判斷力課堂

Chapter 1 The Wrong First Question About AI in Higher Education

The first question most universities ask about generative AI is some version of this: can students use it? It is an understandable question. It is also the wrong one.

I have now heard the question in enough rooms to know its usual choreography. Someone from administration asks whether a university-wide rule should permit, limit, or prohibit AI use. One faculty member worries about cheating. Another worries about being left behind. Someone else raises equity: not every student has the same tools, the same paid tier, the same device, or the same tolerance for technical friction. The conversation feels serious because everyone is anxious. Then the room does what institutions often do under pressure. It compresses a design problem into a permissions problem.

That is the beginning of the mistake.

The Familiar Meeting

Imagine a typical faculty meeting in the second or third month after a new AI release gains traction. The room is full of sensible people. A dean wants clarity. A committee chair wants consistency. A younger faculty member wants the institution to be less performative and more realistic. A senior faculty member wants to know whether students are still writing anything on their own.

The draft policy on the screen says something like this: “Students may use generative AI only with instructor permission and proper acknowledgment.” Everyone nods because the sentence sounds balanced. It is balanced in the way a chair with three legs is balanced.

The sentence does not answer the questions that actually govern learning. May students use AI before they have read the assigned case? May they use it to summarize the case they were supposed to read? May they use it to generate counterarguments after they have produced a first position of their own? May they use it to clean up grammar after the reasoning work has already been done? May they use it to build a role-play scenario that will later be debated in class? Each of these uses sits at a different point in the learning process. Each creates a different risk. Each may be educationally sound or unsound depending on the task.

Yet the meeting wants one answer. That desire is bureaucratically understandable and pedagogically disastrous.

Why the Permission Question Fails

The permission question fails for three reasons.

First, it treats AI as one thing. It is not one thing. It is a bundle of capabilities: summarizing, drafting, questioning, role-playing, classifying, translating, searching, comparing, visualizing, coding, and stylistically imitating. A student who asks an AI system to produce a polished final memo before thinking has not done the same thing as a student who asks the same system to simulate a hostile regulator after the student has already written a recommendation. Both are “using AI.” Only a lazy institution would stop there.

Second, the permission question treats the assignment as fixed. It assumes the task is already well designed and that the only matter left is whether a tool should be allowed near it. But AI has changed the meaning of weak tasks. A vague essay prompt once produced mediocre student prose. Now it produces mediocre machine prose with better punctuation. The tool did not create the weakness. It exposed it.

Third, the permission question encourages institutions to speak morally where they should speak operationally. They tell students to “use AI responsibly” and faculty to “maintain academic integrity.” Fine. Responsible how? Integrity with respect to which steps of the process? What evidence should remain visible? What part of the work is still expected to remain stubbornly human? A moral posture without workflow design quickly degenerates into ritual scolding.

Good teaching has never been merely a matter of whether a student arrives at a product. It has always depended on sequence, struggle, feedback, revision, and visible improvement. AI changes none of that. It simply makes bad sequencing faster, prettier, and easier to hide.

The Better Question

The better question is this: where does AI belong in the work of learning?

That is a different question because it asks about placement rather than permission. Once placement becomes the issue, the teacher can begin designing. What must happen before AI enters the task? What may AI do once it is allowed in? What evidence of process must survive? What should the student still have to defend without assistance?

That shift immediately produces better course questions:

Moving from permission to placement

The teacher who asks the better question is already doing something important. The teacher is defining the task as a sequence of intellectual acts rather than as a finished object. That matters because AI is strongest when the task is treated as a finished object. It is much weaker when the task is treated as a chain of reasoned moves.

A Tale of Two Students

The difference becomes clearer if we compare two students in the same class.

Both students are assigned a strategic case about entering a foreign market. Both are told to submit a recommendation memo. Student A opens a model immediately and types: “What is the best market entry mode for this company?” The model returns a crisp answer: perhaps a joint venture because the market is risky but promising, local knowledge matters, and full acquisition may be too costly. The answer is readable. It may even be reasonable. But the model has already made the first conceptual cut. It has decided what kind of problem this is. Student A experiences relief and mistakes that relief for progress.

Student B is told to do something different first. Before using AI, the student has to identify the decision criteria: control, capital burden, reversibility, speed, institutional risk, partner risk, and learning value. Only then is AI allowed in, and only for a specific purpose: generate two competing frameworks for choosing an entry mode, compare what each framework highlights and ignores, and produce one strong counterargument to the student’s initial leaning. Student B still uses AI. But now the tool enters after the problem has been framed, not before.

In administrative language, both students “used AI.” In educational reality, they did not do the same work at all.

The Cost of Vagueness

This is why generic campus guidance so often feels both noisy and thin. It is noisy because it speaks in the register of disruption, crisis, and innovation. It is thin because it does not specify learning design. A student may use AI to widen a comparison, surface a blind spot, generate a rival stakeholder position, or prepare a structured rebuttal. Used in those places, AI may improve learning. The same system may be used to produce a passable first draft before the student has read closely, chosen a framework, or identified the real tension in the case. Used that way, AI may hollow the task out from the start.

Vague policy cannot distinguish these uses well. Good course design can.

Vagueness has other costs too. It pushes faculty into covert inconsistency. One professor quietly tolerates AI summaries because the reading load is heavy. Another bans AI but never redesigns the assignments, so students simply use it more discreetly. A third wants to experiment but has no clear language for process documentation, source control, or oral defense. The institution then wonders why “policy compliance” looks uneven. The answer is simple. The policy is pretending to do the work of pedagogy.

What the Better Question Reveals

Once the question changes, the design space opens. The teacher must now decide at least five things.

First, what intellectual work must be done before the machine enters. In some courses that means reading the primary text. In others it means framing the problem, defining criteria, or generating one provisional claim without assistance.

Second, what kind of help the machine is allowed to provide. Is it being used to widen the set of plausible perspectives? To organize raw evidence? To simulate opposition? To support revision? These are different uses and should not be collapsed.

Third, what evidence must remain visible. If the final product is now easy to polish, the teacher has to ask what traces of process are necessary: logs, revision notes, oral explanation, annotated drafts, or structured checkpoints.

Fourth, what counts as authorship. If the model produces fluent prose, what makes the resulting work the student’s rather than simply the student’s submission? The answer will not come from intuition. It has to be designed.

Fifth, what part of the work still belongs to judgment. This is the deepest point. If the machine may summarize, compare, and draft, what exactly remains the teacher’s and student’s irreducible responsibility? The answer, I will argue throughout this book, is judgment: the right to define the problem, choose the frame, weigh the evidence, reject the convenient answer, and take responsibility for the conclusion.

What This Book Will Not Do

This book will not offer therapeutic slogans about the future of education. It will not pretend that every faculty member needs to become an expert in every model release. It will not sell prompt tricks as a substitute for course design. And it will not tell teachers to return heroically to a pre-AI classroom that was often less rigorous than memory suggests.

What it will do is more demanding and, I hope, more useful. It will treat AI as a design pressure on higher education. It will ask where the machine belongs, where it does not belong, how assignments should change, how discussion should change, how assessment should change, and how students can still produce work that shows judgment rather than merely polished completion.

That is a less glamorous conversation than most technology launches prefer. It is also the only one worth having.

Chapter 2 AI × Discipline

The phrase “AI literacy” is useful up to a point. Beyond that point it begins to mislead. It suggests that there is a portable skill called using AI and that once students acquire it they can carry it unchanged from marketing to engineering, from teacher education to finance, from international business to public policy. That is convenient language for workshops. It is weak language for curriculum.

The more accurate view is that serious AI use is always nested inside a discipline. It depends on what counts as a worthwhile question, what counts as acceptable evidence, what counts as a sound explanation, and what kind of mistake is unforgivable in that field. A student in corporate strategy does not ask the same kind of question as a student learning qualitative methods. Nor should they.

Three Rooms, One Model

Consider three classrooms using the same underlying generative model.

In the first room, students are studying international trade and the rules of origin that determine whether a product qualifies for tariff treatment under an agreement. The teacher asks the model to become a novice trade coach and explain the problem through a campus shopping district analogy. Later the same exercise becomes more technical. Students feed in the cost structure of an e-bike. The question is no longer “What is trade?” It becomes: can the model help calculate regional value content without skipping the logic or inventing arithmetic?

In the second room, students are studying the collision between Hollywood, streaming platforms, and creator ecosystems. The model is asked to find five facts that violate common intuition about the Oscars, Netflix, and YouTube. Then students are told to verify the sources, sort official data from journalistic claims, and generate an opposing explanation for each supposedly counterintuitive insight. The educational problem here is not explanation in the abstract. It is how to use the machine to investigate platform change without becoming the victim of its first neat narrative.

In the third room, students are doing qualitative case work. They are not asking for a finished answer at all. They are trying to design system instructions for a research assistant that can move from interview transcript to keywords, from keywords to hypotheses, and from hypotheses to evidence without inventing facts. Later, the same students compare a live case with IDEO method cards to ask what research methods were used, what was missing, and what another methodology would surface.

All three rooms are “using AI.” That description is almost useless. The real educational work is different in each case. Trade reasoning demands precision about rules, thresholds, and calculations. Media and platform reasoning demand source scrutiny, multiple interpretations, and a sensitivity to changing business models. Qualitative work demands disciplined interpretation, evidence handling, and protection against over-reading. One model, three disciplines, three different standards of seriousness.

Generic Fluency Is Not Enough

A student can be fluent in prompt syntax and still be intellectually empty. The student may know how to ask for bullet points, role-play, tone shifts, charts, or polished summaries. Fine. If the student cannot distinguish a mechanism from a symptom, or a framework from a slogan, or evidence from commentary, prompt fluency simply accelerates confusion.

This is the central reason I insist on AI × discipline rather than AI alone. In one classroom, students used AI to compare market-entry frameworks. In another, they used it to pressure-test industrial policy and de-risking. In another, they used it to contrast research methods in a digital-transformation case. The prompts looked different on the surface. The real difference sat underneath: each task required disciplinary standards, not general cleverness.

That is why the best prompt in the room is often written by the teacher who knows the field well, not by the person most impressed by the tool. The teacher knows where the naive answer will fail. The teacher knows which variable is decisive and which merely sounds technical. The teacher knows what a field treats as evidence, and what it treats as noise.

Case: Trade, Rules, and the Cost of Vagueness

Take the trade-compliance exercises built around free trade agreements and rules of origin. A superficial AI interaction might ask: “Can this product qualify for tariff exemption?” The model will likely answer quickly and confidently. That confidence is educationally dangerous because the real work lies in specifying the product architecture, the origin of components, the labor allocation, the threshold rule, and the logic of the agreement itself.

In one classroom sequence, students were first asked to use a campus alliance analogy so they could understand why origin rules exist at all. That was the accessible layer. Then the exercise became concrete. One student group worked on Bluetooth earphones. Another worked on e-bikes. They had to explain what happens when batteries come from one country, frames from another, and assembly work from a third. They discovered that the apparent simplicity of the rule hid a demanding evidentiary problem: in real supply chains, upstream cost data are often incomplete, strategic, or hard to verify.

This is what AI × discipline means in practice. The machine can help stage the explanation, visualize the comparison, or even draft a computational plan. But the disciplinary demand remains. The student has to know that a calculation without a clearly specified threshold is worthless, that a tariff outcome without verifiable cost components is suspect, and that legal compliance cannot be reduced to rhetorical clarity.

Case: Hollywood, Streaming, and Counterintuitive Inquiry

Now compare that with a classroom exploring Hollywood, streaming, and creator media. Here the educational risk is different. The model’s default behavior is to return median narratives: streaming changed audience behavior, awards matter less, creators are rising, studios face pressure. None of that is necessarily false. All of it may be boring.

So the teacher redesigns the task. Students must ask for counterintuitive facts. They must force web search. They must run the same task through multiple settings: conventional search, deeper research, and tail-end distribution prompts that try to pull unusual but still defensible hypotheses from the model. Then they must verify source quality and produce opposite explanations.

That workflow teaches a disciplinary habit. In business and media analysis, the first plausible explanation is often not the one worth trusting. What matters is not whether the model can speak fluently about Netflix or the Oscars. What matters is whether students can investigate power, attention, and strategy without confusing a tidy explanation for a tested one.

In the classroom record, students surfaced insights they did not begin with: low-budget films outperforming assumptions about prestige production, ad-supported streaming tiers succeeding more than ideology suggested, mobile gaming functioning as an attention-retention strategy rather than an identity shift, legacy intellectual property re-entering strategic importance after original-content enthusiasm cooled. Those are not just “AI insights.” They are examples of what happens when a teacher uses AI to create investigative discipline.

Case: Qualitative Analysis and Methodological Restraint

A third example comes from qualitative teaching. Students working with interview materials were not told to “analyze the transcript.” They were shown how to build system instructions for a specific research job. The assistant had to know its goal, its steps, its constraints, and its output format. It had to avoid inventing facts. It had to say when uncertainty remained. It had to transform raw material into something like a commercial-insight brief without pretending to have discovered more than the transcript could support.

Later, this work was extended with method comparison. Students took a case, then compared its existing qualitative method to other options drawn from external method cards. The question was not whether the model could sound methodological. The question was whether students could learn that methods are design choices, each of which sees something and misses something.

This too is AI × discipline. The machine does not replace the methodological judgment of the field. It makes that judgment visible.

The Unit of Design Is the Learning Task

If this sounds obvious, good. Universities have still spent much of the last two years ignoring it. The common mistake has been to treat the prompt as the unit of design. The better unit is the learning task. Start there and the rest follows.

What problem is the student solving? What framework should the student consider? What evidence must the student marshal? What position must the student defend? What ambiguity is worth preserving rather than flattening? Only after those questions are answered should the teacher ask what AI is allowed to do.

In this sense, prompting is not primarily a linguistic activity. It is a translation activity. The teacher translates disciplinary judgment into instructions, constraints, checkpoints, and comparison structures. The student eventually learns to do the same. That is why strong AI use is genuinely educational. It forces hidden disciplinary logic to become explicit.

What This Means for Curriculum

The curricular implication is demanding. We should not try to make students “generally good at AI.” We should try to make them better at disciplinary work in an environment where AI exists. That means at least four things.

First, each course should define its non-negotiables: the kind of evidence it trusts, the errors it refuses, and the reasoning moves it wants to see.

Second, each course should state where AI is helpful and where it is corrosive. In one class, AI may be excellent for generating stakeholder views. In another, it may be too risky to use before close reading because it collapses interpretive work too early.

Third, assignments should be redesigned so that the model is forced to serve disciplinary method rather than replace it.

Fourth, faculty development should move beyond tool demos. A department does not need forty prompt tricks. It needs a clear conversation about how different fields want to use or constrain the same family of tools.

The problem is not that AI exists across disciplines. The problem is pretending that this makes disciplines less important. It makes them more important.