AI-Resistant Assignments: Design for Evidence, Not Detection
A process-evidence stack for designing assignments that reveal student thinking without making the false promise of AI-proof homework.
No take-home assignment is reliably AI-proof. If a final product can be produced outside the teacher's view, restrictions alone cannot show how it was made. The defensible response is to collect several forms of learning evidence across time.
An AI-resistant assignment makes unearned outsourcing less useful and genuine assistance easier to disclose. It does this through process evidence, local or individual context, brief explanation and a changed task—not through surveillance theater.
Begin With the Claim the Assignment Must Support
Write one sentence: “From this assessment, I want to infer that the student can ___ under ___ conditions.” If the target is source evaluation, a polished essay is indirect evidence unless the teacher can see how sources were selected and weighed. If the target is revision, only the final draft hides the learning.
The U.S. Department of Education's report on AI and the future of teaching and learning calls for human priorities and humans in the loop. Assessment design should start from the intended learning, not from the capability of the newest tool.
Use the Six-Layer Process-Evidence Stack
Local context
Use class data, a shared experience, a student observation or a defined audience.
Staged artifacts
Capture question, notes, plan, partial attempt, feedback and revision at meaningful points.
Decision log
Ask students to record consequential choices, alternatives and reasons.
Source verification
Require claim-to-source links and checks of authority, context and contradiction.
Brief defense
Students explain one choice, limitation or revision in their own words.
Transfer check
Change a condition and ask the student to apply the same thinking again.
Not every assignment needs all six. Choose the smallest combination that supports the intended inference.
Add Context That Matters to Reasoning
“Write about climate change” invites generic output. “Using our two-week shade and temperature observations, advise the school where to add one tree and explain the limits of the data” requires students to connect shared evidence to a decision.
Personalization should be optional and bounded. Never require sensitive family, health or identity disclosure simply to make AI use harder. Local context can come from a common demonstration, class text, provided dataset or public place.
Collect Stages That Represent Real Decisions
More checkpoints can become more paperwork without improving validity. Capture moments when thinking changes: problem definition, selection of evidence, first model, response to feedback and final limitation.
Use short artifacts—a photographed sketch, three annotated sources, a 90-second explanation or a paragraph explaining one revision. Sample some stages rather than grading every mark. The goal is a trace of authorship and learning, not a continuous recording of the student.
Make AI Use Inspectable
If AI is allowed, name the permitted job: clarify instructions, generate counterarguments, give feedback on an existing draft or help format a bibliography. Require a short use receipt listing the tool, purpose, one output accepted or rejected and how important claims were verified.
Do not require entire chat logs by default; they can contain personal data and create excessive review. A targeted receipt plus an ownership check is usually more educational.
Use Oral Defense as a Sample, Not an Interrogation
Ask two or three predictable questions: “Why did you choose this evidence?” “What is the weakest part?” “How would your answer change if this condition changed?” Students may respond live, in a short recorded explanation or through an accessible alternative.
Defense should confirm reasoning, not reward quick speech or punish anxiety, language development or disability. Share the question types in advance and apply documented accommodations.
Prefer Transfer Checks to Style Policing
A student can imitate a teacher-approved style or edit AI text until it appears ordinary. A changed task asks whether the underlying knowledge travels. After a data analysis, alter one assumption. After an argument, introduce a credible conflicting source. After a math explanation, change the representation.
The check can be brief and partly in class. It often reveals more than another page of final prose.
Do Not Make Detection the Evidence
AI-detector scores do not demonstrate which parts a student authored or whether learning occurred. They can be one uncertain signal for a conversation, never the sole basis for a consequential accusation.
UNESCO's guidance on generative AI in education emphasizes human-centered validation, privacy and pedagogical design. Use direct process evidence, the school's due process and a repair opportunity when evidence conflicts.
Redesign One Existing Assignment
Ten-minute redesign
- Circle the exact capability being assessed.
- Delete requirements that add length but no evidence.
- Add one common local input or constrained audience.
- Capture one early decision and one meaningful revision.
- State which AI actions are allowed and how they are disclosed.
- Add one two-minute defense or transfer question.
- Decide what will be sampled rather than fully graded.
The result should often be shorter than the original assignment and stronger as evidence.
Resistance Comes From Better Assessment
An assignment is AI-resistant when the easiest route to success is to participate in the learning process and retain ownership of key decisions. Students may still misuse tools; no design eliminates dishonesty. But the teacher no longer relies on a single polished artifact to carry every inference.
That is the deeper shift: from trying to prove what software touched the page to gathering credible evidence of what the student can understand, decide and do.
Frequently Asked Questions
-
Not reliably for unsupervised work. Teachers can make outsourcing less useful and collect stronger evidence through local inputs, staged decisions, brief defense and transfer checks.
-
Use the learning target to decide. Some assessments need an unaided condition; others can assess verification, revision or responsible AI collaboration. State the condition explicitly.
-
A detector score is not proof of authorship and should not determine a grade or misconduct finding by itself. Follow school policy, review direct evidence and give the student a fair chance to explain.
-
It can if every artifact is graded. Collect short stages, sample strategically, use quick conferences and grade only the evidence aligned to the target.
-
Assess reasoning rather than speech speed. Share question types, allow appropriate response modes and implement IEP, 504, language and other documented supports.