What Should Students Learn for Careers Where AI Is Commonplace?

A student asks why they should learn to do something a chatbot already does. A program advisory group says employers want graduates who can use AI at work. Your syllabus has to satisfy both, and neither of them has said what the graduate should be able to do unaided.

Short answer to give a colleague: teach the whole professional task, not the tool operation, and assess both the supported product and the student's independent capability. The evidence supports treating readiness as disciplinary knowledge plus judgment plus verification plus transfer, and it does not support any list of skills as a guarantee of employment. A survey of 500 US college instructors against an employer panel found the two groups agreeing on the importance of exactly one of 26 AI-related skills, setting realistic expectations for AI-augmented work; instructors weighted critical and responsible use while employers weighted workflow, automation and human-AI teaming (Fried, 2026). Those are reported priorities, not hiring evidence. The sharper finding for a curriculum is elsewhere: in a survey of 455 final-year students, AI literacy reached perceived employability mostly by raising career adaptability, with an indirect effect of β = 0.173 alongside a direct β = 0.447 (Hendrawan and Lestari, 2026). Adaptability is a capability you can teach. A tool tip is not.

The short version

1. Write the capability as a claim about what a graduate can do unaided, then add the supported version beside it.

2. Teach the complete task in six parts: frame, know, delegate, verify, communicate, adapt.

3. Assess the product and the judgment separately, and make the independent check proportionate.

4. Keep independent practice where it is doing work, and say why in the syllabus.

5. Distinguish immediate outcomes from later ones when you evaluate the program.

6. Treat career anxiety as a real concern and answer it with investigation, not with predictions.

What employers and instructors each emphasize

Ithaka S+R's report asked 500 US four-year-college instructors to rate 26 skills from a framework spanning seven categories, and compared their ratings to a panel of employers surveyed by the framework's authors. The disagreement is systematic rather than random: instructors ranked the limits of AI, attribution and human accountability at the top, while employers ranked workflow design, automation and human-AI teaming higher.

Two readings of that gap are tempting and both are wrong. The first is that instructors are behind the labor market. The second is that employers want compliance. A more useful reading is that each group is describing a different half of the same job, and a graduate who can only do one half is underprepared for both. A student who cannot explain why a result is wrong is no use to an employer who trusts the workflow; a student who can only critique AI output is no use to an employer who needs the task completed.

The limits of this evidence matter. These are stated priorities from surveys, and the report does not show that any curriculum produces better hires, better performance or better pay. What it does give you is a defensible reason to ask whether your students ever practice completing and checking an AI-supported workflow, as opposed to discussing it.

Teach a complete professional task

The curriculum question worth writing down is: what would a capable beginner need to do, explain and verify in this situation? Then build the task in six parts.

Frame the problem. Identify the goal, audience, constraints, available evidence and the cost of being wrong. A student who cannot recognize an underspecified request will pass it to a model and accept whatever comes back.

Bring disciplinary knowledge. Students need enough understanding to inspect the result: interpret a case, estimate a plausible answer, notice a misleading comparison, spot what is missing. Define that baseline explicitly instead of treating all independent practice as a legacy requirement.

Delegate deliberately. Decide which steps AI may assist and which need independent work or professional authorization, and be able to give the reason. Drafting routine wording and deciding what a consequential record should say are different responsibilities with different accountabilities.

Verify against evidence. Check calculations, claims, sources, assumptions and compliance with the requirement. Teach students to record how a check resolved the question. Asking the same model whether it is sure resolves nothing, a point verification practice develops directly.

Communicate and collaborate. Explain the decision to a colleague, client or supervisor: what changed, what remains uncertain, when someone else must be involved. AI-supported work still has an audience and a chain of responsibility.

Adapt and transfer. Repeat a comparable task with a different tool or a new constraint, then ask what stayed useful and what had to change. This is what makes adaptability observable rather than an adjective in a graduate attribute.

Build it into existing course work

These are adaptations of the research themes, not tested employment interventions.

Composition and business communication: have students critique and revise an AI-drafted reply to a customer complaint. The evidence of capability is naming the unsupported promises, explaining the audience choices, and defending the final wording.

A quantitative course can do the same with an AI-produced estimate for a realistic planning problem, where students construct their own estimate, state the assumptions and test plausibility.

For a spreadsheet or computing course, hand out a generated formula and synthetic data, then ask for test cases, an explanation of the references, and one failure condition.

Health professions: students evaluate a generated explanation for a fictional case against approved sources, name its limitations, and state when a qualified professional must decide.

Career exploration has its own version: compare the task composition of two occupations using credible occupational information, then plan one concrete skill-development or help-seeking step.

For asynchronous delivery, ask for an initial attempt, a short verification record and a two-sentence note on one consequential decision. In a lab or seminar, the explanation can happen in conversation. Use synthetic or authorized materials whenever the alternative is confidential workplace or student data.

Assess supported performance and individual understanding

A polished artifact tells you what was produced under the permitted conditions. It does not tell you what the student can do. Add a proportionate independent check: explain a decision, repair an error, adapt the work to a changed condition, or perform one foundational subtask unaided.

Choose those requirements because they support a stated capability, and say which one each assessment is measuring. Students accept independent practice when it is connected to the thing they are being prepared for; "you will need this someday" does less work than showing how the calculation is what lets them catch a wrong estimate later.

The support side has its own risk. Shojaei et al. (2026) found GenAI use positively associated with GenAI dependence, and dependence negatively associated with both critical thinking and employability in their survey model. That is an association rather than a causal claim, and it points at the design question: a course that only ever assesses supported performance may be training the dependence it reports.

For program review, separate the measures. Course performance, demonstrated AI literacy, internship-supervisor evaluation, employment status and job quality are different outcomes with different time horizons, and a single intervention cannot establish all of them. The AI literacy evidence page is the honest inventory of what intervention studies actually support.

Address career anxiety without promising certainty

Students arrive with real fears, and the survey literature on them is a description of the fears rather than a forecast of the labor market. Duan et al. (2026) found AI anxiety directly and negatively predicting career decisions, which is a finding about decision paralysis, not about jobs. The wellbeing picture covers the wider evidence.

What helps is investigation at the scale a student can act on: which tasks does this role actually contain, what does a week look like, what would I need to be able to do, who can tell me. Career services, alumni and placement staff hold that information. A tool demonstration, an instructor's impression or a vendor's presentation is not a labor-market forecast, and presenting one as a forecast is how programs lose their credibility with students who then check.

What will come back at you

"Students just need better prompting." Prompting is one clause of one of the six parts. A well-specified request does not supply the knowledge needed to recognize a bad result, and the employers in the Ithaka panel were not ranking prompt craft first either.

"If workplaces allow AI, coursework should allow it." A course also builds the capability that makes responsible use possible. Say which of the two you are assessing in each task, and the objection becomes a design question rather than a contradiction.

"The tools change too fast to teach any of this." Teach tool-specific operations where they help and assess the durable practices: framing, checking, testing, documenting, communicating, escalating. The evidence above is about those practices, and none of it depends on which model is current.

"An AI-literacy course will improve their job prospects." The mediation result is the caution here. AI literacy reached perceived employability directly and through adaptability, and perceived employability is not a job offer, a performance rating or a salary. The survey design cannot rule out that confident students rate themselves higher on everything.

What to do this week

1. Choose one task graduates in your program will actually face, and write the six parts for it in five lines.

2. Mark which parts the current assessment actually measures and which it assumes.

3. Add one independent check to an existing assignment: a decision explained, an error repaired, or a subtask done unaided.

4. Tell students which capability that check is for, in the assignment instructions.

5. Check whether any task in the course assesses only supported performance, and decide whether that is deliberate.

Where to go next

Career Development and Readiness is the concept page for employability, adaptability and workforce evidence, and AI Literacy covers what literacy means beyond tool operation. For building it into a specific course, see incorporating AI literacy and what competencies faculty need.

For the assessment design that makes supported and independent performance distinguishable, see redesigning assessment in the AI era and AI grading. For how to keep support from becoming dependence, see reducing over-reliance.