More than calculators: Why large language models threaten learning, teaching, and education
I remember getting my first personal calculator in 3rd grade. It wasn’t the first time I’d used one — that was with my mom’s calculator, which she used to balance her checkbook and had the most incredible built-in dot matrix printer—but my 3rd grade calculator was the first that was mine. It was blue, solar powered, and had basic arithmetic functions. When the power ran out, the single row LCD display would dim, and I would run to the window to recharge it with the sun.
The calculator wasn’t very instrumental in my math learning, though. My 3rd grade math teacher had spent months making us do arithmetic in Roman numerals to reinforce the power of Arabic numerals. Once we got to drill and practice multiplication tables, we were amazed at the speed with which we could do our multiples in Arabic, relative to the Roman number system. We were actually faster at recalling our multiples than we were typing them into the calculator, and so for most of us, the calculator was a novelty, a toy. The most entertaining thing about it was typing 1001 to make a Homer Simpson face, or typing 8008 to make the word “BOOB”. Other than that, it was just an occasional tool for doing bigger calculations than we could memorize.
Of course, calculators became more powerful. By the time I was in middle school, graphing calculators could do far more than arithmetic. They could do any function in math, solve for x, even compute derivatives and integrals. They could plot 3 dimensional functions. We even used them to write interactive games. They still didn’t play much of a role in our ability to do algebra, geometry, or calculus, because our grades were always based on showing and explaining our work towards the right answer, not the right answer itself. And even if my TI-82 could show its work, it could not explain it. The scare in education about calculators threatening math education was mostly a false alarm. They’ve caused no apparent harm to math aptitude and when students are allowed to use them to augment their learning, they help in some cases, mirroring my own experiences (see Ellington 2003, A meta-analysis of the effects of calculators on students’ achievement, J. Math Education).
Today, a new technology is raising similar concern: large language models (LLMs). This time, however, rather than threatening math education, they threaten any education that involves writing, with their ability to generate likely sequences of words that mimic human writing. And they followed a similar progression of history. The dream of digital computers, which first began in the 19th century, became a reality in the 1940’s, and took 40 years before they made it to the classroom as calculators. And the earliest language models, proposed in the 1980’s, took about 40 years to find their way to schools. And just as with calculators, their entry into classrooms has caused fear, panic, curiosity, optimism, and every other emotion, just as calculators did. Teachers and schools are left, yet again, to sort through a technological mess they didn’t ask for (and at the worst possible time, after a pandemic, and record declines in teacher certification enrollments).
I have been asked — by teachers, school leaders, state superintendents—for my opinion. Are large language models like calculators, where they’ll have either no effect, or some positive effects? Will they have even greater effects, heralding a transformation of education by freeing teachers from having to teach content? Are they an unqualified threat to modern education that will have devastating consequences to student learning? Or is it the standard “it depends”, where there will be some good and some bad, depending on how they are used?
Obviously, there is no way to predict any of these futures with certainty. But we can make predictions, and we can inform those predictions with the best knowledge we have about what LLMs are, how school works, and what learning is. So here is my informed prediction: LLMs will ultimately be a net harm to student learning and schools, accelerating the collapse of public education. Let’s step through why I predict this, one fact at a time, and see if you arrive at the same conclusions.
Why collapse?
First, let’s start with what LLMs are. The most visible ones (e.g., ChatGPT, and the various models released by Microsoft and Google), essentially predict likely sequences of symbols given some prior sequence of symbols, where likelihood is based on documents posted publicly on the internet. That includes all of the widely known sites on the internet — Wikipedia, Reddit, newspapers—but also all of its obscure random corners, such as these essays I occasionally post. It does not include the “dark” web, such as corporate intranets, private Discords and Slacks, private discussion forums, private mailing lists, or most of the world’s books. It doesn’t include a large amount of the web that’s written in languages that these LLM providers have deprioritized. It doesn’t include non-digital collections like that at the U.S. Library of Congress. It doesn’t include most of the audio content on the web, such as podcasts and video, as transcribing that volume of content, and doing it correctly, isn’t quite feasible. And it doesn’t include the recent web, since it takes so much time to process and train models on the entire public internet. So when we talk about “likely sequences of symbols”, the word “likely” means “the collective writing of mostly English-fluent adults likely to write publicly online in the past decade”, not “all of human knowledge or thought”.
With those limitations, let’s take a moment and talk about what I mean when I say “student learning”. I’m not talking about standardized test outcomes, graduation rates, grades, college placement, or other measurements of learning. I’m talking about learning itself. This certainly includes all of the skills that these assessments intend to measure. But also the human ability to communicate in their own voice, to comprehend others’ reading, to reason with logic, to comprehend human emotions, to navigate uncertainty, to problem solve, to resolve human conflict, to believe that they can dive headfirst into a complex wicked problem and find a path forward that reconciles conflicting human values, to understand who they are, what they value, and how that’s changed over their lifetime, to form relationships and maintain them, to respond to change. Learning is all of these things and and inherent part of being human.
And what about “school”? In the most ideal sense, school is community and relationships that enable learning, as defined above, to happen, most often through the relationships of teachers and peers. I’ll use my own learning as an example. School certainly taught me everything on the tests: reading, writing, mathematics, history, science. But it also taught me that there was a world of knowledge well beyond books and the internet and that it would take struggle to find it. It taught me that people are very different, and struggle to understand each other. It taught me that safety isn’t about metal detectors, but trust and mutual respect. It taught me that wealth is destructive. It fostered my artistic taste. It shaped my values, but challenged me to question them too. These were not ideas on exams or in lesson plans, but I learned them nonetheless, because my teachers in the 1980’s taught my whole self, not yet constrained by high stakes exams.
Now let’s talk about school and learning in the less ideal sense. School, as it is often structured in most countries, is not about the things above. It is unfortunately about the tests, the grades, the credentials, because the public has largely come to frame schools in utilitarian, capitalist, industrialized terms. In our modern institutions, learning math and science is not about learning math and science, it is about achieving high scores on indicators of learning. And teaching math and science, as much as teachers would love it to be about teaching math and science, is often about ensuring students achieve high scores. This is because observing the kind of learning that actually matters is too difficult at scale, and teachers are too often not trusted to observe it “objectively”, and so teacher performance, compensation, and school resources are often tied to these indicators of learning. Teachers know this, as do students, and they have responded accordingly, by tragically shifting their focus to indicators of learning instead of learning.
The student reaction to this shift is important to understand further. In essence, all students are put in a bind. If they are interested in the broad kinds of learning I described above, they are certainly not being rewarded or valued for it by our education institutions and so they are left to erase a part of themselves in order to survive in the systems that define the majority of their waking hours. And if they aren’t interested in the kind of learning I described above, they are left to play a game, in which achieving freedom to do the things they are interested in requires finding the shortest path to winning the game (e.g., “cheating”), so they can preserve any capacity for the things they value. The only students for whom this system works are the ones who value the learning narrowly defined within learning standards, or simply are most interested in meeting others’ expectations. In these rare cases, incentives are aligned, and all is “well”. (For the record, I was in this latter group: I was interested in what schools taught and wanted to be compliant).
And so most “school” is one in which students earn their freedom by gaming summative assessments. Learning is one option for gaming, and the intended and most fruitful one; but the many other paths offer less resistance, less struggle, and more freedom, and so they are the choice that most students take.
This is the status quo into which LLMs have been inserted. And it it is having predictable effects. First and foremost, any student whose interests aren’t aligned with what is being measured already had an incentive to find the shortest path through an assessment. Before LLMs, that menu included: 1) copying a peer’s answer, 2) finding an answer online, 3) avoiding struggling through the inherent challenge of learning by simply not doing homework. After LLMs, the path is quite clear: copy the prompt, copy the answer. The only risk in this is that a teacher attempts to detect LLM-generated content (which isn’t accurate enough to justify doing), and accuses the student of cheating. But I think most students have quickly learned that these accusations can’t be proven. And so avoiding learning is what students do.
I don’t blame students at all for this behavior: they have other things they want to do, and there are often other more important things to learn, and so getting an unwanted task out of the way is a very rational thing to do. (We’ll set aside the question of ethics, integrity, and honesty for a moment, since most schools are not really places that teach those things anymore). And so we should expect that the vast majority of disinterested students are only going to have greater incentives and ability to avoid learning. At best, they might learn some skills in transcription and editing, as they try to mask their reliance on generated content.
But even for students whose interests are aligned with school, LLMs change the equation in problematic ways. Consider, for example, a student who sincerely wants to learn algebra. They get a problem from the teacher, and struggle with it, because algebra and learning are hard. So they seek assistance from an LLM, and it steps through an explanation of the solution. The student diligently reads the solution, and writes down the solution in their own words, ensuring they understand how it was solved. They get a great score, they have a sense of having learned, the teacher has evidence that they learned, and so all is well, right?
Well, wrong. The most problematic aspect of this scenario is that comprehending an explanation of a solution is not the same skill as finding a solution. The student, without anyone knowing, shifted the task from one of learning to solve algebra problems to one of reading comprehension. They are not learning algebra problem solving, they are learning to read and evaluate algebra solutions. If they will only ever encounter problems that LLMs can answer, and they will always have an LLM, and LLMs will always generate right answers, then may be this is fine. That position essentially posits that because LLMs exist, humans knowing algebra does not matter. But this is a lot of ifs, and so more likely, the student will encounter problems they haven’t encountered and they won’t be able to describe them to an LLM, because they don’t know algebra. In the same way that we have watched human ability to navigate decline as GPS has become ubiquitous.
But there are more subtle problems in both of these scenarios, aside from not learning algebra. One is that LLMs are wrong. A lot. And especially when prompted on uncommon problems or an uncommon natural language. Students that aren’t fluent in English, or have creative teachers that try to avoid common assessments, are more likely to get wrong answers, and yet LLMs will still happily generate a “likely” answer. Students, not knowing algebra, will have few skills to detect these flaws, introducing noise and randomness into their comprehension, shifting learning from a process of responding to targeted feedback from a trained to teacher, to a process of searching a space of probable answers that they are not yet capable of judging. LLMs serve as an inequitable wrench in the already complex orchestration of self-regulated learning skills that so many students already struggle with. They amplify these inequities by rewarding “normal” and punishing difference.
Another issue with these scenarios is that teachers themselves often do not know the subtle and often profound limitations of LLMs. I hear teachers—or state superintendents— excitedly talking about how they have free teaching assistants, how they can more quickly generate lesson plans, how they can easily detect cheating with the right tools. But none of these things are actually possible. LLMs cannot teach; they can only generate lesson plans that people have shared online, and not particularly good ones; and there is no reliable method to detect what was generated. The hype machine has readily deceived educators and education leaders into thinking that LLMs are some kind of intelligent education aid, when in fact, they are just generating likely things that other educators have posted online, whether they are good or not. This lack of LLM literacy, coupled with the hype, is leading educators to give up the irreplaceable human insight required to structure learning for their students.
The broader trends amongst students, teachers, and even policy makers, is that every group is willingly adopting a tool that supplants the broad scope of human learning and ability with a narrow tool that generates a pastiche of whatever was posted online. This vision of education is one that says “we don’t need to know things, or create things, or write, or think; we can delegate all of that to the English-fluent adults who came before us, and then just edit their work if it seems wrong or bad.”
What is education for?
I question whether this vision for human experience is desirable, or even viable. In the case of calculators, the loss of manual arithmetic skills seemed relatively innocuous. The only reason it was seen as valuable before calculators was because we valued the ability to calculate and we valued the reasoning it required. Losing the ability to do manual arithmetic did not supplant our ability to do arithmetic, nor did it supplant the need to reason. We still had to do math, we just stopped having to do arithmetic algorithms.
LLMs, however, within the context of schools, are supplanting thought. They are short circuiting students’ opportunities to write, synthesize, reason, and struggle by perfectly reinforcing a vision that equates learning with correct answers. And they are short circuiting teachers’ opportunities to craft learning opportunities for their students that respond to who they are, what they know, and what they want to learn. What we will be left is teachers who cannot do instructional design, students who cannot learn, and an inescapable dependency on whatever was published on the internet circa 2022, when humans used to write. I saw this at scale with the 200 new college students I taught this past quarter. Many believed there was no reason to learn anything anymore, because LLMs could do it all. Many had stopped writing, and balked at the idea that I expected them to. Most didn’t come to class, or came to class and just watched TikTok instead of engaging in active learning.
Some might think I’m overselling LLMs capabilities. Won’t students and teachers ultimately reject these tools, because they just aren’t very good at supporting learning and teaching? That might be true if schools met the ideal I described before. But at least in the United States, where public schools are chronically underfunded, teachers are being attacked for acknowledging their students’ diverse identities, and students are under immense pressure to maximize indicators of learning in order to access basic resources like housing, health care, and safety, no. Our system of education, deeply linked to our late stage capitalist system of exploitation, does not particularly care about the abilities of future generations. It wants the cheapest possible route to return on investment right now, which means low taxes and hiring from abroad, where learning is viewed as an investment. I see this in the way that teachers and education leaders are eagerly trying to integrate LLMs into learning, with the false hope that they will help them overcome resource scarcity and finally address inequities by giving every learner access to a personal teacher. Of course, it will not. Instead, LLMs will accelerate the collapse public education in the U.S., by amplifying these existing forces that were already taking us there.
Some might think I’m underselling LLMs capabilities. Won’t LLMs get good enough that they really can help with teaching and learning? That presumes that the inherent nature of LLMs are actually a viable tool for learning. It also presumes that LLMs will get better. (I’ll tackle the former here, and leave debates about whether LLMs will actually improve to others). Here’s why that premise is suspect: learning requires struggle. It’s hard. It takes focus, constructive feedback, confusion, repetition, someone caring that you continue to do all of these things, despite its difficulties. It can be fun, but it’s not always pleasant. And there’s nothing about a tool that promises to replace all of these things with answers we can lightly edit that acknowledges this essential struggle. The whole point of an LLM, when applied to learning, is to remove struggle, which by supplants learning. So no matter how good LLMs are or become, they will always prioritize a future that replaces human learning with convenience.
If we care about people having knowledge and abilities—and we should, because abilities are the only things that keep us alive, not to mention LLMs functioning —then education must stay focused on them, and not let LLMs subvert the purpose of the institution.
What do we do?
I’m sure some of you disagree. And I’ll be excited to read your rebuttals. But for now, let’s assume you agree. What should we do? Should we ban LLMs from schools? Should we ban them in general? Should we teach teachers and students who to use them more responsibly? Or perhaps we should put guardrails on LLMs themselves, holding model providers accountable?
There are issues with all of those proposals (e.g., banning doesn’t work, responsible use requires incredible self-control, capitalism doesn’t really hold anyone accountable). But ultimately, I think all of these different recommendations miss the point: schools, as they often work today, do not actually incentivize learning and there is no use or disuse of LLMs that can change this. Putting our attention on LLMs, therefore, is mostly a distraction. We should be focused on how to change teaching and learning so that LLMs and all of the other technological distractions that have dominated youth attention are simply irrelevant. If schools were already in collapse because of their broken incentives, and LLMs are just accelerating it, then it’s about time we took seriously school reforms that have better incentives, before that collapse comes. And LLMs aren’t going to be a tool to get us there, because reform doesn’t need technology, it needs vision, values, and political will.
There are many education philosophies to choose from. I’m partial to Dewey, Freire, Foucault, and hooks, who all have important things to say about knowledge, learning, and liberation. But none of them account for a world in which freely available tools comprehensively devalue human ability. I think we need to begin imagining school reforms that radically examine what we want humanity to be, what we want our relationship to tools to be, what rights we deserve, and who gets the power to decide. These philosophies all ask these questions, and offer visions for how to engage them in school, and were relevant before LLMs become public.
Of course, they do not tell us how to practically manage the availability of technologies that subvert human capacity to learn. When I try to bridge these philosophies to practice, I think of the most radical visions of responsive teaching. Ones that are ruthlessly centered on relationships, trust, and individual student identities and interests. Ones that dismantle the industrialized, structured nature of education, and make space, very early in education, for youth liberation from their limiting situations. Ones that ask students to critically examine everything: their home lives, the technologies in their homes, their parents’ rules, the way their school is structured. Deconstructing the social worlds around them would be the only way to build a critical foundation on which they could judge for themselves what role they want any technology to play in their lives, but also what world they want to build with each other, and with the adults who support them. This kind of liberatory vision of learning is the only one that prevents youth from passively accepting the world its latest innovations as immutable fact and inevitable future.
Of course, this is exactly the kind of education that is being outlawed in the U.S. now. Do not question authority. Do not question parents. Do not question markets. Do not question God. Do not question gender. Do not question race. Do not question anything. In a way, this broader oppressive political vision of schools mirrors the underlying philosophy of an LLM: the only things that can be said are that which has been said before by our collective definition of “normal,” and at the expense of anyone on the margins.
So what, then, should a teacher do about LLMs, while we fight for this freedom? I think the only thing to do is resist. Refuse to use them. Highlight their flaws. Demonstrate that they aren’t the intelligent life their creators claim; that they are frequently wrong; that they will not save humanity, or bring forth eternal human prosperity. Celebrate human achievement by highlighting and validating that it requires struggle that can be playful and rewarding. And reinforce that LLMs, like most modern technologies, are deeply cynical about the rich capacity of humanity to create, surprise, commune, and love, and that we should not give up these things so willingly, lest we give up our humanity. I’ve tried these pedagogies myself, and found that while some youth ignore me in favor of feeds, most perk up at the mention of another way of seeing the world. And then they start asking questions.
