On What Teaching Is and Isn't

Higher ed's rush to embrace AI is a catastrophic failure to recognize what education means.

It’s been quite a while since I last wrote a post for this newsletter, mostly because I’ve been tremendously busy. My last post was about what it’s like to teach again, and I’m still pretty shocked that I get to do what I do. It’s summer break now, and although I’m doing some summer teaching, work is not quite as all-consuming as it was during term, and I’ve had more time to think about what exactly this thing I’m doing is, why it’s very often misunderstood even by people who claim to value it, and why it is especially threatened by college and university administrators’ alarmingly fast embrace of AI. If you’re a working teacher at the secondary or college level, you’ve probably already thought about a lot of this; I make no claims to originality, and indeed I’d be extremely alarmed if I were the only person to be thinking this way. But I do apologize in advance if this feels like well-trodden territory to those who are with me in the educational trenches.

Here I want to highlight an old but useful distinction between education and training. When we talk about “training” we don’t often run the risk of being misunderstood: there’s pretty strong and widespread agreement that training means sharpening particular skills in ways that are testable. The success or failure of a course of training may depend on many things, but it is always measurable: we can know if someone has been successfully trained or not, because that success or failure depends on their being able to meet certain metrics. Many college courses involve a certain amount of training, and this is not a pejorative statement: training students in the notation, language, and procedures of a field is an important part of helping them learn. Indeed, any course that has a topic at all will, on the surface, appear as a form of training: one learns the norms and knowledge of (a) particular field(s), one is tested on them, and on that basis one is judged to have been effectively trained or not.

There is, however, another process going on, and whether that process is going well or badly is essentially orthogonal to students’ successes or failures at meeting the metrics of their courses. This is the process that I refer to when I talk about “education.” Let me give an example. A student’s grade might tank spectacularly, to the point that they are forced to withdraw from the course. Such a student has obviously failed the course qua training, but we have no way to know, from the description given, what that failure has contributed to their education, because education as I describe it here is the process by which we become more fully human and more fully ourselves. To educate someone is to help make them more conscious of their human capacities, and above all of their capacities for reason and for action, which are united in our capacity for language.

Because we have language, the capacity to invest arbitrary signs with meaning, there is also a communicative dimension to all our actions. Once we start saying things with words (or, in the case of signed languages, with hand-signs), in a sense we can’t ever stop saying things: our knowledge that communication is possible (indeed, inevitable) makes us hungry for it, and our minds seek after the meaning of others’ actions and gestures as naturally as our bodies seek after nourishment, and indeed for the same reason, because it is only through the economy of meaning—through its creation and apprehension, a sending-forth and taking-in—that mind is cognizable as mind at all. If we are always saying something, we had better learn how to say what we wish to and avoid saying what we wish not to. This is a simple description of an almost inconceivably complicated process: it involves awareness of our deliberations and our reasons for wanting something, our prudential choice to communicate it a certain way, awareness of others and imagination of their internal life, interpretation of their words and gestures…if we didn’t do it every day, a description would make it sound impossible.

Education, as I said before, is the process of becoming more conscious of our faculties of reason and action and thus more intentional in their use. These two ideas (consciousness and intentionality) are conceptually distinct but so closely bound together in practice that it makes no sense to speak of one without the other: both are always in play in our use of reason. In a mathematics course, we “train” to deploy mathematical language and concepts in a way that makes sense within the discourse of mathematics, but at the same time that we do that, we become more conscious of the way we think about and approach problems of cogency and coherence; we learn about the form that our reasoning takes and about its various kinds of fallibility. In language courses, and especially advanced literature courses where a language foreign to use is studied as a medium of art, we train in the fluent use of vocabulary and syntactic structures, but we also learn about the structure of our linguistic faculty and its inseparability from our thinking. The difference in languages’ systems of categorization teaches us about the arbitrariness of signs and broadens our thinking about the modes and kinds of meaning that human beings can create. What’s important to note about this kind of learning is that it has very little to do with one’s “mastery” of the subject. One can fail to learn the many inflections of the Greek verb while still learning a great deal about how human beings use language; one can fail to master the conventions of mathematical proofs even as one arrives at a far greater appreciation for careful and systematic reasoning. When someone says that we learn more from failure than from success, they are in fact making a statement about the relative value of training and education, decisively in favor of the latter.

The problem presented by Large Language Models (LLMs, which is the more accurate term for what many people call AI) is that sufficiently powerful computation can in fact detect the statistical patterns in our language use. This isn’t surprising; language is a medium for the creation of meaning precisely because it is ordered, and anything ordered exhibits patterns that can be recognized and replicated. LLMs have been trained on an enormous corpus of linguistic data (much of it obtained illegally) and thus have developed extraordinarily sophisticated statistical models for how people say or write things, especially in English. We have now reached a stage of computing development in which these programs are able to produce plausible-sounding English (that is, English that does not violate the rules of grammar and that appears to stay on topic) in response to input. The language they produce does not display any overt signs of being generated by a statistical procedure; it appears, at first glance, to be the same kind of English that a human being who spoke or wrote fluent English would produce. And yet it isn’t the same: these programs have no concept of truth or falsehood, and they seem to recognize patterns rather than actually thinking through problems. A recent paper from Apple ran both LLMs and the more complex LRMs (Large Reasoning Models) through a series of logic puzzles that the models had never seen before. All began to fail catastrophically on the most complex problems, and no amount of computational resource allocation allowed them to recover.

Perhaps the most remarkable finding in the paper is that the models were not coachable. Even when given an exact solution algorithm, they broke down. They could not be walked through the task, the way one might walk an enthusiastic six-year-old through the steps of baking cookies. This is a serious fault in a technology whose boosters claim that it has “Ph.D.-level” intelligence, and it suggests quite strongly that whatever is being modeled here, it is not the kind of thinking that human beings do. The paper seems to vindicate the arguments made by John Searle in his now-famous paper “Minds, Brains, and Programs,” which proposes the so-called “Chinese Room” thought experiment to demonstrate precisely these differences. Among Searle’s most persuasive points in that paper is that, in the experiment, he is receiving instructions in English (which he does understand) on how to manipulate the symbols of written Chinese (which he does not). His own experience of the two is qualitatively different: the manipulation of Chinese is experienced as a completely abstract system of rules, applicable only to this set of symbols. English, on the other hand, is embedded within an entire sensory life: it is impossible to abstract it from that context. The relationship between sounds or signs mediates for us a relationship of things, a set of experiences and imaginings; it is part of an irreducible being-in-the-world. The abstract rules for manipulating the Chinese symbols cannot be learned as abstract rules: they are learned only within the context of this language that is itself within sensory life.

The problem that these models present for teaching and learning is that they do not use language the way human beings do. The lack of intentionality is built into the foundational structures of their programming: they’re trained on reward structures that encourage certain kinds of responses, and the idea of “truth” isn’t built into those reward structures because it turns out that “truth” is one of those concepts governed by the Potter Stewart “I know it when I see it” principle: we have a very strong intuitive grasp of what it means for something to be true, but describing the quality that separates a true statement from a false one is unbelievably difficult, and a couple thousand years of philosophy hasn’t been able to fully resolve that question. An AI model cannot intend to tell the truth, and it cannot intend to improve a person or to tell them things they didn’t know. A teacher can be wrong, and can also be told as such, apologize for it, and correct their mistake for the benefit of the class. An AI might correct itself if you say it’s wrong, or it might double down, and this will have nothing to do with whether what it’s saying is true or not. Even though the “economy of meaning” that sustains our minds will latch onto the ordered language that AIs produce as if it meant something, this is a mistake. Whatever is going on within these models is much further removed from thinking than Clever Hans, because at least the horse had a sensory life. It seems like human language, but it cannot do the things that human language does.

The ultimate reason for this inability is that education is a two-way process. It is the less measurable and more important element of school and university life because it is a human relationship that changes all parties involved. In brushing up against a definite other self we learn what we are and what we are not and what we might be: we see ourselves more clearly as one self embedded among many. One cannot see one’s humanity more clearly by treating as human something that fundamentally isn’t. An AI doesn’t make us more conscious of human faculties; it encourages a murky understanding of the two because its programmers want to create the illusion of human thinking. It has no self for us to brush against. It cannot imagine its way into our heads; it does not see us more clearly than we ourselves do and give us what we need, the way a great teacher seems to do. It is ineducable and therefore cannot educate, because the teacher and student are engaged in the same activity: we know this from the fact that this activity can be the basis for friendship between them.

What AI-boosting admins are doing, then, is introducing something that structurally cannot educate students and trying to use it to replace the only people who can. Students’ lives have already become a great deal more isolated: many cannot afford to live at or near their universities, so they commute in and out for class and have very little time to get to know other students. Others skip in-person lectures, watch the recordings in their rooms, and wonder why finding friends seems so difficult. All are constantly on their phones, and all are reading far less than students used to. The places where this isn’t happening, or at least is not nearly as big a problem, are those places (like St. John’s College or Deep Springs or, if I may be so bold as to say, the University of Tulsa Honors College) that have made deliberate efforts to show students the connection between their work and their educations: these students see that their work helps to make them more human, and as a result they have an easier time seeing why it’s worth doing themselves.

I think there isn’t an easy fix to this that doesn’t involve a serious re-evaluation of what colleges and universities are doing. If they see themselves solely as places where students are equipped with “skills,” then there is probably no saving them: AI promises efficiency and cost savings, and although these things will probably not materialize (since the AI companies will jack up their prices once they’ve achieved market saturation), colleges and universities that embrace that model will probably cease to exist in any form that we would recognize. The only way through is to embrace the university’s educational mission, over and above the training it provides as a means to that end. I am not sure how many institutions exist in which a critical mass of faculty believes in this. I am extremely fortunate that my particular division is one such place, but our view is not necessarily shared by all of our colleagues. I am convinced, however, that faculty and administrators who fail to come around to this view will end up destroying the institutions that they claim to be helping.