Teaching students to work with AI

What AI literacy actually means

“AI literacy” gets used to mean everything from knowing what a transformer is to being able to write a decent prompt. Here’s a narrower definition that you can actually teach and assess in a normal course.

Most AI literacy frameworks fail in one of two directions. Either they’re technical - how models are trained, what a parameter is - which is interesting and almost never load-bearing for a student’s actual work. Or they’re so broad as to be untestable: “understands the ethical implications of AI.”

The useful definition sits in between, and it’s behavioural rather than conceptual.

That framing has a property the others lack: every part of it is observable. You can watch a student do it, or fail to.

What it breaks into

1. Knowing what the tool is bad at

Not in general - in your discipline. A history student should know models confabulate citations. A statistics student should know they fumble base rates. A nursing student should know the dosage arithmetic needs checking every time. Generic warnings don’t transfer; discipline-specific failure modes do.

This is the single highest-value thing you can teach, and it takes one well-designed assignment.

2. Asking a question that deserves a good answer

Most poor output is a response to a poor question. Students routinely paste an assignment prompt and accept whatever comes back. The skill is supplying context, constraints, and a sense of what a good answer would contain - which requires understanding the problem first, and is therefore not a workaround for understanding it.

3. Verifying before relying

The habit of checking one claim before building on it. This is the skill most likely to matter in a student’s first job and the one least likely to develop on its own, because the output looks finished. Checking has to be assigned before it becomes a habit.

4. Pushing back

Treating the first answer as a draft rather than a verdict. A literate user challenges a claim, asks what would make it wrong, notices an unstated assumption. An illiterate one accepts and reformats.

5. Deciding what to keep

Taking the useful part, discarding the rest, and being able to say why. Wholesale adoption and wholesale rejection are both failures of the same judgment.

6. Saying what you did

Accounting honestly for the tool’s role in your work. This is partly integrity and partly professional norm - in most workplaces your colleagues need to know which parts of a document were machine-drafted, because it changes how carefully they read it.

Teaching it without adding a course

Nobody has a spare three weeks. The good news is that all six capabilities can be built inside assignments you were already going to set, by changing what gets assessed rather than what gets covered.

• Failure modes: one assignment where the model predictably gets it wrong in your discipline, and the student’s job is to find out how.
• Question quality: ask for the prompt alongside the work, and grade it.
• Verification: require one checked claim with the source, every time. It becomes automatic in about three assignments.
• Pushing back: any assignment with an AI character who argues a position.
• Judgment about what to keep: a three-sentence disclosure note on what was changed and why.
• Honest accounting: the same note, plus a policy that makes it safe to be truthful. Policy language here.

Twelve assignment shapes covers the mechanics in more detail.

Why this version and not the broader one

There’s a reasonable objection: this definition leaves out the societal questions - labour displacement, bias, environmental cost, concentration of power. Those matter, and in many courses they belong.

But they’re a different subject. A student can hold sophisticated views on algorithmic bias and still accept a fabricated citation because the paragraph read well. Conversely, a student who has learned to verify, push back, and account honestly has the habits that make the broader critique actionable rather than theoretical.

Teach the judgment first. It’s the part that doesn’t expire when the models change, and it’s the part employers are currently finding absent.

Common questions

Isn’t this just critical thinking with extra steps?

Largely, yes - and that’s the point. The capabilities aren’t new. What’s new is that a tool now produces fluent, confident, plausible output at zero cost, which means the gap between students who apply judgment and students who don’t is no longer visible in the finished work. The skills are old. The need to assess them directly is new.

At what level should this be taught?

All of them, with different emphasis. First-year students need failure modes and verification. Advanced students need judgment about what to keep and honest accounting, because they’re closer to contexts where the stakes are real.

Do students think they already have this?

Usually. Fluency with the interface reads as competence with the tool, and they’re different things. The fastest way to demonstrate the gap is a seeded-error assignment - most students are genuinely surprised by how confidently wrong the output can be.

How do I assess it?

A six-criterion rubric mapped to the capabilities above, weighted at 15-30% of the assignment. The rubric, with what each level looks like.