Teaching students to work with AI

Should you let students use ChatGPT on this assignment?

The question is usually framed as permission. It’s more useful as a design question: what is this assignment actually measuring, and does a chatbot measure it too?

Most instructors asking whether to allow ChatGPT on an assignment are really asking something narrower - will allowing it destroy what I’m trying to assess? That question has an answer, and it depends far less on your feelings about AI than on the structure of the assignment in front of you.

Here is a way to decide in about two minutes.

The two-minute test

Put your assignment prompt into a chatbot exactly as your students would receive it. Read what comes back. Then ask three questions.

So: yes, no, or it depends

Allow it when the AI does the part that isn’t the point

If you’re teaching market analysis and the assignment involves building a spreadsheet model, the formula syntax isn’t the learning objective - the judgment about which assumptions to make is. Let students use AI for the syntax. Assess the assumptions.

Same logic for code scaffolding in a course about system design, for literature search in a course about synthesis, for prose polish in a course about argument.

Restrict it when the tool replaces the skill you’re building

Foundational courses are different. If a student never struggles through writing a paragraph, they don’t develop the ear that later lets them tell a good paragraph from a plausible one. There’s a real case for restriction in first-year composition, in introductory programming, in any course whose job is to build the judgment that later lets someone supervise the tool.

But restriction only works if you can observe the work. Which brings us to the uncomfortable part.

If you want to restrict, move the assessment somewhere you can see it: in class, on paper, in a short oral defence, in an activity that records the process rather than collecting a finished file. Otherwise you’re grading two populations against one standard and telling yourself otherwise.

Require it when the skill is working with the tool

This is the case most instructors skip, and it’s the one your students will most need. Working well with AI is a learnable skill with identifiable failure modes: accepting a confident wrong answer, asking a question so vague the output is useless, failing to notice a missing assumption, never pushing back.

Those are all things you can assign and grade. Teaching students to work with AI covers the capabilities in detail.

The middle ground most courses land on

In practice most instructors end up somewhere specific rather than at a blanket yes or no:

• Open on drafting, closed on final prose. Brainstorm, outline, argue with it. Write the submitted words yourself.
• Open with disclosure. Use what you like, tell me what you used and what you changed. See policies you can copy.
• Open on some assignments, closed on others, labelled per assignment so nobody has to guess.

All three are defensible. What isn’t defensible is leaving it unstated, because students will resolve the ambiguity in whichever direction suits them and you’ll have no basis for objecting.

A better version of the question

“Can students use ChatGPT on this assignment?” assumes the assignment is fixed and the only variable is permission. Flip it:

That question has better answers. It leads to assignments where AI is present and the student still has to think - defend a position to someone who pushes back, decide what to do with incomplete information, catch the error the model introduced, explain why the confident answer is wrong.

Those assignments are harder to write from scratch, which is most of why people don’t write them. They’re also the ones that will still be worth assigning in three years.

Common questions

Won’t allowing AI mean students learn less?

It depends entirely on what you then assess. If you allow AI and keep grading the artifact, yes - students will submit better artifacts having learned less. If you allow AI and grade the reasoning, the decisions, and the ability to catch the model’s mistakes, they’ll learn more, because those are harder skills than the ones the artifact was testing.

What about fairness - some students are better at prompting than others?

True, and that’s an argument for teaching it rather than ignoring it. Uneven prior access is a reason to build the skill explicitly in class, the same way you’d handle uneven prior exposure to academic writing.

Is it cheating if the syllabus doesn’t say?

From a student’s point of view, an unstated rule isn’t a rule. From an integrity-office point of view, an unstated rule is very hard to enforce. State it.

How do I know if they used it when I said not to?

Mostly you can’t, and detectors won’t save you - here’s why. Design the assessment so it doesn’t depend on knowing.