understood the material perfectly but struggled to verbalize it in real-time English. The oral assessment, which I use as my primary fraud detector, almost failed an honest student. The workaround was straightforward. I told her explicitly that using AI for writing support was acceptable and we would assess her understanding orally rather than her prose. She passed comfortably. The lesson was that the system needs to be explicit about what it values or it will accidentally punish the people it should be supporting most. The counter-intuitive part nobody wants to hear: making assignments robot-proof often makes them worse for learning if you overdo it. When I started requiring process logs and oral checks across the board, student satisfaction dropped in the mid-terms. They felt micromanaged. The fix was transparency. I showed them the data from my previous cohort where AI use correlated with lower performance on the final comprehensive exam. Most students agreed with the policy once they saw the evidence. A few didn't and that was fine. They were taking the class for a grade they could get elsewhere. Another nuance that matters: AI detection tools are not reliable enough to base grading decisions on. Turnitin's AI category and GPTZero both have false positive rates that are high enough to be dangerous. I stopped relying on them entirely and switched to the process-based methods I described above. The detection tools detect patterns in text. They don't detect learning. Those are different things. The real limitation of this approach is time. Oral assessments and process documentation multiply your grading workload by roughly 40 percent compared to traditional essay grading. If you teach a course with 120 students, you need to scale this differently. I use a hybrid model for large sections. Students submit a brief video explanation instead of an oral meeting. They record a three-minute screen share walking through their process. I sample 30 percent of those videos randomly and the rest get checked on suspicion triggers like unusually polished prose or mismatch between the process log and the final output. This brings the grading overhead down to something manageable while keeping the accountability intact. The other hard truth: some courses will resist this more than others. Labor-intensive hands-on disciplines like lab sciences and studio arts already have built-in process assessment. The fields that need to adapt the most are the theory-heavy ones with long-form writing assessments. My advice for those instructors is to start small. Pick one assignment per term and redesign it using the process documentation model. Once you see how it goes, expand. Don't overhaul an entire syllabus in one semester. You will burn out and the students will resent the whiplash. I should also note that the students who benefit most from these changes are the ones who actually need the education. The students who were gaming the system with AI before were not learning anything anyway. The ones who put in the work were already succeeding. Making the process visible just makes the success honest. There is no downside to that except the grading time, and that is a resource problem, not a pedagogical one. If you are starting from scratch with a new course, write the assessments first and then build the lectures around what students need to know to complete them well. That ordering makes a noticeable difference. Most instructors do it backward and then wonder why the exams feel disconnected from the readings.