The previous essay in this series closed with a question: if knowing requires formation, and formation requires resistance, and AI removes resistance, what does the pathway from novice to expert look like in an AI-saturated world?

This essay attempts to answer that question with an account of what formation actually requires, what AI does and does not threaten in that process, and what institutions can do to preserve the conditions under which knowledge develops.

The argument is not that AI in education is bad. It is that AI, used without deliberate pedagogical intent, defaults to removing the friction that forms foundational knowledge and thinking structures in the brain. This default, if left uncorrected, produces a particular kind of learner: fluent, confident, and intellectually thin.1

-

The lecture hall is a knowledge delivery mechanism. It is efficient. A single expert can transmit information to hundreds of students simultaneously. As a formation mechanism, it is remarkably weak. The student who sits in a lecture hall and absorbs content has been informed, not formed. The difference, as the previous essays in this series have established, is not trivial.

Seminar culture is something else entirely. For those not from academia, the seminar is an instructional technique where professors, students, and researchers present scholarly work and enter into subject matter through discussion and debate.

The seminar room is an arena. It is a space in which something real is at stake, in which participants must show up prepared and defend their thinking under questioning, in which the act of being wrong in front of others—and being helped to understand why—is the primary mechanism for learning.

The research tradition behind this is substantial. The psychologist Manu Kapur has spent two decades studying what he calls productive failure—the finding that students who struggle with a problem before being taught the solution develop deeper, more transferable understanding than those who receive instruction first. The struggle is not inefficiency. It is the pedagogy. What Kapur’s work establishes empirically is what the seminar tradition understood intuitively: that the mind forms itself against resistance, and that removing the resistance prevents the formation.

Robert Bjork’s research on desirable difficulties extends this finding. Conditions that make learning feel harder in the short term produce more durable and flexible knowledge structures than conditions that make learning feel easy. The student who retrieves information under pressure, who must reconstruct rather than recognize, who encounters material in varied contexts—that student is building something the student who receives frictionless access to information is not.

Ultimately, that is the definition of modern pedagogy: how to move a learner through subject matter in a way that requires them to struggle with it productively and sit in the problem space for just the right amount of time.

This is also one of two primary premises behind games-based learning. Bernard Suits, the philosopher of play, defined a game as “a voluntary attempt to overcome an unnecessary obstacle.” Read that again.

It is not a requirement that the player overcome the obstacle. They attempt to overcome it voluntarily, because they are in the flow state and the game is fun (i.e., it makes the obstacle worth it). Flow, per Mihaly Csikszentmihalyi’s research, is “the state of optimal engagement that occurs when challenge and competence are precisely balanced.” Flow is not comfort. Flow is formation. It is the feeling of being at your own edge, working at the exact boundary where effort is required and mastery is not yet complete. You get flow from two places: fun and curiosity. The seminar room, at its best, is engineered for exactly that state. So, it turns out, is a well-designed game.

-

This challenge of pedagogy isn’t only technical. It also has a cultural component, which is observed through a historical lens when we study previous academic institutions and pedagogical structures.

Culture is a catalyst. It multiplies the efficacy of every other intervention. The same student, with the same tools, in the same classroom, will use those tools to accelerate their learning or to avoid it depending on the context in which they are embedded. A student who grows up in an academically oriented environment will use the internet (or AI) to deepen their understanding. A student in a different cultural context will use the same tool to distract themselves or to avoid real work.

The Jesuit colleges of early modern Europe understood this. They did not inherit a culture suitable for learning. They built one deliberately by manipulating the social hierarchy of their institutions so that the students who commanded respect were those who embodied the values the institution was trying to instill. When the students who were looked up to were serious scholars, scholarship became the ambient norm. The culture did more pedagogical work than the curriculum.

Montessori had a similar interpretation. The prepared environment—the deliberate arrangement of the physical and social space of learning—was not decoration. It was the primary instrument of formation. The culture of a Montessori classroom, once established, was transmitted not by instruction but by imitation. Older students modeled the norms and younger students absorbed them.

The questions for AI-era pedagogy are not how to ban tools that remove friction. It is how to build cultures in which friction is valued, in which the struggle to understand, rather than the appearance of understanding, is what earns respect. It is how to teach in a way that forces students into genuine friction regardless of access to AI tools. It is how to teach in a way that encourages students to use AI to promote intellectual friction, even if the student is not able to articulate that process. These are likely harder problems than purely technical interventions. The solutions feed directly on an inherently human quality: curiosity.

What does this culture look like in practice? It looks like a classroom in which the question “how did you arrive at that?” is asked more often than “what is the answer?” It looks like assessment that rewards the visibility of reasoning over the polish of output. It looks like a seminar in which students are expected to have genuinely encountered the material (not summarized or synthesized), and to bring that encounter into the room with other peers and practitioners.

-

So, how do we design learning experiences that require genuine encounter in a world where AI is literally producing simulacra of human experience on demand?

The answer is not to make assignments harder to complete with AI. It is to make them genuinely unanswerable without authentic engagement. Design for the kind of response that can only come from someone who has actually been through something.

Consider the difference between these two prompts:

1. Summarize the primary beliefs of the main character through chapter 15.

2. Find a passage in this book that made you stop. Not because you agreed with it, not because it was the most important passage, but because something in it stalled you—because it clashed with something you already believed or felt. Bring that passage to the seminar and be prepared to discuss your experience.

The first prompt is answerable by summary and synthesis. The second is not. The response it demands is personal: it requires the student to have read with enough presence to notice their own resistance, to have taken that resistance seriously, and to be able to articulate what it revealed. A student who did not actually read cannot fake this well enough to survive a seminar discussion. A student who did read will find that the prompt makes the reading matter in a way that drives introspection and meta-awareness of the formation process.

This is the design principle: pedagogy and assessment that requires the student’s specific, situated experience is assessment that AI cannot replace.

The seminar is the natural home for this kind of assessment, which is part of why leveraging it as a primary pedagogical form matters. In seminar, reasoning is visible. The student must show up in real time and think in front of others. They must respond to questions they did not anticipate. They must revise their position when someone pushes back with something they had not considered. None of this is replicable by proxy. The seminar is structurally resistant to AI. This is not because AI cannot produce fluent text about the material, but because the seminar is not about producing text or demonstrating polish. It is about showing up in the arena.

-

There is an even more challenging problem underneath the assessment design question.

The students who would benefit most from using AI as a Socratic tutor—asking it to challenge reasoning, to steelman the opposing view, to identify the weakest link in the argument—are the students least likely to do so spontaneously. Novice learners do not yet know what they do not know. They cannot evaluate AI output critically because they have not yet developed the knowledge structure against which such evaluation occurs. Left to their own devices, they will use AI to fill gaps rather than to deepen understanding, not always out of laziness but out of the entirely reasonable assumption that getting the right answer is the whole point.

Teaching these students is less about knowledge transfer and more about unlocking their minds (fueled by their own curiosity). This means teaching question formation, how to inquire, how to productively fail, how to explore a conceptual space quickly with no a priori knowledge.

This means metacognition: the capacity to think about one’s own thinking and to evaluate the quality of one’s own understanding. The explicit teaching of metacognition in an AI-saturated environment includes teaching students how to use AI in ways that build rather than bypass their developing knowledge structures.

This is genuinely new curriculum territory. No established pedagogical tradition has addressed it, because the tools are new. Vygotsky’s zone of proximal development provides a framework. This is the insight that learning happens most powerfully at the edge of what a student can do with support but not without it.

The instructor’s job, in this framing, is not to deliver knowledge, but to identify where each student’s edge or ceiling is, and to create the conditions under which this boundary can be expanded. AI, used well, can expand that boundary rather than shortcut it.

For my time instructing at the CIA, I used to describe my job in the classroom as meeting the student where they are and helping them take one additional step forward. Every artifact in the classroom, every assignment, every lecture topic was an invitation to the student to speak up, ask a question, and show me their ceiling. My job as the instructor was to go meet the student in the place they identified. It is a challenging thing to do both with a classroom and with each student individually, but good instructors are masterful in this very practice.

“Once you see the boundaries of your environment, they are no longer the boundaries of your environment.” – Marshall McLuhan

So what does this metacognitive instruction look like in a classroom? It looks like assignments that require students to document their AI interactions alongside their work: not just what they produced, but how they used AI to produce it, what they asked, what came back, what they accepted, and what they questioned and why.

It looks like seminar discussions that begin not with the assignment output but with the process: where did you get stuck, what did you reach for when you got stuck, and what happened when you reached for it, what subsequent questions did you ask as a result, and what other sources did you then consult as a result of this inquiry?

It looks like an instructor who models the use of AI as an intellectual interlocutor rather than an answer machine—who demonstrates, in public, what it looks like to push back on a generated response, to test it against independent knowledge, to treat it as a starting point rather than a conclusion.

The goal is to build a student who has internalized the distinction between using AI to think and using AI to get answers. That discernment between genuine understanding and its simulacrum is the epistemological question this series began with. It just turns out to be a pedagogical question as well.

What Institutions Should Actually Do

Redesign assessment around process, not product. The essay that can be produced by AI is not a useful instrument of formation. The process that produces the essay is—the encounter with the material, the struggle to organize a response, the revision with feedback. Assessment should be redesigned to make the process visible: oral examinations, process portfolios, in-class argument, documented (and graded) AI interaction logs.

Restore the seminar as a primary pedagogical form. The lecture is both efficient and weak. The seminar is costly and strong. It requires smaller class sizes, more adept faculty, and more instructor preparation time. But the research university chose the seminar for reasons that have not changed: it is the form of learning that most closely approximates formation-through-encounter. It is the one pedagogical form that is structurally resistant to AI substitution because it requires presence, real-time reasoning, genuine interaction, and discernment in a way that no proxy can replicate.

Teach metacognition explicitly. Make the distinction between using AI to think and using AI to know things a named, discussed, assessed part of the curriculum. Assign students to document and reflect on their AI use. Model expert AI use and analyze these interactions in class. Design seminars around the process of learning, not only its products.

Reframe the instructor’s role. The instructor’s irreplaceable function in an AI-saturated environment is not to convey content. AI does that well enough. It is to model expert thinking, to make the reasoning process visible, and to create the conditions under which students can develop their own knowledge structures through genuine encounter. This is closer to what the best seminar instructors have always done. It requires a shift in how institutions hire, train, and evaluate teaching, perhaps away from content expertise as the primary criterion and toward intellectual environment design and facilitation.

Be deliberate about the productive struggle. Design courses so that the formative struggle cannot be bypassed, not by making AI unavailable, which is both futile and philosophically incoherent, but by making the struggle itself the object of assessment. Give students problems before they have the tools to solve them. Give students impossible tasks to maintain focus on process, not outcome. Kapur’s productive failure model is the empirical validation of this principle: the student who grapples with a problem before receiving instruction develops a more robust and flexible understanding than the student who receives instruction first.

-

The previous essay offered an observation that is worth returning to here:

“You can live on an inheritance for a while. You cannot educate a civilization on inherited judgment forever.” - Brendan McCord

The generation of students now entering higher education is the first to grow up with ambient access to AI. They are also the generation that will eventually be responsible for producing the judgment that the next generation will inherit. If that judgment is not formed, if it is accessed rather than earned, the inheritance will be hollow.

The student who emerges from a well-designed educational encounter with AI is not a student who knows less than their pre-AI counterpart. They are potentially a student who knows differently, one who has learned to use extraordinary access as a tool for both creating and deepening understanding. That student is the goal. Building the institutional conditions that produce them is the future of higher education.

-

This is the third essay in a series on AI, knowing, and what it means to think in the age of intelligent machines. The first essay examined AI and expert cognition. The second examined what it means to know something as opposed to merely having access to it. The fourth will turn to a different question: what AI does at the frontier of existing expertise, and what it means for practitioners who are already deep inside a domain.