AI in the Classroom: Start With the Assignment, Not the Tool

AI in the Classroom: Start With the Assignment, Not the Tool

Every campus conversation about AI eventually collapses into the same two camps: the enthusiasts who want a tool in every syllabus and the skeptics who want a ban in every catalog. Both camps share an assumption — that the technology is the decision. It isn’t. The assignment is the decision. The tool is a consequence.

The question that reframes everything

Here is the exercise I run with faculty, and it takes ten minutes. Pick one assignment from your course. Now answer honestly: what is the thinking this assignment exists to make students do? Not the deliverable — the thinking. The essay is not the point; the argument construction is. The problem set is not the point; the modeling judgment is.

Now ask the second question: if a student used AI on this assignment, would the thinking still happen?

Three answers are possible, and each one tells you exactly what to do.

The thinking still happens. Some assignments survive AI untouched because the work is in the doing: a lab performed, a patient interviewed, a weld inspected, a presentation delivered live. If AI can only help around the edges — cleaning up a summary, suggesting structure — then the tool is a typing upgrade, and your policy can say so in one sentence.

The thinking gets skipped. This is the uncomfortable middle where most writing-adjacent assignments live. If the essay was the only evidence that the reasoning occurred, and AI can produce the essay, the assignment now measures access to a chatbot. The answer is not detection software — it is redesign: move the reasoning somewhere visible. Drafts with tracked decisions. In-class defense of the argument. A process log alongside the product. Oral follow-ups on a random sample. None of this is new pedagogy; it is what writing instructors always did when they could afford the time. AI just made the time cost unavoidable.

The thinking changes. The most interesting category: assignments where AI in the loop creates a harder, better task. Critique the model’s answer and find its errors. Get three AI drafts, then write the version all three missed. Use the tool to generate the counterargument you must now defeat. Here the syllabus line isn’t “permitted” — it’s “required, and here’s the skill we’re actually grading.”

Policy that fits on an index card

Once you’ve sorted your assignments into those three bins, the AI policy writes itself, and it fits in 150 words: which uses are open, which require disclosure, which are closed — per assignment, not per course. Blanket bans read as unenforceable theater, and students treat them accordingly. Assignment-level clarity reads as design, and in my experience students respect design even when they test it.

Disclosure is the piece most policies miss. Normalizing “state what you used and how” does two things: it converts a cheating conversation into a craft conversation, and it generates the data you need to see how the tool is actually being used — which is never quite how you guessed.

What this asks of faculty (and what it doesn’t)

The honest cost of AI-resilient teaching is assessment redesign, and redesign is labor. It is fair to say so out loud, and fair to ask institutions to fund the time. But the scale of the redesign is smaller than the panic suggests. You do not need to rebuild the course. You need to find the assignments in bin two — where the thinking gets skipped — and move the evidence of thinking somewhere AI can’t fake it. For most courses that is two or three assignments, not twenty.

What faculty should refuse to do is outsource the problem to surveillance. Detection tools produce false accusations at rates no disciplinary process should tolerate, and an arms race with your own students corrodes the thing the classroom runs on. Design beats detection, every time, at lower cost in trust.

The skill under the skill

One more reframe, because it changes the stakes. When employers say they want graduates who “know AI,” they do not mean prompt tricks. They mean judgment: knowing what to delegate to a model, what to verify, and what to never hand over. That judgment is teachable, and the classroom that requires disclosed, deliberate AI use is teaching it. The classroom that bans everything is teaching students to use it in secret, which is the one fluency nobody wants.

Start with the assignment. Name the thinking. Protect it where it’s threatened, extend it where the tool makes more possible, and write the policy last. The technology will change again by spring; the method survives every model release.