Prompting is becoming a basic skill. Knowing when AI is wrong is becoming rare.
Two final-year students get the same assignment: analyze why a regional retail chain’s online sales are falling, and recommend what to do. Both use AI. Both submit polished, confident reports within an hour.
The first student’s report is excellent. The second student’s is equally polished, and wrong. The AI misread the question, assumed the chain sold mostly to urban customers, and built an entire recommendation on that assumption. The student didn’t notice, because nothing in the report looked wrong.
Both students can prompt. Only one can judge. In a few years, that difference will decide who gets hired, promoted, and trusted with decisions.
Prompting is becoming table stakes
Two years ago, “prompt engineering” sounded like a career. Today, most students can get useful output from a chatbot, and AI tools increasingly write and refine prompts themselves. Asking well still matters, but it’s quickly becoming a baseline skill, like typing or searching the web.
What doesn’t get commoditized is the ability to look at a fluent, well-formatted answer and decide whether it’s right, good enough, or dangerously plausible.
The assumption that broke: Knowing how to use AI is the skill that matters.
The jagged frontier
The best evidence for why judgment matters comes from a field experiment with 758 consultants at Boston Consulting Group. On tasks within AI’s capabilities, consultants using GPT-4 completed 12.2% more tasks, 25.1% faster, at more than 40% higher quality. But on a task deliberately chosen to sit just outside what AI could do well, consultants using AI were 19 percentage points less likely to reach the correct answer than those working without it.
The researchers called this boundary the “jagged frontier”: tasks that look similar can fall on very different sides of what AI handles well. The study used GPT-4 on bounded exercises in 2023, so the specific numbers won’t hold forever, but the management problem remains: past success with an assistant can increase the chance that a person relies on it in a nearby task where it is weak.
The consultants weren’t careless. They were experienced professionals, and the AI’s answer looked just as convincing when it was wrong as when it was right. The only protection is knowing enough about the domain to notice when the problem has changed.
That’s the uncomfortable conclusion for education: you can’t judge what you don’t understand. Judgment isn’t a replacement for knowledge. It’s what knowledge makes possible.
How judgment gets built
You can teach judgment, but not through lectures on AI ethics or a module on prompt techniques. It’s built through practice with feedback, in four habits.
Hunt for the error. Give students AI output that contains a flaw, such as a wrong assumption, a fabricated citation, or a subtle logic error, and grade them on finding it. This reverses the usual relationship: the AI produces, the student evaluates.
Compare and defend. Have students generate three different AI answers to the same problem, choose the best, and explain why the others fall short. Choosing well is the professional skill; generating is the easy part.
Trace it to the source. Check every important claim in AI-assisted work against original evidence: a dataset, a textbook, a standard, or a real document. Students learn quickly how often confident AI claims have nothing behind them.
Know when not to use AI. Some tasks sit outside the frontier. Students should practice recognizing them by predicting where AI will struggle before they try it, and then checking whether they were right.
The scale problem
Judgment is learned through feedback on judgment calls: an expert saying “you trusted this, and you shouldn’t have, and here’s why.” That’s one of the most time-intensive forms of teaching.
A teacher with 120 students can’t design fresh flawed examples every week, read 120 critiques, and give each student detailed feedback on their reasoning. A Training and Placement Officer (TPO) preparing 600 students can’t personally probe each one’s judgment in a mock interview.
This is where AI can help teach the skill it makes necessary. It can generate realistic flawed outputs matched to each syllabus topic, evaluate students’ critiques against a rubric, challenge their reasoning through Socratic dialogue, and show teachers which students consistently over-trust AI. Mentors then spend their scarce time on the hardest calls and the students who most need a human conversation.
AI for everyone’s practice, and human judgment where it matters most.
What each stakeholder can do now, and why it’s hard
Students
Do now: Keep a short verification log for every AI-assisted assignment: what the AI claimed, what you checked, and what you changed. Once a week, deliberately ask AI something in your subject that you already know the answer to, and see whether it gets it right.
Why it’s hard: AI answers look equally confident whether they’re right or wrong, and your current grades reward almost none of the time you spend checking.
Teachers
Do now: Add one “critique the AI” question to every internal test: give students a flawed AI answer and ask them to find and explain the problem. Grade the reasoning, not just the conclusion.
Why it’s hard: Designing good flawed examples takes time, and marking reasoning for 120 students is far slower than marking answers.
Training and Placement Officers (TPOs)
Do now: Add a judgment round to mock interviews: show students AI-generated code, analysis, or a draft, and ask what’s wrong with it and what they would do next.
Why it’s hard: Mock interview capacity is already stretched, and few question banks test judgment rather than recall.
Institution leaders
Do now: Make critical evaluation of AI output an explicit learning outcome in every program, mapped to existing outcome frameworks such as problem analysis and professional ethics.
Why it’s hard: Current syllabi and accreditation language don’t name it yet, so it falls between departments, and nobody owns it.
Industry leaders
Do now: Replace one recall question in your interviews with a review task: give candidates AI-generated work with a planted flaw, and assess whether they find it and how they reason about it.
Why it’s hard: Standard interview banks are built for recall and speed, and reviewing reasoning takes more interviewer time.
What to look for in a solution
The goal is to build judgment deliberately, for every student, not just the few who get expert feedback. The right platform should offer:
Assessments that test evaluation, not recall: questions mapped to the higher levels of Bloom’s taxonomy, including tasks where students must analyse, critique, and correct AI-generated work.
Job-grounded judgment tasks: review-and-critique exercises built from real job descriptions, so students practice the calls they’ll actually face at work.
An AI that questions and argues back: a Socratic tutor that challenges a student’s reasoning instead of confirming it, and a debate partner that pushes back on weak arguments.
Mock interviews that probe reasoning: AI interviewers that ask “why?” and “what would you check?” with instant feedback, available to every student.
Evaluation at scale: rubric-based assessment of critiques and explanations, with an AI first pass, so teachers can grade reasoning for 120 students without losing their weekends.
Signals on over-trust: analytics showing which students consistently accept flawed AI output, so mentors know who needs a conversation.
Skills verified, not claimed: judgment demonstrated in controlled assessments before it appears on a certificate or profile.
In short, look for an employability operating system that trains students to judge AI, not just to use it.
The bottom line: Soon everyone will be able to get a fluent answer from AI. What will separate graduates is the ability to tell a good answer from a convincing wrong one, and that ability rests on knowledge, practice, and feedback. Teach students to prompt, and they’ll keep up. Teach them to judge, and they’ll lead.
What’s the most convincing wrong answer you’ve seen AI give in your field?
See how a connected learning-to-employment ecosystem works → http://7seers.ai
