Learn · Education

For Teachers: Assignments in the AI Era

The essay did not die. The essay as proof of work did.

Teachers got the hardest version of the AI transition: the same year the tools upended what assignments prove, they also became the profession's most useful assistant — and both changes arrived mid-semester, with no training day. This guide takes the two halves in turn: redesigning assessment now that a machine can produce competent homework, and putting the machine to work on the teacher's side of the desk, where the hours have always been the scarcest resource.

What actually broke

Precision helps here, because less broke than the panic suggested. AI did not break *teaching writing*, *teaching problem-solving*, or *the value of practice*. What broke is narrower: the unsupervised artifact as proof of learning. A take-home essay or problem set used to certify two things at once — that the student produced it, and that the student could. The first certification is gone, unverifiable, and not coming back. Every workable response starts from accepting that, because strategies aimed at restoring it — [detector regimes](/school/ai-and-academic-integrity/) with their documented false accusations, surveillance software, the arms race — spend enormous trust for a certainty they cannot actually deliver.

Redesign patterns that are working

Assess the process, not just the artifact. Proposals, annotated bibliographies, drafts with revision visible, reflection memos on what changed and why. Process is where learning is visible, it is dramatically harder to outsource wholesale — and it was always better pedagogy; AI merely made the upgrade urgent.
Add a live component. A five-minute conversation about the submitted essay — why this argument, what did you cut, defend this claim — verifies authorship better than any detector, and doubles as the most instructive five minutes of the assignment. Scaled versions: in-class writing sprints, presentations, oral spot-checks of a random sample.
Anchor in the particular. Assignments tied to *this* classroom's discussion, *this* lab's data, local observation, personal experience, or this week's material resist generic completion — a model can imitate an essay on a theme; it was not in the room.
Assign the AI itself. Have students generate an AI answer and then critique it — find the errors, the hallucinated sources, the shallow reasoning, and improve on it. This converts the forbidden tool into the object of analysis, builds [exactly the evaluation skill](/school/critical-thinking-with-ai/) they need for life, and quietly demonstrates that the median AI essay is a C.
Say what the assignment is for. Students outsource work whose purpose is opaque to them. 'This essay exists to train X, and doing it with AI trains nothing' is not naive — paired with a clear [per-assignment AI policy](/school/ai-and-academic-integrity/) spelling out what is allowed, it measurably shifts behavior, because most students are not looking to cheat; they are looking for the point.

The other half: AI as the teacher's assistant

The same capability that disrupted assessment is the best planning-and-differentiation tool teaching has ever had, and teachers deserve to use it as unapologetically as [any other professional](/work/everyday-ai-at-work/): generating problem sets and their variants, drafting rubrics and model answers, re-explaining a concept five ways for five students, adapting a reading's difficulty for different levels, producing quiz banks from your own materials, drafting the parent email and the recommendation-letter first pass. The professional norms transfer directly — verify content before it reaches students (AI errors in a worksheet are still your errors), keep student personal data out of consumer tools, and let AI draft while you decide. Hours reclaimed from production are hours available for the part of the job AI cannot do, which was always the actual job: attention to particular students.

The classroom that comes out of this

The uncomfortable, energizing conclusion most thoughtful educators reach: AI did not break assessment so much as call a bluff on assessment that was already weak — the unsupervised generic artifact was always a lossy proxy for learning. The redesigns that survive contact with AI (process visibility, live defense, particularity, critique) are the ones assessment research recommended before AI existed. And the students, who will graduate into [workplaces where these tools are simply ambient](/work/bringing-ai-to-work/), are better served by classrooms that teach discriminating use than by ones that simulate 2019. None of this makes the transition less work. It does make it work in a defensible direction.

Common questions

How do I make assignments AI-resistant?

Shift weight from the unsupervised artifact to things AI cannot fake: visible process (drafts, revisions, reflection memos), live components like brief oral defenses or in-class writing, and anchoring in the particular — this classroom's discussion, local data, personal observation. Resistance through surveillance and detectors is weaker than redesign, and detectors carry false-accusation costs.

Should teachers use AI detectors?

With heavy caution if at all: they are probability guesses with documented false positives that disproportionately flag careful conventional writing and non-native English speakers, and a wrong accusation costs more trust than most cheating does. A five-minute conversation about the work verifies authorship better than any detector, and teaches something while doing it.

How can teachers use AI for their own work?

As a production assistant: generating differentiated problem sets and readings, drafting rubrics, model answers, and quiz banks from your materials, re-explaining concepts multiple ways, and first-drafting routine communication. Verify everything before it reaches students and keep student personal data out of consumer tools — then spend the reclaimed hours on individual attention.

Should students be allowed to use AI in class at all?

A blanket ban prepares students for a world that no longer exists and pushes use underground where no skills are taught. The emerging consensus pairs clear per-assignment rules with assignments that use AI as an object of critique — generating, then dismantling, its answers — which builds the evaluation skills graduates will actually need.