How to redesign class assignments in the AI era

TLDR: Students have access to AI and that will not change: the question is no longer how to prevent it but what to assess now. Automatic detectors produce false positives and cannot sustain an accusation. What works is redesigning: assessing process and oral defense, designing tasks where AI is allowed and judgment is graded, and verifying comprehension in the classroom. This guide uses the same tools students use, to redesign with full knowledge.

This guide is for secondary, vocational and university teachers grading homework and no longer sure what they are grading. The discomfort is legitimate and the reflex response (ban and detect) does not work, so the honest move is redesigning.

The starting position matters: this is not about surrendering or about hunting. It is about returning to the eternal question, what evidence do I need that this student learned, knowing the context changed.

1. Accept the real starting point

Before redesigning anything, do the uncomfortable exercise: take your three usual assignments and ask the AI for them. YesChat serves for this on its free plan, and it also lets you compare what GPT and Claude produce with the same brief, which is exactly what your students will do.

The exercise’s result is usually this: the generic text commentary, the summary and the documentation-based essay come out at a solid pass level in seconds. What comes out wrong or empty: everything demanding classroom data, personal experience, justified decisions and knowledge of what happened in class.

That map is your redesign material. The first group of assignments no longer assesses what you thought: it assesses who used the tool best, and that demands deciding whether you accept it as part of the task or change the task.

2. Redesign toward process and defense

The assignment submitted as a single final product is the most vulnerable format. The formats that hold up share one trait: they make the process visible and require the student to stand behind their work in person.

The concrete pieces: staged submissions with commented drafts, where the impossible jump between draft and final exposes itself. The decision log, two lines per session on what was done and why. And above all the brief oral defense: three questions about their own work distinguish in two minutes the student who did it from the one who submitted it.

To manage defenses in large groups, Curipod lets you run sessions where the whole group answers written questions about their own work simultaneously, on its free teacher plan. It does not replace the individual conversation, but it screens where one is needed.

3. Make AI part of the assignment where it makes sense

In many subjects, banning the tool students will use their whole professional lives is preparing them for a world that does not exist. The alternative is the task where use is allowed and what gets graded is judgment: the part the tool does not supply.

The formats that work: requiring the work with AI plus a critique of the result (what is wrong, what is missing, what you would improve and why). Comparing two generated answers and arguing which is better. Starting from a generated text with planted errors and correcting it. In all of them, the grade sits in the student’s judgment about the material, which cannot be delegated.

This approach demands explicit rules: what is allowed in each task, how use is declared and what happens when rules are broken. The current ambiguity, where each teacher assumes something different, is unfair in both directions.

4. Verify in the classroom, not with detectors

The detector-software temptation deserves a clear dismissal: generated-text detectors produce documented false positives, flag second-language writers disproportionately and cannot sustain a formal accusation. Building disciplinary decisions on that base is unjust and, moreover, indefensible.

The verification that does work is pedagogical and in person: regular in-class writing, which gives the real sample of each student’s level, and questions about the submitted work itself. For subjects with exams, Quizgecko generates quick checks from the assignments’ content, closing the loop between what was submitted and what is known, from its free tier.

The whole system’s goal is not catching anyone: it is making copying without learning unprofitable, because work submitted without understanding collapses at the defense. Formal unit assessment gets built with the exams and quizzes guide, and the rest of the sector lives in AI for education and training.

Common mistakes

Trusting detectors. Documented false positives, bias against second-language writers and zero evidentiary value. An accusation built on a detector is a problem waiting for its first appeal.

Banning without verifying. A ban without oral defense or in-class writing only penalizes the students who respect it.

Redesigning everything toward the in-person exam. It is the comfortable reaction and it impoverishes assessment: long work with visible process measures things no exam measures.

Not making the rules explicit. If students do not know what is allowed in each task, half will use too much and half too little, and neither half learns what was intended.

Frequently asked questions

Are AI detectors useless?

As a signal for a conversation, weak. As proof for sanctions, worthless: their false positives are documented and they punish second-language writers disproportionately. Valid verification is in person.

How much work does this redesign add?

Less than it seems: three-question defenses fit in one session, drafts get skimmed and this guide’s tools generate the checks. What disappears is the long grading of assignments that no longer informed anything.

From what age should AI be allowed in assignments?

That is a school and stage decision, not this guide’s. The general criterion: base skills first (writing, summarizing, arguing), then the tool with explicit rules and tasks designed for it.

What do I do if I suspect a specific assignment?

Conversation and questions about the work, not accusation. The student who did it answers naturally about their decisions. The one who did not shows it in two minutes. And that evidence, unlike the detector’s, holds up.