AI-Proof Assignments: What Still Works in 2026
There is no assignment a determined student can't run through a chatbot. The realistic goal isn't AI-impossible — it's AI-unhelpful: designs where using AI either doesn't save effort, produces obviously weak work, or is visible in the process. These eight patterns hold up in real classrooms.
1. Grade the process, not just the product
Require the artifact trail: outline, messy first draft, revision with tracked changes, and a short reflection on what changed and why. A polished essay with no believable drafting history is its own red flag — and students who do the process learn the thing, which was the point. (Our writing guides give students a process worth showing.)
2. Anchor part of the work in class
A 20-minute handwritten or lockdown-browser response on the same topic as the take-home essay gives you a writing baseline per student. When the take-home reads three grade levels above the in-class sample, you have evidence — not vibes — for a conversation.
3. Demand local, personal, or current specificity
Models are weakest where training data is thinnest: this week's class discussion, your town's specific policy debate, an interview the student conducts, data the class generated in lab. "Argue about social media" is a chatbot's home turf; "argue about the phone policy our principal announced Tuesday, citing the student-survey data from Friday" is not.
4. Attach a two-minute oral defense
Randomly select students (or flag-based) to answer three questions about their own essay: "Why this source? What did you cut? Defend your weakest paragraph." Students who wrote it answer instantly. The mere existence of the defense changes behavior more than any detector.
5. Require sources AI can't fake — and verify them
Class texts with page numbers, a librarian-verified database source, an interview. Then actually check bibliographies: invented references are the most objective AI tell there is, because models generate citation-shaped text without looking anything up. Our fake-citation checker verifies a bibliography in seconds — and telling students you run it changes what gets submitted.
6. Make AI use a graded, disclosed step
The judo option: require students to get AI feedback on their draft, include the transcript, and write about what they accepted and rejected. This converts the covert shortcut into visible critical thinking — and teaches the tool-literacy they'll need anyway. Clear disclosure rules beat blanket bans that push use underground.
7. Use formats models still fumble
Annotated maps, podcast episodes with show notes, whiteboard explainer videos, data-analysis of a dataset the class collected. Multimodal and process-heavy formats aren't AI-immune, but the AI-assisted path stops being the lazy path.
8. Shrink the stakes-per-deadline
One 40%-of-grade essay due in three weeks is a machine for producing desperation shortcuts. Four small deliverables with feedback between them lower the payoff of cheating on any single one — and produce better writing besides.
Where detection fits (a realistic policy)
Detectors are a signal, not a verdict — the position we hold while running one. A sensible classroom policy: publish that you use detection tools, run flagged work through a sentence-level detector, pair any flag with process evidence (drafts, in-class baseline, oral defense) before acting, and never let a score alone decide a grade. False positives are real — here's what students experience when policy skips that step — and non-native writers are flagged disproportionately. Detection works best as one leg of a stool whose other legs are assignment design and knowing your students' voices. More teacher workflows on our teachers page.
Frequently asked questions
Can any assignment be truly AI-proof?
Fully proof, no — supervised, handwritten, or oral work comes closest. But "AI-unhelpful" is achievable: process requirements, local specificity, and verification steps make honest work the path of least resistance.
Should teachers ban AI or teach with it?
The emerging consensus is both, by assignment: some work is explicitly no-AI (skill building, assessed in conditions you control), some is AI-disclosed (drafting with documented assistance). What fails is ambiguity — students treat unstated rules as permission.
How do I check if a student used AI?
Triangulate: a detector scan for the statistical read, a citation check for fabricated sources, the writing baseline from in-class work, and a short conversation about the essay's choices. Any one alone is weak; together they're fair and hard to argue with.