If a chatbot can finish the assignment in thirty seconds, the fix isn't a better detector — it's a better assignment. Here's exactly what that looks like, task by task.

You assign a 750-word essay. A student asks ChatGPT and has a polished draft in thirty seconds. For a task that goes home with the student, you really have two options: try to detect the AI use after the fact, or redesign the task so the shortcut no longer works.

Detection is the weaker of the two. In studies where researchers slipped fully AI-written work into real exam systems, roughly 90 to 94% of it passed undetected — and detection tools flag plenty of honest students by mistake, which is why even K-12 guidance now advises against relying on them.

This guide is four real before-and-after redesigns — a history essay, a math problem set, a science lab, and a novel response. Each follows the same six steps so you can copy the move, not just the idea.

Old assignment: "In 750 words, explain the main causes of World War I."

Why AI breaks it: This is recall and summary of public information — exactly what a chatbot does best. It writes a competent 750 words instantly, and the finished essay gives you no way to tell.

Redesigned assignment: "Using the three sources we annotated in class this week, argue which cause mattered most. Reference at least two points from Tuesday's class debate, and explain how your view changed from the start of the week to now."

What students hand in: Their annotated sources, the essay, and a short note on how their thinking changed.

What you grade: The argument and the use of the specific class sources — plus the change-of-mind note, which shows real thinking, not a polished product.

Why it works: The task now depends on things the AI was never part of: those specific annotated sources and that specific debate. A student can still use AI to check grammar, but it can't reconstruct Tuesday's class.

Old assignment: Twenty practice problems for homework, graded on correct answers.

Why AI breaks it: Chatbots now solve most standard problems and show their work. When problems are graded only on final answers, the assessment measures whether students found a tool, not whether they learned the method.

Redesigned assignment: Keep a short practice set that isn't graded (it's for them). Then assess understanding with an "find the error" task: give students an AI's worked solution that contains a mistake, and ask them to find it, fix it, and name the rule that was broken. For example:

What students hand in: Their corrected solution and a one-sentence explanation of the rule the AI got wrong.

What you grade: Whether the student can spot the error and explain the underlying rule — which proves they understand the method, not just the answer.

Why it works: Catching someone else's mistake is harder to fake than producing an answer, and it tests exactly the judgment students now need. (AI gets things wrong often enough for this to be real: in one study a popular model was wrong about a third of the time in the subjects tested.)

Old assignment: "Write up the experiment: hypothesis, method, results, conclusion."

Why AI breaks it: Given the experiment's name, AI generates a flawless-sounding write-up — including a method and results it never observed. The polish can hide whether the student understood anything.

Redesigned assignment: Students write the results and conclusion from their own group's actual data (which differs from every other group's, and from anything AI would invent). They include a photo of their setup and answer one follow-up they couldn't predict: "What would you change if you ran this again?"

What students hand in: The write-up built on their own data, the setup photo, and the follow-up answer. A two-minute talk with the group confirms they can explain their own numbers.

What you grade: Whether they can interpret their own results and reason about the method — not whether they can produce a clean report.

Why it works: A generic AI write-up doesn't fit data only that group generated, and the quick verbal check surfaces who really understands it. Teachers who use short verbal checks report fewer integrity problems and say students find them more fair.

Old assignment: "Write a 500-word response to the themes of To Kill a Mockingbird."

Why AI breaks it: The themes of a famous novel are all over the training data. AI produces a tidy thematic essay instantly, and it reads fine.

Redesigned assignment: Build the response in stages, some on paper in class. Students annotate the courtroom scene, draft their own thesis, and write. As a final layer, they bring an AI-generated reading of that same scene and write about where they disagree with it and why.

What students hand in: Their in-class annotations, their thesis, the essay, and their written disagreement with the AI's reading.

What you grade: Their own interpretation and the quality of their disagreement with the AI — that's the graded object, not a fresh essay.

Why it works: The staged, partly-in-class work makes the thinking visible step by step, and arguing against an AI reading is much harder to outsource than writing one.

Across all four, the shift is the same: from the finished product toward the thinking behind it, and from basic recall toward judgment. When a machine can produce the product on demand, the product proves less. What proves more is whether the student can build it, defend it, and catch its mistakes.

A useful way to think about it: split your assessment into two kinds. Some work should happen under conditions you can trust — in-class writing, a quick oral explanation. The rest can stay open-ended and even AI-assisted, because there your goal is to grade how well a student thinks with the tool. And keep some basics genuinely AI-free — students still need knowledge they actually own, not just the ability to direct a tool.

A fair worry: you can't sit down one-on-one with every student, or redesign every task — especially if you teach several classes a day. You don't have to. Add one small proof-of-thinking step to the handful of assignments where an AI shortcut would do the most damage — a change-of-mind note, a setup photo, a two-minute talk.

Take the single assignment where AI shortcuts hurt your students most, and run it through the six steps above: name the old version, see why AI breaks it, redesign it, decide what students hand in and what you grade, and check that it works. The shift from "did you use AI?" to "show me your thinking" is the whole game.

📎 Free download: "Four Homework Redesigns, Before & After" — the four examples on one printable page.

Disclosure: Hossein works in AI and builds AI-related products. AI by Age takes no AI-vendor sponsorships. Full disclosure on our About page.