A detector tells you, with some probability, whether a finished document looks machine-generated. It cannot tell you whether the student understood the source, whether the argument is theirs, or whether the writing represents growth from where they started the term.
Quick answer
The most reliable defense against AI in writing assignments is not detection. It is assignment design that produces visible evidence of student engagement: in-class scaffolding, process portfolios, source-specific prompts, oral defense, multi-modal artifacts, and peer review checkpoints. Each adds modest faculty time and produces better learning outcomes whether or not AI is in the room.
None of these patterns require a new platform, a new policy, or a tool purchase. They require thinking about the writing assignment as a sequence of observable acts rather than a single submitted document.
Why Detection Alone Doesn't Hold
Paraphrase attacks compound the problem. A 2023 paper by Krishna and colleagues showed that running AI-generated text through a paraphraser substantially degrades detector accuracy. The detector arms race is not one teachers can win by buying better software.
The pedagogy argument runs in the opposite direction. If you can see the student doing the work, you do not need to prove they did it after the fact. The Inside Higher Ed essay from April 2026 makes this point bluntly: the best defense against AI cheating is teaching that is harder to phone in.
The redesign that survives next term's AI tooling is the one you would have wanted to do anyway.
Editorial board, Working Educators
The Six Patterns
The first pattern is in-class scaffolding. Break the writing into stages with at least one stage completed in the room. A brainstorm session on paper, a thesis draft written in fifteen minutes during class, an outline produced during a workshop. The artifact does not need to be polished. It needs to exist, with a date and a context, so that the final paper has visible roots in observed student work.
The second is the process portfolio. Students submit not just the final essay but the trail: notes from the readings, an annotated source list, two draft versions with timestamps, a reflection on what changed between drafts. This is what some process-based assessment research calls evidence-of-thinking, and it works because fabricating a coherent process trail is meaningfully harder than generating a single polished output.
The third is source-specific prompts. Ground the assignment in a text or dataset the class read together in a particular week, and require quotation with page numbers and analysis of specific passages. A generic prompt produces a generic answer. A prompt that asks how the protagonist's silence in chapter seven complicates the reading of the final scene cannot be answered without the book in hand.
The fourth is oral defense. Five to ten minutes, one-on-one or in small panels, where the student walks through the argument of their paper and answers a couple of clarifying questions. Our piece on oral defense as assessment covers the logistics.
The fifth is multi-modal artifacts. Pair the essay with a short recorded explanation, a slide deck, an annotated map, or a hand-drawn diagram. The multi-modal piece does not have to carry the grade. It anchors the written work to a creative act the student can speak to.
The sixth is peer review checkpoints. Structured peer review at draft stage, with specific questions students answer about each other's work, produces a paper trail of engagement and improves the writing whether or not AI is involved. The Harvard generative AI guidance for faculty includes peer review as one of its core recommendations.
What Each Pattern Costs
The faculty time question is the one most pedagogy advice ducks. In-class scaffolding costs roughly fifteen minutes of class time per stage, with no additional grading load if you collect the scaffolding as participation. Process portfolios add maybe ten minutes of grading per student because you are skimming a trail, not deeply reading two drafts.
Source-specific prompts cost nothing in time and may actually save grading time because the answers are more comparable to one another. Oral defense is the most expensive: five minutes per student times twenty-five students is roughly two hours per assignment. The mitigating factor is that for many teachers, those two hours replace time previously spent adjudicating suspected AI cases.
Multi-modal artifacts cost setup time the first semester and almost nothing thereafter. Peer review checkpoints cost the class time it takes to run them, usually one class period at the draft stage, and produce better drafts that are faster to grade.
A Two-Week Redesign Sequence
For a teacher with one writing assignment to redesign before next term, here is a sequence that fits into two weeks of preparation time. Week one, day one: write down what you actually want the student to demonstrate. Not the topic, the capacity. Day two: choose two of the six patterns above that align with that capacity.
Day three through five of week one: rewrite the assignment prompt to require those two patterns. Day one through three of week two: rewrite the rubric. The rubric needs to grade the patterns you added.
Day four and five of week two: pilot the new prompt with a colleague. Have them read it as if they were a student and tell you what they would do. Their confusion is data. Adjust the prompt before students see it. The whole sequence is six to eight hours of work spread over two weeks.
The Pedagogy Argument Independent of AI
Each of these patterns was advocated by writing teachers long before ChatGPT existed. Process portfolios are a forty-year-old tradition in composition pedagogy. Source-specific prompts are basic close reading. Oral defense is how graduate humanities programs have assessed their students for a century.
The reason to adopt them now is not that AI forced your hand. It is that AI gave you the political cover to adopt teaching practices that were already better. A redesigned assignment produces students who can speak to their own arguments, who have a record of their own thinking, who have practiced revision rather than first-draft submission.
If a redesigned writing curriculum also happens to be more resistant to AI generation, that is a side effect. The primary justification is the one a teacher should be able to make in any decade: this produces better writers.
Frequently Asked Questions
What if I do not have time to redesign all my assignments?▼
Pick one. The piece walks through a two-week sequence for redesigning a single assignment. Start with the assignment you most suspect is being generated.
Does this mean I should stop using AI detectors entirely?▼
No. Detectors are still useful as one signal among several, particularly when combined with the kind of process evidence these patterns produce.
Won't students just use AI to write the process artifacts too?▼
Some will try. The friction is meaningfully higher. Generating a single polished essay is one prompt. Generating two timestamped drafts, a reading reflection, and an annotated source list that all hang together coherently is several prompts and substantial assembly work.
What about students who do worse with oral defense because of anxiety?▼
Pair oral defense with a written option for students with documented accommodations, and keep the stakes modest. The point is to gather evidence of engagement, not to run a high-pressure exam.
How do I explain this redesign to skeptical colleagues?▼
Lead with the pedagogy argument, not the AI argument. Process portfolios, source-specific prompts, and peer review have decades of research behind them as practices that improve writing instruction.
The Bottom Line
The most useful frame for any teacher thinking about AI in writing assignments is to stop treating it as a discipline problem and start treating it as a design constraint. Detection is a downstream patch on an upstream design choice.
Six patterns hold up: in-class scaffolding, process portfolios, source-specific prompts, oral defense, multi-modal artifacts, and peer review checkpoints. None require a tool purchase. All produce better writers, with or without a chatbot in the room.
The pedagogy argument was already correct before AI made it urgent. AI is the deadline, not the reason.
