Assignment Design for STEM: Making Lab Reports and Problem Sets AI-Resistant

A chemistry instructor I know stopped trusting AI detectors the week a polished lab report sailed through as “human” and a struggling student’s honest one got flagged. The problem wasn’t the tool—it was the assignment. Reports that ask for the same five paragraphs every term are the easiest thing in the world for a model to fake, and the hardest thing for a detector to judge.

The better move in STEM isn’t more surveillance. It’s designing tasks where a copied answer is obviously empty, because the real answer depends on something only that student has: their data, their reasoning, their mistakes.

Key takeaways

• Detectors are a weak backstop in STEM—they output probabilities, misfire on short or formulaic text, and have been shown to flag non-native English writers more often. Design beats surveillance.
• Anchor tasks to something the student personally generated: their own measurements, a specific dataset, a local constraint, or a recorded in-class moment.
• Problem sets resist AI best when they ask for reasoning, error analysis, and justification—not just a final number.
• Build process into the grade: raw data, intermediate steps, revision history, a short defense.
• Tell students plainly what’s allowed. Clear policy plus authentic tasks beats a quiet detector arms race.

Why detectors fail STEM specifically

STEM writing is close to the worst case for AI detection. Methods sections are short and templated. Equations and standardized phrasing keep perplexity low—and detectors read low-perplexity, formulaic text as machine-made, whether a person or a model wrote it. So a careful, by-the-book human report and a generated one can land in the same suspicious zone.

It gets worse. A 2023 study led by James Zou’s group at Stanford found AI detectors flagged writing by non-native English speakers far more often than native speakers’ work—more than half of the TOEFL essays they tested were wrongly tagged as AI-generated. A large share of STEM students are writing in a second language. And OpenAI pulled its own AI Text Classifier in July 2023 for low accuracy. If the company that builds the models can’t reliably detect them, a checkbox in your LMS won’t either.

None of this means students aren’t using AI. It means you can’t outsource the judgment to a score. You have to build the judgment into the task.

Make the work depend on the student’s own data

The single strongest design principle for STEM is this: tie the deliverable to something the student personally produced and you can verify.

For lab reports, that means grading the analysis of *their* numbers, not a generic write-up of the procedure. Two students who ran the same experiment will get different readings, different error bars, different anomalies. A model handed a lab prompt has no idea that this particular group’s spectrophotometer drifted, or that their reaction ran two minutes long.

Mini-scenario. Instead of “Write a lab report on the synthesis of aspirin,” try: “Report the percent yield *you* measured. Explain at least two specific steps in your procedure that most likely cost you product, using the masses you recorded.” Now the report can’t exist without the student’s own bench data. A generated paragraph about “potential sources of error such as incomplete filtration” reads as filler the moment it isn’t tied to the actual grams they lost.

Practical ways to anchor work to real data:

• Have students record raw measurements in a shared sheet or photographed lab notebook *during* class, then submit those alongside the report.
• Assign each student or group a slightly different dataset, sample, or parameter, so there’s no single “right” report to copy.
• Require one graph built from their own numbers, with a caption explaining one feature of it.

Grade reasoning, not the final answer

A model is great at producing a correct final number. It’s much weaker at justifying a specific choice, estimating uncertainty, or diagnosing a particular mistake. So shift the points there.

On a problem set, the answer “v = 14.2 m/s” is trivial to get from any chatbot. But “explain which assumption you dropped to get a closed-form solution, and what error that introduces” forces reasoning that’s hard to fake convincingly—and easy for you to assess.

Small edits to existing problems do most of the work:

• Add “justify your choice of method” or “state your assumptions” to problems you already use.
• Ask for uncertainty: “estimate the error and say where it comes from.”
• Use real, messy numbers from a dataset instead of clean textbook integers.
• Invert the task: “A classmate got 9.8 × 10⁴. Find and explain their error.” Debugging someone else’s reasoning is a skill models do unevenly, and it’s genuinely useful.

Mini-scenario. A statistics instructor swapped “calculate the p-value” for “you ran this test and got p = 0.06. Your lab partner says the result is significant. Write three sentences telling them whether to trust that claim and why.” The numeric work is tiny; the reasoning is the assignment. Copied answers came back generic and sometimes contradicted the specific value in the prompt, which made them easy to spot.

Build process into the grade

Authentic assessment treats the final document as one artifact among several, not the whole grade. When intermediate work counts, a clean final answer with no supporting trail becomes its own red flag—and honest students get a record that protects them if they’re ever wrongly suspected.

Things worth grading or collecting:

• Raw data and notebooks. Photographed bench notes, instrument printouts, or a timestamped data sheet.
• Intermediate steps. Partial derivations, discarded approaches, a sketch of the setup.
• Revision history. Drafting in a Google Doc or version-controlled file shows the work taking shape over time. A document that appears fully formed in one paste is at least worth a conversation.
• A short defense. Two or three minutes explaining one design choice, or a single written paragraph answering a question you pose *after* submission.

The oral defense is the great equalizer. A student who did the work can explain why they chose a log scale or dropped an outlier in fifteen seconds. A student who didn’t, can’t—and you learn this without ever opening a detector. Keep it low-stakes and quick so it stays manageable for a full section.

A redesign worked example

Here’s a standard physics lab turned AI-resistant without inventing a whole new course.

Before: “Measure g using a pendulum. Submit a report with introduction, methods, results, and conclusion.”

After:

1. In class, each pair times their own pendulum and logs ten raw periods in the shared sheet (data is personal and witnessed).
2. The report asks: “Your measured g was [value]. The accepted value is 9.81 m/s². Account for your specific discrepancy—was it timing reaction lag, amplitude, or length measurement? Use your numbers.”
3. They submit the raw timing sheet and one plot they built.
4. A one-paragraph follow-up question, posted after submission: “If you halved your pendulum length, predict what your measured period would do, and check it against your data.”

Same equipment, same concept, same grading time. But every layer now leans on something a model can’t supply from the prompt alone. The detector never enters the picture, and a falsely accused student has a paper trail that speaks for itself.

Be transparent about AI use

Resistant design works better when paired with a clear, specific policy—not a vague “no AI” that students ignore. Name what’s allowed (checking grammar, getting a concept explained, debugging code they wrote) and what isn’t (generating the analysis or write-up), and say *why* the assignment is built the way it is. A short disclosure line where students note any AI use removes ambiguity and keeps the focus on learning instead of catching people.

This isn’t about trusting students less. It’s about removing the incentive to cut corners by making the corners not worth cutting, and removing the fear of false accusation by grading evidence instead of vibes. If you want to understand how the tools you’re tempted to rely on actually behave, there’s more on AI detection and writing worth reading before you mandate anything.

Frequently asked questions

See the structured FAQ above.

Closing

You won’t make a STEM assignment perfectly AI-proof, and chasing that with detectors mostly produces false positives and stressed students. What you can do is design tasks where the real work lives in data, reasoning, and process a model can’t borrow—so a copied answer is just obviously hollow. That’s a better use of your time than refereeing a score.

If you want to see how a piece of writing reads to a detector before you trust one in your classroom, check how a sample draft reads and judge the signal for yourself.