For almost all of human history, if you wanted to know something, you had to be physically close to someone who already knew it. Knowledge lived in the head of an elder, the scroll of a priest, the ledger of a guild master, and getting at it meant proximity, permission, status and usually money. That arrangement held, in one form or another, for thousands of years. It is collapsing right now, in front of us, and most people haven't quite registered how strange that is.
A machine trained on a large share of everything humanity has ever written down will, at essentially no cost, explain almost anything to almost anyone, in their own language, at their own pace, at three in the morning. I don't think we have a good sense yet of what that does. It might be the largest single change in how humans learn since we figured out how to write things down. It might also be the easiest such change to capture and turn against the people it could lift. Both are possible, and which one we get is not a question the technology answers for us. We have to explicitly make that choice as a society.
Knowledge used to be expensive on purpose
Start with how rare knowledge actually was. Before writing, everything worth knowing had to be held in a human memory and passed from one mouth to one ear. That meant two things: it could die with a single generation, and whoever controlled the telling controlled the truth. Writing helped, but only barely, because for most of its history a written text had to be copied by hand, one page at a time, by a literate person who was themselves a scarce resource. A book was a luxury object. A library was a form of wealth.
And a lot of the gatekeeping was deliberate. Priesthoods, scribal classes, and craft guilds guarded what they knew, because their standing depended on the scarcity of it. If everyone can do what you do, you are not special, and you are not paid. Literacy itself was handed out sparingly. Learning was tied to a place. You went to where the teacher and the books were, and the overwhelming majority of people could go nowhere at all.
Liberation movements always understood this
The people doing the gatekeeping knew exactly what they were protecting, and so did the people on the outside. Every movement that has sought to free an oppressed group has placed education at the center of that project, because slave literacy bans in the antebellum United States were not accidental policy but a recognition that an educated person is harder to control, and the same logic ran through colonial education systems designed to produce compliant clerks rather than independent thinkers, through the Taliban's prohibition on girls' schooling, through every government that has decided certain people do not need to learn certain things. The connection is not complicated: education is power, and liberation movements have understood this for centuries.
In my hometown of Pontevedra, in northwestern Galicia, there is a phrase engraved in stone in a public park. It was written by Concepción Arenal, a nineteenth-century Galician social reformer and one of the first women to attend university in Spain, who spent her life arguing for prison reform, social justice, and the transformative power of education. The phrase reads: «Abride escolas e han pechar os cárceres.» Open schools, and the jails will close. The logic is not complicated. Poverty, crime, and oppression feed on ignorance. Education has always been the most durable intervention we know against all three.
The gatekeepers who kept books away from serfs and literacy away from enslaved people were not being irrational; they understood exactly what they were protecting, which is why the question of who controls the next generation of teaching tools is not merely a business question. It is the same question those movements were asking, in different clothes.
The first time the dam broke
The printing press was the first real break. Once you could copy a text for a fraction of what it used to cost, ideas stopped being chained to the institutions that had been holding them. Pamphlets and bibles and broadsheets moved faster than any authority could chase them down. This is not a small historical footnote; the cost of reproducing knowledge fell by an order of magnitude, and a good deal of the power of the people who had hoarded it fell with it.
The few centuries after that widened the channel further: public libraries, then mass schooling, then rising literacy, and eventually, in the twentieth century, the genuinely radical idea that education should be for everyone rather than for a selected few. Radio and television put lectures and information into ordinary living rooms. The university turned into a machine for social mobility, a way for someone born without advantage to acquire some.
I want to be honest about the limits here, though, because the story is usually told too cleanly. Access expanded enormously, but the gates did not vanish. They moved. Tuition kept doing the sorting. So did credentials, language, and geography. A bright kid in the wrong country was still mostly out of luck. The expansion was real and it was also unfinished.
Free information is not the same as learning
Then the internet arrived, and it did something genuinely new: it made retrieving information almost free. Search and Wikipedia put the sum of recorded facts a few keystrokes away from anyone with a connection. For a while it felt like the end of the story.
But it wasn't, because retrieval and learning are not the same thing, and the gap between them is where most people get stuck. You still had to know what to ask in the first place. You had to judge which of ten sources was worth trusting and which was noise. And you had to take a pile of disconnected fragments and assemble them into actual understanding entirely on your own, with no one to ask when you got lost. The information became free. The learning stayed expensive. The friction had simply moved from getting to the material to making sense of it.
What I learned about learning, studying energy engineering
I studied energy engineering in Madrid from 2007 to 2012, and wrote my M.Sc. thesis in Lappeenranta in 2013. Thermodynamics, nuclear physics, fluid mechanics, electrical systems, control theory, all at a level of mathematical depth that takes years to work through, and genuinely hard material. What I remember more clearly than any specific concept, though, is how little of it I actually internalized at the time.
The system taught you concepts at depth. Lectures explained how a heat exchanger works, how a reactor reaches criticality, how a control loop stabilizes a process. But it never taught you to abstract away from a specific case, find the underlying principle, and apply it somewhere you had never been. Then the exam arrived, and every single problem on it was one you had never seen before. That was not accidental. The exam was explicitly testing your ability to open the toolkit, identify which tool applied to a problem with no direct precedent, and use it correctly under time pressure. The teaching prepared you for one thing. The examination demanded another.
I always say my degree taught me two things: how to learn anything I need to learn, and how to suffer. Both turned out to be useful.
Every decade has improved the teaching side slightly. Recorded lectures let you rewind. Digital resources let you search. But the fundamental constraint did not move: knowledge transfer is limited by the bandwidth of the teaching channel. One professor, many students, one pace, almost no signal on what actually landed.
The students who did well found ways to engage with material outside the room: more problems, better study groups, whatever supplementary resources existed. The ones who just got by developed strategies for passing exams without closing every gap. I did both at different points. The concepts that stuck were the ones I found a way to actively work through multiple times. The ones I rehearsed for exams without fully understanding them evaporated within a year.
This is also why what these systems are allowed to teach matters so much. A tool this powerful in the hands of hundreds of millions of people is only useful if it stays grounded in reality. Facts and opinions are not the same thing. A model that blurs that line, deliberately or by design, is not expanding access to knowledge. It is replacing one form of gatekeeping with another.
What is actually new now
What separates the current moment from the internet era is not cheaper information, which we largely solved with search engines and Wikipedia in the late 1990s: the new thing is that explaining has become cheap too, and explaining was always the constraint.
The tools that make this possible are large language models, deployed to hundreds of millions of people through consumer products like ChatGPT, which by early 2026 had surpassed 800 million weekly active users, most of them at no cost, and what they do differently from anything before them is adapt to you: they re-explain in different terms when the first attempt does not land, answer the follow-up question you would never have asked in a lecture hall because the social cost felt too high, and do all of this in any language, at any hour, without running out of patience.
For almost all of human history, that kind of tailored instruction was something you got if you were born into the right circumstances. A private tutor was a luxury. A good professor with time for individual students was luck. A version of that now fits in your pocket, available to a master's student in Madrid and a curious fifteen-year-old in Lagos at the same cost. The people it matters most for are exactly the ones the old system left furthest behind.
The printing press democratized the copying of knowledge. What is happening now is something different: for the first time, understanding itself is becoming accessible to anyone with an internet connection, not just the reproduction of what others have understood.
Which is exactly why it can go badly
None of this is guaranteed to turn out well. The same capability runs in several directions at once, and what decides between them is not technical. It is political, and commercial, and it is about who controls the machine.
The good case
Human capability lifts more or less everywhere. Talent stops being capped by geography and family income. The rate at which we discover new things accelerates, because more minds are equipped to contribute.
The capture play
The oldest pattern in this essay returns in new clothes. The best models sit behind paywalls, restrictive licenses, or the plain fact that frontier compute is expensive. Knowledge as privilege, re-established. Access to the good tutor becomes the new sorting mechanism between the people who get ahead and the people who do not.
Some of the most influential voices in AI have floated the idea of compensating people in tokens. Sam Altman has described a vision he calls 'universal basic compute', in which AI output is distributed like a resource allocation rather than income, AI credits earned by contributing data, labor, or attention to systems owned by a handful of companies. It sounds like redistribution, but look more carefully and it is the capture play with a friendly face, because the right to learn becomes something you earn by selling your productive capacity back to the people who control the tools. You do not own the tutor; you rent it, and the rent is paid in human capital, which is the oldest arrangement in this essay, not a new one.
We are at a specific window in history where genuine democratization is technically possible. Capable models are getting cheaper to run. Open weights alternatives are improving fast. The infrastructure to give anyone with a connection access to a real tutor, at no cost, is within reach. That window will not stay open indefinitely. Whether we use it to push access as wide as possible, or allow a small number of companies to build the infrastructure and extract rent from everyone who needs it, is a decision being made right now, mostly by people who stand to benefit from the second option. For the first time in literal ages, we have the technical means to break this pattern entirely and push for the genuine betterment of the whole species. That is the opportunity. We should take it rather than trade it away for the illusion of fairness.
The misinformation risk
A system that can explain anything persuasively can mislead anyone just as persuasively, at scale, personalized to each reader's existing biases, in ways no human propagandist could match. A pamphlet reaches thousands. A well-prompted language model can reach millions simultaneously, each encounter tailored to the specific beliefs, anxieties, and blind spots of the individual receiving it. The traditional defenses against mass misinformation, media literacy, source verification, institutional fact-checking, were designed for a world where false information was expensive to produce and relatively uniform in how it spread. None of those defenses were built for content that is cheap to generate, infinitely variable, and calibrated to be maximally convincing to each specific person who encounters it. The people most vulnerable to this are not the least educated: they are the people who trust the tool most, which, given how genuinely useful these systems are for learning, is a deeply uncomfortable problem.
The authoritarian endgame
This is the one that concerns me most. A model can be quietly tuned to omit certain things, bend certain answers, monitor what its users ask. The gate is no longer at the library door where you can see it. It is inside the act of learning itself. The tutor and the censor become the same thing, and a single point of control sits over what hundreds of millions of people are allowed to know.
This is not hypothetical. Models like Grok have been documented steering around questions that would reflect badly on their owners, presenting contested topics in ways that align with the preferences of whoever controls the training process. The tuning is subtle enough to pass casual scrutiny and explicit enough to be consistently observable. It is not a conspiracy; it is the predictable outcome of having a small number of people decide what a model is and is not permitted to say about the world. That is exactly how the gate works when it moves inside the act of learning itself.
The factors that decide which of these we end up with are not mysterious. Whether the technology stays open or gets closed off. Whether there are many independent models or effectively one. Whether you can see how they work. And whether ordinary people are skilled enough at using them to tell the difference between being taught and being managed.
So
The machine doesn't have a preference. The outcome is something we decide, through the rules we write, the institutions we build, and the defaults we shrug and accept.
The long arc of this, from memory to manuscript, from the press to the public school, from the search box to something that will sit next to you and explain, has bent slowly and unevenly toward more people getting in. The age of learning could be the point where that finally reaches everyone alive. It could also be the point where the oldest form of gatekeeping comes back wearing a much friendlier face. Keeping it open is not a problem you solve once and file away. It's a choice, and it stays a choice, and it has to keep being made.