Teaching students to work with AI

AI readiness for students

Employers say graduates aren’t AI-ready. They rarely say what that means. Here’s the gap they’re actually describing, and what a course can do about it in one term.

“AI readiness” is doing a lot of work in a lot of strategic plans right now, usually without a definition attached. When you press employers on what they mean, the complaint is surprisingly consistent, and it isn’t about tool familiarity.

New hires are fluent with the interface. They can prompt, they can iterate, they produce output fast. What they don’t do is notice when the output is wrong, push back when it’s thin, or take responsibility for what they pass along.

What employers are actually reporting

Broken down, the complaint has four parts:

• Passing along unverified work. A junior hire sends a client-facing document containing a figure nobody checked. The figure was confident, formatted, and wrong. Nobody along the chain treated it as a draft.
• Accepting the first answer. Treating output as a verdict rather than a starting point. No follow-up, no “what would make this wrong,” no noticing the unstated assumption.
• Not knowing what the tool is bad at in this domain. Generic awareness that “AI can hallucinate” doesn’t help someone spot a plausible but invented case citation, or a dosage calculation off by a factor of ten, or a cost that should scale non-linearly.
• Unclear accounting. Colleagues need to know which parts of a document were machine-drafted, because it changes how carefully they read it. New hires often don’t volunteer this, sometimes because nobody told them it mattered.

Every item on that list is a judgment habit, not a technical skill. Which is good news, because judgment habits are what a course is for.

Three things AI readiness is not

It’s not tool training

Students learn the interface in an afternoon and the specific tool will change twice before they graduate. A workshop on “prompt engineering best practices” teaches techniques that are already half-obsolete and none of the underlying judgment.

It’s not banning AI so they learn the fundamentals

There’s a real case for protecting foundational skills, and it applies to specific courses at specific points. But a student who reaches graduation having been told only that AI is prohibited has been prepared for a workplace that doesn’t exist. The ban, where it’s justified, has to sit inside a curriculum that also teaches the tool.

It’s not a standalone module

A one-credit AI literacy course produces students who can discuss AI readiness. It doesn’t produce students who verify a number under deadline pressure, because the habit only forms where the stakes of the discipline are present. It has to happen inside the major.

What a single course can do

You don’t need a curriculum overhaul. Four changes, each attachable to assignments you already set:

1. Teach your discipline’s failure modes, once, concretely

One assignment where the model predictably fails at something specific to your field - the base-rate problem, the invented citation, the cost that doesn’t scale linearly, the edge case in the code. Students are usually genuinely surprised, and surprise is what makes the lesson stick.

2. Require one verified claim, every time

Every submission includes one claim the student checked, with the source and what they found. It takes them five minutes and it becomes automatic in about three assignments. This is the single highest-return change on the list.

3. Make them defend something to someone who pushes back

Fluency collapses under a good follow-up question. An exchange where the student has to hold a position against a sceptical counterpart - and concede where they should - builds the habit that reading alone doesn’t.

4. Ask for honest accounting

Three sentences: what tool, what you asked for, what you changed. It’s a professional norm before it’s an integrity mechanism, and framing it that way gets far better compliance. Policy language for this.

Talking to employers and accreditors about it

If you’re writing this into a program review or an advisory board conversation, the defensible version is narrow and observable. Graduates of this program can:

• name the specific failure modes of AI in this discipline;
• verify a claim before building on it;
• challenge a confident answer and change position on evidence;
• state clearly what a tool contributed to their work.

Those are assessable. “Students will be prepared for an AI-enabled workplace” is not, and an accreditor will ask you how you measured it.

Common questions

Is this different from AI literacy?

Mostly framing. AI literacy is the capability; AI readiness is the same capability described from the employer’s side. The capability breakdown is the more useful version for course design.

Which courses should carry it?

Ideally the ones where students already do applied work - capstones, labs, case courses, practicums. Those have the domain stakes that make verification feel necessary rather than performative.

Won’t this be obsolete when the models get better?

The specific failure modes will shift. The habit of checking before relying won’t, because the models will still be confident and still occasionally wrong, and the person passing the work along will still be responsible for it.

How do we measure it across a program?

Pick one artifact per course - a verified claim, a disclosure note, a transcript of a defended position - and score it on a common three-point scale. It aggregates, and it’s cheap enough that faculty will actually do it.