Imagine a final-year public-health student advising a city clinic on whether to adopt a new appointment system. An AI-generated briefing confidently recommends a rollout, citing a pilot that supposedly reduced missed visits. Buried in the source packet, however, the pilot involved a different patient group, its comparison period was unusually short, and the briefing treats an association as proof of causation. The prose is immaculate. The recommendation is shaky. Would the student spot the problems before a real employer acted on the advice?
That is a question universities can test. When a polished answer is easy to produce, grading the answer alone tells a lecturer too little about the student's judgment. A June 2026 assessment report from Australia's Tertiary Education Quality and Standards Agency emphasizes evaluative judgment, critical thinking, and ethical reasoning as capabilities students need in AI-integrated workplaces. Those capabilities deserve evidence in the assessment itself.
Research offers a reason to take the problem seriously, though it should be interpreted carefully. In a 2026 Organization Science study of 758 management consultants, AI assistance helped on tasks that suited the technology and worsened performance on a task outside its capabilities. The participants were working professionals, and the experiment did not test university teaching. Still, it illustrates a practical danger: a person can receive a fluent recommendation and miss the moment when independent scrutiny becomes necessary.
Give Students an Answer Worth Questioning
I suggest a three-stage assessment that a capstone instructor or professional-placement supervisor could pilot within an existing course. It permits AI use while making the student's own reasoning visible. This is a proposed teaching method, not an intervention with demonstrated learning benefits.
First, give students a realistic disciplinary problem before showing them any AI response. In ten minutes, have them write an initial recommendation, the assumptions behind it, two facts they would need to verify, and one finding that would change their mind. The point is to capture a baseline for reasoning, rather than reward speed or perfect recall. Students can consult the ordinary materials specified for the exercise, but the initial response should reflect their own judgment.
Next, provide a prepared AI-generated answer alongside a small, instructor-curated source packet. Include sound evidence, at least one source whose relevance is questionable, and a claim the packet cannot establish. Ask students to annotate the AI answer sentence by sentence: supported, contradicted, or unresolved. For each contested claim, they must point to the particular passage or dataset, explain how it bears on the decision, and identify what additional evidence would settle the issue.
Return to the clinic example. A strong student notices that the cited pilot excluded patients whom the clinic principally serves. Another finds that an apparent improvement may reflect changes in the appointment population. A weaker response merely calls the AI answer “biased” or “unreliable” without demonstrating why. Such distinctions matter more than whether a student can recognize a spectacularly invented citation.
Finally, ask for a short revised recommendation and a two-minute explanation, or an equivalent written defense when that better meets accessibility needs. Students should show which conclusions survived scrutiny, which changed, what remains uncertain, and the point at which they would seek a qualified supervisor's advice in actual practice. The exercise thus tests both the decision and the student's ability to explain its limits.
Grade What the Student Actually Demonstrates
A simple rubric can score four behaviors on a 0–2 scale: source checking, uncertainty, revision, and escalation. Give zero for an absent response, one for an assertion without a defensible explanation, and two for specific reasoning tied to the supplied material. A student earns the highest source-checking score by matching a consequential claim to evidence and identifying a mismatch. For revision, the student must explain why new evidence changed, or reasonably failed to change, the initial recommendation.
This rubric measures visible performance in one exercise. It should not be treated as a diagnosis of someone's general intelligence or a measure of long-term learning. An instructor can compare the rubric results with other course work and discuss the reasons for disagreements between markers. An earlier TEQSA assessment-reform paper likewise treats assessment design as a set of context-dependent trade-offs, rather than a universal formula.
Universities can adapt the method without pretending every discipline faces the same mistakes. An engineering class might examine an AI specification that omits a safety constraint; a history seminar might inspect an argument that confuses primary and secondary evidence; a business capstone could challenge a market projection based on the wrong customer group. TEQSA's 2025 examples of assessment reform underline the value of putting assessment principles into practical course settings.
There is also a developmental benefit worth testing, rather than assuming. UNESCO's AI competency framework for students includes critical judgment of AI solutions, while its teacher framework calls for pedagogies that preserve students' agency and independent thinking. Instructors could pilot this exercise in one course, refine ambiguous source packets, and observe which reasoning steps students struggle to explain. They should report what the pilot reveals, without claiming improved graduation outcomes on the basis of a classroom exercise.
A university cannot promise to produce graduates who never trust a bad AI answer. It can, however, ask them to demonstrate how they recognize one, check the evidence, correct their judgment, and know when the stakes call for help. That is a much more useful graduation test than another polished answer.
About the Author
Gleb Tsipursky, PhD, a behavioral scientist, CEO of Disaster Avoidance Experts, and author of The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026) https://disasteravoidanceexperts.com/aibook. He taught at Ohio State University and UNC-Chapel Hill for fifteen years and currently advises organizations on practical AI adoption across industries.