Teaching students to work with AI

12 AI assignment ideas

Twelve assignment shapes that assume students have AI and still require them to think. Each one is described concretely enough to adapt to your discipline this week.

The common feature is that AI is present and insufficient. The student can use it, will use it, and still has to do something the model can’t do for them - decide under uncertainty, defend a position against pushback, or catch the thing the model got confidently wrong.

Assignments about argument and judgment

1. The seeded error

Give students a scenario where the model reliably goes wrong - a base-rate problem, a case with a buried conflict of interest, a calculation with a unit trap. Ask them to produce an answer with AI, then identify where the model went wrong and why. Grade the diagnosis, not the answer.

Works in: statistics, ethics, finance, law, health sciences.

2. Defend it to someone who disagrees

The student takes a position and then argues it against an AI character briefed to push back hard from a specific viewpoint - a sceptical CFO, an opposing counsel, a review board. Grade whether the position survived, and whether the student updated when they should have.

Works in: business, policy, law, any course with contested questions.

3. The steelman exchange

Ask AI for the strongest case against the position the student holds. Then have them write a response to the best version of that argument, not the weak version they’d have constructed alone. The model is genuinely useful here, and the work is still theirs.

4. Where the consensus is wrong

Students ask a model for the standard account of a topic, then use course sources to identify where that account is oversimplified or outdated. This one directly rewards knowing the material better than the average of the internet.

Applied and quantitative work

5. Build the model, defend the assumptions

Students use AI to build a spreadsheet model - runway, break-even, dosage, load. The formulas can come from anywhere. The assessment is on the assumptions: why this discount rate, why this growth figure, what breaks if you’re wrong.

Works in: finance, operations, engineering, public health.

6. The code review

Give students AI-generated code that works but is wrong in a way that matters - a subtle off-by-one, an unhandled edge case, a security hole, a structure that won’t scale. They find it, explain it, and fix it. This is what most professional programming with AI now looks like.

7. The data interrogation

Give a dataset and an AI assistant that will run any analysis asked of it. Students have to decide what to ask. Grade the questions, not the charts - a student who asks whether the sample is representative before asking for a trend line is demonstrating exactly the thing you want.

8. Build the tool you need

Students describe a small interactive tool for a specific purpose - a calculator, a decision aid, a visualiser - and have AI build it. Grade whether the tool actually serves the purpose, and what they changed after seeing the first version.

Assignments that capture process

9. The annotated transcript

Students submit their work plus one exchange they found useful and one where the model was wrong, with a sentence on each. Minimal extra grading, and it surfaces reasoning the artifact hides. See how to grade this.

10. The staged assignment

Break the work into stages with a checkpoint between each. The student can’t generate the whole thing in one pass because stage three depends on feedback from stage two. This also happens to be good pedagogy independent of AI.

11. Group deliberation, then individual position

Each student takes a private position with a reason. Positions are revealed to the group. The group discusses, then each student writes a final position that must address the strongest objection raised. AI can help anyone prepare; the exchange itself is between people.

More on the structure: discussion and deliberation formats.

12. Teach it back

The student uses AI to learn something, then explains it to an AI character playing a confused novice who asks the obvious follow-up questions. Grade the explanation. This surfaces shallow understanding fast, because the novice keeps asking why.

What to stop assigning

Three formats are now close to unassessable as take-home work:

• Summarise this reading. Models do this well, and the student learns nothing you can verify.
• Define these terms. Never taught much anyway; now it teaches nothing.
• The general-topic essay. “Discuss the causes of X” with no specific case, no course framing, and no position to defend is the single most automatable thing in higher education.

Each can be rescued by adding specificity - your case, your dataset, your week-six framework, a position the student has to hold under questioning.

Common questions

How long do these take to write?

Longer than a prompt, shorter than you’d think. The seeded error takes the most effort because you need to find where the model reliably fails in your domain - budget an hour the first time, much less after that.

Do these work in large sections?

Numbers 1, 5, 6, 9 and 12 scale well because the assessment target is specific enough to grade quickly or automatically. Numbers 2 and 11 need either small sections or a platform that runs the exchange for you.

What if students use AI to do the meta-work too - the reflection, the error-finding?

They’ll try. It works poorly, because the model doesn’t know which of its own outputs you saw or what your course emphasised. Ask for specifics tied to your materials and the shortcut stops paying.