LLMs and the Library Card Fallacy

the LLM tutor story is the Khan Academy story is the MOOCs story

… no

I think Jay Caspian Kang is guilty of what I call the Library Card Fallacy, which has consequences for education that go far beyond current debates about AI’s role in the future of teaching and learning. But I don’t blame him in particular, as there’s an awful lot of that sort of thing going around.

The rise of LLMs has brought “Is college dead???” back to the forefront of the discourse lately. Though it’s now framed as an AI conversation, this kind of chin-stroking higher education doom-saying has been going on for far longer than the LLM era; the late 2000s and early 2010s in particular saw a ton of this Clay Shirky-style, “we have Google now so college is going to disappear” stuff too. As the cost of college became more and more of a scandal, and especially after the financial crisis of 2008 temporarily crushed the entry-level job market, these predictions were expressed more and more often in terms of simple economic self-interest : eventually students would rationally conclude that it made more sense to simply forego college and begin their working years four years earlier, without taking on a mountain of student loan debt. The crash would come and a ton of colleges would close and going to college would go back to being a rare curio that a small slice of Americans took advantage of.

This did not happen.

preorder now

Why? Well, for one, 18-year-olds are not exactly known for a surfeit of practicality. That might sound insulting, but if you think about what “starting your career journey early” means, it’s quite sensible: for most people, it means spending four years of your prime young adulthood squeezing into a sad cubicle job instead of spending them getting drunk and trying to get laid and, yes, having stimulating conversations with cool professors. (My considerable experience in higher education tells me that college kids care way more about that last part than the stereotype allows.) These conversations always revolve around hypothetical self-directed autodidacts with grand ambitions and a future filled with influencing and investing and innovating…. Most people are nothing like that. Indeed, even most people in the educated, higher-earning classes aren’t like that, don’t go out and forge careers blazing new trails of self-employment or even in high-status but conventional jobs like lawyer or doctor; a huge number end up in stable and relatively well-paying but unfulfilling generic white collar work. I have no interest in engaging in the pejorative discussion about “bullshit jobs” here. I am asking you: if you were 18, would you really rush to begin your decades-long “employment journey” in miscellaneous low-glamour cubicle work four years older, in the prime of your youth? I wouldn’t.

Now, a lot of people are suggesting (as Kang does) that young people won’t feel the need to go to college because they can just sit in front of their computer in their sad apartment and learn all they need to know from ChatGPT. Education is about accessing information, LLMs have the information, QED. In one entry in his series of essays on the topic, Kang asks whether “a fifteen-year-old hellbent on a journalism career be best served by working himself to the bone both academically and extracurricularly to get into Harvard, or should he just start a Twitch stream and get to work?” Let’s set aside the fact that, as Kang himself concedes, most people just aren’t like this; again, even when we restrict our perspective to the more ambitious and proficient young people who might succeed in college, we’re mostly talking about people who just want to get a good white collar job at a health insurance company or similar and have time for Netflix at home in the evenings. Even beyond that, Kang’s supposition makes perfect sense if you assume that people are perfectly efficient utility maximizers, but of course we aren’t, and crucially, that fifteen-year-old doesn’t want to sit alone in a sad apartment with a ring light blasting him in the face for his entire youth. Perhaps the Gen Z stereotype suggests he would, but I’m telling you, that’s not what most young people want to be doing. Because it’s depressing!

It should go without saying that none of the people who make these predictions have actually done what they say young people should do; they all went to college themselves. And of course they did. College is great. You’re only young once, and going to college is fun and fulfilling. It’s always easy to be perfectly actuarial about someone else’s theoretical best interest. It’s quite another thing to decide that you’re going to IncomeMaxx in a way that denies you social opportunity, unstructured time, intellectual fulfillment, and living in a way consonant with the great American id’s vision of a successful life.

That Kang makes the same predictions that David Brooks and Tom Friedman made before him is just one of those things, the way it goes. What Kang and Friedman and Brooks and Shirky and Hollis Robbins are all seemingly incapable of understanding is that there is a vast gulf between what young people say about how they feel and what actually motivates their behavior. Kang writes

In 2013, seventy-four per cent of eighteen-to-thirty-four-year-olds polled by Gallup said that a college education was “very important.” By 2019, three years before the public adoption of ChatGPT, that number had dropped to forty-three per cent; it fell again, in 2025, to thirty-five per cent, a decline that represented the steepest drop among all age groups that were surveyed.

OK. That’s what they said in a poll. But what did those kids actually do? Did they really forego college? The enrollment percentages in 2013 and 2019 and 2025 suggest that the college-ready, college-oriented students did not actually act in accordance with those attitudes. Because when push comes to shove, our personal resentments and our own perceived best interest are two different things. It turns out that a lot of young people who find to be a college dubious prospect in the abstract are not actually rushing to get that job at Geico at 18 themselves. American colleges do have an enrollment problem, but that’s the product of demographic change. And Shirky’s relentless doomsaying has failed to come true in anything like the scales he’s declared, over and over and over again. Colleges and universities have proven themselves to be arguably the most tenacious and adaptable of all human institutions; there are far more higher education organizations extant that are more than 300 years old than there are governments that have existed in the same form for that long. The reality is that, as traditional media continues to suffer its never-ending sickness, the attraction to The End of X or The Beginning of Y framing grows and grows. But most things stay mostly the same most of the time, no matter how badly prestige media wants to declare everything to be on the brink of revolution. In reality, most of human history is not a matter of apocalypse or genesis but damp continuity, which is precisely why journalists keep trying to set things on fire: they need the smoke to get attention.

---

OK. But I was suggesting that there’s a deeper reason to be very dubious about these claims that new technologies will enable people to educate themselves and thus destroy all manner of schools. It comes down to a basic reality that many people really struggle to accept: education has almost nothing to do with access to information. Which brings us to the Library Card Fallacy.

We have very recent evidence that helps us understand why the gatekeeping and incentive structure of higher education aren’t so easily replaced with student-directed technological alternatives, and that experience demonstrates how the Library Card Fallacy misleads. The Library Card Fallacy is the mistaken notion that the purpose of education is to transfer information from teacher to student, and thus that schools and teachers are subject to disruption when any technology comes around that democratizes access to information. The trouble with this theory is that information has been very broadly available for a hundred years or more; depending on how exactly you want to define things, most Americans have enjoyed public library access since sometime between the 1890s and the 1920s. In the late 1990s, people started saying that Google was an existential threat to colleges and universities - you can just get the knowledge from Google! But most people already had access to an immense amount of knowledge before Google, in the form of their public library. You certainly can give yourself quite a self-education with a library card, but the plain reality is that almost no one actually does. Most people aren’t busy little self-starters who will diligently learn on their own. That’s why schools exist, because people need someone looking over their shoulder to force them to learn the material! And even then it often doesn’t work. Most people resist being educated, and the assumption otherwise is part of why policy discussions about education are so unhelpful.

That’s why I call it the Library Card Fallacy: if it was true that education was about access to information, then anyone with a library card would become educated. But that’s just not what education is about. Education is about being challenged to learn things you don’t particularly want to and about creating an incentive structure that forces you to do so. The much-ballyhooed prediction that Google would create a nation of busy little autodidacts has clearly not come to pass. Of course it hasn’t! Most people aren’t Googling “explain the factors that led to World War I,” they’re Googling “Sydney Sweeney nude” or “Batman torrent” or “fantasy football rankings.” Some people love to learn; many, many, many more love to waste time with trivial bullshit. This is why, for example, the famous NBER study that distributed PCs randomly to homes showed no sign of educational gains for the kids whose families received one. Those kids weren’t reading Wikipedia entries! They were playing Farmville on those computers! Sometimes I wonder if these big-think types have ever met an actual child. And the same thing goes for our 18-25 year olds - how many of them, honestly, do you think are going to be sitting there having Gemini come up with a lesson plan to learn about something they find boring? That is not how human beings function.

Remember the MOOC hysteria of the early 2010s? When MOOCs (massively open online courses) burst onto the scene around 2010 or so, the rhetoric was loud and utopian. The New York Times declared 2012 “The Year of the MOOC,” with excitable types insisting that such courses were a “revolution” and a “tsunami.” Time ran a cover declaring “College is Dead.” Sebastian Thrun, a MOOC evangelist who ran a Stanford AI course that had 160,000 students, predicted that within fifty years there would be only ten universities left on earth. The premise was simple: take elite lectures, put them online for free, and you would dissolve the geographic and financial barriers that had kept higher education out of reach for the global poor. As this piece in Science put it, MOOC advocates and their excitable fans in the media “imagined a disruptive transformation in postsecondary education” where video lectures from the world’s best professors “could be broadcast to the farthest reaches of the networked world, and students could demonstrate proficiency using innovative computer-graded assessments, even in places with limited access to traditional education.” This was all based on the Library Card Fallacy - access to information was everything, or so the utopians/Cassandras thought, MOOCs provided it, so college was done for.

The data has been merciless in correcting that fantasy. Across hundreds of courses, completion rates always hovered in the single digits to low teens and stubbornly refused to budge. Katy Jordan’s meta-analysis of 221 MOOCs found median completion rates of just 12.6%, while Hew and Cheung found that MOOCs were typically completed by only 10-20% of students. That Science analysis I mentioned above looked at six years HarvardX and MITx data and was even more damning: “the vast majority of MOOC learners never return after their first year” and “the bane of MOOCs—low completion rates—has not improved over 6 years.” All of this, despite enormous hype and immense investment in pedagogy, nudges, gamification, and platform design. Defenders sometimes argued that many enrollees never intended to finish their courses. This seems straightforwardly damning to me in and of itself, but even we’re inclined to be generous, the reality is grim. This EDUCAUSE research controlled for student intent and still found that among students who explicitly intended to complete a course, only 22 percent actually did so. In other words, even when you filter the sample down to people who said they wanted to finish, almost four in five failed to do so. The technology was there; the lectures were free; access was granted. What was missing the sustained desire to grind through twelve weeks of problem sets when nothing external was forcing the issue.

you don’t say

MOOCs, in other words, turned out to be incapable of supplying what their students needed the most: the capacity for self-regulated learning, realistic goal setting, real interest, perseverance, time management….. These are, for the record, precisely the traits that already correlate with class, prior education, and stable life circumstances. And the people who have those things, whether they’re privileged or not? Yeah, they’re very unlikely to want to skip college and grad school in the first place. The people who are best equipped to be autodidacts are very often the people who have the least interest in dropping out of school and the least to gain by doing so. A free lecture can’t teach a student who won’t watch it, and the students least likely to watch are exactly the ones the technology was sold as saving. This is what that Khan Academy’s Sal Khan, quoted in the piece excerpted in that image, just cannot seem to wrap his mind around: you can lead a horse to water but you can’t make them drink. The sunny, supposedly egalitarian vision of a world full of people hungry to learn just doesn’t fit the reality. Look around you. How many people are spending their free time learning? And even among the people who are, how many of them are learning things that are genuinely boring and frustrating to learn, instead of what’s fun to learn?

Now we’re watching the cycle repeat with the ChatGPT, and somehow with even more bravado than last time. The people predicting that ChatGPT will achieve in 2030 what Coursera couldn’t achieve in 2015 are wrong in the exact same way and for the exact same reasons. They’re confused about what education supplies; they think it’s a matter of access to information, which has been ample for some time, when it’s really a matter of institutional accountability, incentives, and personal inspiration. And they’ve ignored the demand side problem, which has always been the binding constraint. An LLM that can patiently walk you through the causes of the Thirty Year War doesn’t matter if almost nobody wants to be walked through the causes of the Thirty Year War. The marginal student who wouldn’t crack open a textbook at school won’t bother to type a smart LLM prompt, either… and in fact will happily type a prompt asking the bot to write the paper for him, which is the use case actually playing out in every classroom in America right now. Indeed, if LLMs prove anything, it’s how widespread the desire to cheat and cut corners really is; that’s not a condition conducive to autodidacticism. Belief in MOOCs presumed a belief in student willingness to work. The LLM era is, if anything, a regression, a technology sold as the engine of unprecedented self-education that in practice serves as a tool for unprecedented evasion of it. Anyone who’s spent five minutes around an actual teenager could have predicted this outcome.

I don’t know how many American colleges and universities will exist in ten years. Probably fewer than now, but then a little right-sizing has made sense for awhile, and would likely increase rather than decrease the health of the system. The ones that keep existing, which is to say most of them, will go on doing what they’ve always done, which is to supply the external scaffolding that the vast majority of human beings require in order to learn anything they don’t already want to learn: deadlines, grades, embarrassment in front of peers, the looming presence of a teacher who will notice…. That scaffolding is the product and always has been. The lectures are incidental, the textbooks are incidental, and the personalized AI tutor will turn out to be incidental too. What is not incidental is the social and institutional pressure that compels an ordinary late adolescent to sit in a room and slog through the Federalist Papers when every fiber of their being would rather be doing anything else. Maybe we can’t make young people feel that pressure in a meaningful way anymore. Maybe. But that just means that our whole society is doomed anyway, and ChatGPT is not going to be able to fix it.

No chatbot can manufacture the desire to learn. And the people who insist otherwise will, a decade from now, write the same essays they’re writing today about how this time the revolution is really, finally, coming. Damp continuity, like I said. I’ve never been the doomer people have made me out to be, but I confess that in the last couple of years I’ve quietly given up, and if LLMs have done one thing for me, it’s to force me to recognize just how little the average person gives a shit and just how willing the great mass of humanity is to slip into apathy and decline. But I do have hope for individuals, the exceptional and talented people who really give a shit. For them, the ones who need it least, the ability to learn is there. The library card has been in our collective wallet for a hundred years. The whole internet has been in our pockets for fifteen. So go learn something.