See my blog post here: https://blog.stephenturner.us/p/ai-resistant-assignments. This is GPT-5.1's response when asking for an answer on how to actually create AI-resistant assignments.
Short answer: for take-home work you mostly cannot make it “AI proof.” What you can do is 3 things at once:
- Put some secure, AI-free checkpoints in every course.
- Redesign out-of-class work so AI is allowed but cannot substitute for the student’s thinking.
- Grade the process and the reflection at least as much as the final artifact.
That seems to be the emerging pattern in “AI-resilient assessment” practice.
Any unsupervised, out-of-class, text or code based task can be substantially done by AI. Trying to fight that purely with wording tricks is a losing game.
So the realistic premise is:
- Any digital take-home assignment is, by default, AI-susceptible.
- The question becomes:
Given that students can and will use AI, how do I design tasks where the only way to succeed is to actually understand the material?
Shift from “did you use AI?” to “did you use it thoughtfully, and do you understand the result?”
For a genomics or data science course, a single homework could require:
- AI use log. Students paste or upload their prompts and the model’s responses, or at least a cleaned transcript.
- Change history. Students highlight which parts of the code or text came directly from AI and which parts they edited.
- Critique. A short paragraph on where the AI solution was wrong, incomplete, or poorly justified in the context of the data or the biology.
- Justification. For one or two key steps, they explain in their own words why that step is appropriate, referencing course concepts or readings.
Now AI is in scope, but understanding cannot be skipped.
Make the task tightly tethered to artifacts, data, or context that only your class has.
Examples in your space:
- Local or private data. Use a UVA-only dataset, an internal registry, or simulated data you generate each year.
- Course-specific analysis pipelines. Have students build on a starter Quarto notebook, R package, or workflow that you provide and that is not public.
- Recent or unusual readings. Ask students to synthesize a preprint or dataset vignette that you post internally, rather than “explain PCA” at a general level.
Language models are weaker when the task depends on non-public, course-specific context.
Concrete levers:
- Multi-stage assignments.
- Stage 1: proposal or plan (AI allowed, log required).
- Stage 2: draft code and preliminary results.
- Stage 3: final report plus reflection on what changed.
- Version control as evidence. Require GitHub. Grade commit history, issue discussions, and pull request comments.
- Notebook hygiene. Grade labeling of figures, captions, and documentation in R Markdown or Quarto, not just final answers.
AI can help them type faster. It is much harder for it to fabricate a rich, coherent trail of evolving decisions that match your lectures and discussions.
You do not need to move everything into blue books. You just need enough secure assessment that students know their own understanding will be checked without AI.
Ideas for genomics courses:
- Short oral “vivas” tied to homework. Five minutes, once or twice a term.
- “Pick one figure from your homework. Walk me through how you produced it and what it shows.”
- In-class concept quizzes. Very short questions that hit the same concepts as the take-home work, but with different numbers or datasets.
- Whiteboard derivations. For quantitative topics, have them do a small derivation or sanity check in class that echoes a step in their project.
Old assignment:
Explain the difference between RNA-seq and microarray data, and describe how you would normalize an RNA-seq count matrix.
That is pure AI fodder.
Same learning objective, better prompts:
-
Use your own data:
Here is an RNA-seq count matrix from our in-class PBMC experiment and a microarray dataset on the same samples.
Compare the effect of normalization methods A and B on detection of differentially expressed genes.
Which pipeline would you choose for this specific experiment and why? -
Force them to critique AI:
Ask a generative AI tool to describe how to normalize RNA-seq data. Paste the answer, then identify at least three weaknesses or missing considerations in the context of the dataset we used in Homework 2.
-
Tie to professional practice:
You are consulting for a clinical lab that wants to introduce RNA-seq for a minimal residual disease assay.
Based on the readings and our in-class discussion, outline a validation plan.
Where, if anywhere, would you safely use AI tools in this pipeline, and where would you prohibit them?
All of these can involve AI, but success still demands course-specific understanding and judgment.
Make the AI policy assignment-by-assignment, not just syllabus-level.
Example structure for a genomics MSDS course:
-
AI free (certification).
- About 25% of the grade from in-class closed-laptop quizzes and short oral checks.
-
AI conditional (learning how to use it).
- About 35% from weekly labs where AI is allowed but logs and critique are required.
-
AI integrated (authentic practice).
- About 40% from a term project where you encourage tools like Copilot, but grade heavily on design decisions, documentation, and interpretation.
Students see clearly where AI is banned, where it is constrained, and where it is part of authentic practice.
- Stop trying to outsmart the models with clever wording.
- Add a few secure, AI-free checkpoints.
- For everything else, assume AI is in the loop and design so that learning requires:
- specific context or data only your class has,
- visible process and AI-use logs,
- personal judgment, reflection, and oral defense.
You do not get perfectly AI-proof take-home work. You get AI-aware, AI-resilient assessment that still lets you see what your students actually understand.