AI Resistant Assignments: Designs That Make Authorship Visible
The phrase "AI resistant assignments" usually arrives in a department meeting after a detector has flagged a paper and nobody can prove anything either way. This guide is for English teachers and department heads who want a different starting point: assignments built so that authorship is visible while the work is being done, rather than adjudicated afterward from a percentage. It covers staged drafts, annotated reading, blue book exams, oral defenses, process portfolios and disclosure statements, with the cost of each in time, equity and access.
In short: the most durable AI resistant assignments make the student's process the thing being assessed rather than trying to detect generated text: dated drafts with feedback between them, annotated reading, some writing done in the room, a short conversation about the argument, and a portfolio that shows the essay being built over weeks.
What "AI resistant" should mean
An assignment cannot be made immune to a language model; any prompt that can be typed can be pasted. What an assignment can do is change what counts as evidence. A single final document proves only that a document exists. A sequence of drafts, a marked-up text, a five-minute conversation and a reflection on what changed are evidence of a person thinking over time, and no tool produces them on the student's behalf without leaving obvious gaps.
That shift also removes the worst feature of the detection era: the false accusation. A design that never needs the score never has to defend it. A University of Virginia history professor quoted in the Cavalier Daily in May 2026 said she moved to in-class blue book assessment partly because students had told her they feared being falsely accused of using AI. Visibility protects honest students as much as it constrains dishonest ones.
Six designs for AI resistant assignments
Staged drafts with feedback between them
Collect a thesis sentence in week one, an outline with quotations in week two, a full draft in week three and a revision in week four, with brief feedback at each stage that the next must answer. A draft generated at stage three has to explain why it ignores the feedback from stage two; a student who wrote it can point to the sentence that changed. Our guide to essay drafts as evidence is written for students and describes what this design produces.
Annotated-reading submissions
Before the essay, collect the text itself: a photographed copy of the assigned chapters with the student's dated marginal notes and underlinings. For Macbeth, ask for Act 1 Scene 7 marked up with every place Lady Macbeth changes the subject; for Lord of the Flies, every appearance of the conch across four chapters. The annotations are the reading, and an essay that does not grow out of them is visibly rootless. Our Macbeth guide models what a marked-up reading looks like once it becomes an argument.
In-class essays and the return of blue books
Handwritten, timed writing has the longest history and the clearest evidence. The Chronicle of Higher Education reported in September 2026 that sales of exam booklets at the more than 1,000 stores run by campus bookseller Follett were up nearly 40 percent since the fall of 2023, and that Roaring Spring Paper Products, the last remaining producer of blue books, had seen sales climb above pre-pandemic levels as instructors returned to in-person testing. The University of Virginia bookstore reported a roughly 27 percent rise over the same year.
The Chronicle piece also records the objections: a Berkeley computer scientist called the move "a retreat", and a mathematics professor at Rochester Institute of Technology described what is lost when open-ended take-home work disappears, the understanding that the struggle is valuable. An in-class essay proves who wrote these sentences in this hour and little about sustained revision. Use it for one assessment in a term rather than all of them.
Short oral defenses
Five minutes per student, after submission: what is your thesis, which piece of evidence was hardest to fit, what did you cut and why. A student who wrote the essay answers in seconds; one who did not has nothing to say about the cutting. This is the cheapest design per unit of certainty and the most exposing, so it needs a scripted, low-stakes format announced at the start of term rather than sprung on a suspect.
Process portfolios
Bundle everything above into one submission: the annotated text, the staged drafts, the feedback, the final essay and a page of reflection. Grade the portfolio with a rubric that gives real weight to revision. This is the design that most changes what students believe the assignment is for, because the final draft becomes the last page of a longer piece of evidence.
Reflective disclosure statements
Require a short paragraph at the end of every submission stating what tools were used, for what, and what the student did with the result. A grammar checker, a citation generator and a brainstorming conversation are different acts, and a policy that names them lets students be honest about small uses instead of hiding all of them. The disclosure statement template on this site takes a student five minutes.
AI resistant assignments compared
| Design | What it proves | Teacher time | Equity and access concerns |
|---|---|---|---|
| Staged drafts with feedback | The argument developed over weeks in response to comments | High: three or four short reads per student instead of one | Low; works at home and on any device, but needs a stable submission platform |
| Annotated-reading submission | The student read the text closely before writing | Low to moderate; skim rather than grade | Needs a printable text and a camera; screen-reader users and dyslexic students need an alternative format |
| In-class essay or blue book exam | Who wrote these sentences in this room | Low to set, moderate to mark handwriting | Highest: disadvantages slow writers, students with motor or processing differences, and second-language writers; requires accommodations by design |
| Short oral defense | The student can explain and defend the submitted argument | Moderate: five minutes each, scheduled | Stressful for anxious and neurodivergent students; script it and allow a written alternative |
| Process portfolio | The whole arc from reading to revision | High at the end of term, lower during it | Low if pieces are collected as they happen; heavy if assembled in one rush |
| Disclosure statement | What help was used, in the student's own account | Very low | None, provided honesty is not punished |
Trade-offs to plan for
Time is the first cost. Staged drafts multiply reading; afford them by keeping each stage short (a sentence, an outline, a page) and replying in two lines. Portfolios cost less if assembled continuously.
Equity is the second. The designs that prove the most in the moment, in-class essays and oral defenses, also penalise students who write slowly, process slowly, or write in a second language. Build accommodations into the design rather than granting them on request: extra time as standard, a typed option in a supervised room, a written alternative to the oral defense. A department that only adds blue books has replaced one unfair instrument with another.
Access is the third. Annotated-reading submissions assume a printable text and a camera; version-history evidence assumes a laptop and a stable account. Keep one design in the mix that works with paper alone and one that works with a phone alone.
Where the work guides and essay-craft pieces fit
A course built this way needs models of a real reading, because students who have only seen finished essays do not know what to put in the margins. Our work guides are close readings organised by scene and chapter, which makes them usable as annotation targets: assign the guide's reading of one scene, then ask students to mark up the scene and disagree with it in writing. The essay-craft piece on close reading as a method gives the procedure, and the per-work articles in this lane, such as writing a Macbeth essay that is yours, show a generic paragraph beside a close-read one so that students can see the difference before they are graded on it.
Why this beats relying on detector scores
The research on detection is consistent. The 2023 study by Weber-Wulff and colleagues in the International Journal for Educational Integrity tested 14 detection tools and found all of them below 80 percent accuracy, with only five above 70 percent, and documented both false positives and false negatives. A 2023 Stanford study by Liang and colleagues, published in Patterns, found that widely used GPT detectors consistently misclassified essays by non-native English writers as AI-generated while classifying native samples correctly, and cautioned against using detectors in evaluative or educational settings.
A department that leans on scores inherits two problems at once: it will miss a share of generated work and accuse a share of honest students, with the accusations falling hardest on writers who already have the least margin. Our guides to AI detection false positives and what teachers actually see in a Turnitin AI report go through the mechanics. A process-based design sidesteps the question: if the drafts, the annotations and the conversation are there, the score is irrelevant, and if they are missing, the score was never the evidence anyway. The Authorship pillar collects the student-facing side of the same argument.
Frequently asked questions
Are blue book exams enough on their own?
No. A blue book exam proves who wrote a set of sentences in one sitting and nothing about revision, research or sustained argument. Use it as one assessment among several, with extra time and a supervised typed option built in from the start.
How much extra marking do AI resistant assignments create?
Staged drafts and portfolios add reading, but each stage can be short and feedback can be two lines. Oral defenses cost about five minutes per student and disclosure statements almost nothing.
Should we stop using AI detectors entirely?
Use them, if at all, as a prompt for a conversation and never as evidence. Published testing puts every tool below 80 percent accuracy, and the false positives fall disproportionately on second-language writers.
AI resistant assignments as a design decision
The question a detector tries to answer, did a person write this, is one an assignment can be built to answer on its own. Staged drafts, annotated reading, a little writing in the room, a short conversation and a portfolio that holds them together make authorship visible while the work is happening. AI resistant assignments of that kind cost time and care about equity, and they repay both by making the false accusation, and the real one, equally unnecessary.
The essays and guides on this site are study companions for reading and teaching; they are not coursework to be submitted as anyone's own.
Sources
- Haley Zimmerman, "Blue Books Are Back. Is That a Good Thing?", The Chronicle of Higher Education, 4 September 2026: Follett's near 40 percent rise in exam-booklet sales since fall 2023; Roaring Spring as the last producer; faculty objections.
- Lidia Zur Muhlen, "Blue books return in an effort to combat AI usage", The Cavalier Daily, 6 May 2026: UVA's 27 percent rise in blue book sales; a professor's account of students fearing false accusation.
- Weber-Wulff et al., "Testing of detection tools for AI-generated text", International Journal for Educational Integrity (2023): 14 tools, all below 80 percent accuracy, false positives and negatives.
- Liang, Yuksekgonul, Mao, Wu and Zou, "GPT detectors are biased against non-native English writers" (2023; Patterns): detectors misclassify non-native English writing; caution against evaluative use.