The mind as limited space. Knowledge does take up room

The head is a container with walls, and it always was. But now that same container has books, the press, radio, television, cinema, video, the internet, social media, streaming, the phone and, on top of all that, AI dumped into it. Miller's figure hasn't budged since 1956. What has changed is the hose. And the kids born this year are going to grow up with that hose left open from the factory.

The favourite line of whoever designs curricula

There's a line that gets repeated at parents' meetings, in end-of-year speeches and in the brochures of English academies aimed at the over-fifties. The line is "knowledge doesn't take up room". It sounds good. It justifies adding subjects, languages, workshops, after-school clubs, micro-credentials, languages again.

It's false.

Not false in the soft sense of "open to nuance", but false in the measured, experimental, lab-repeated-since-1956 sense. The human head is a container with walls. The walls are close. And we know where they are.

The figure that comes out of the lab

George Miller published that year, in Psychological Review, a paper titled The Magical Number Seven, Plus or Minus Two. The figure came from asking subjects to hold discrete items —digits, syllables, positions— and reproduce them without error. Working memory, that space where the information being processed at this very instant is manipulated, holds seven items give or take two, for seconds, and fewer still when the items have to be combined rather than just repeated. Nelson Cowan, in 2001, reviewed the evidence accumulated since Miller and adjusted the figure downward: under real conditions, without deliberate grouping, the effective limit is closer to four items than to seven. On those limits John Sweller built, in 1988, cognitive load (carga cognitiva): any mental task competes for the same container. Kahneman, from another angle, called the deliberative machinery System 2 and described it as a battery that runs down.

What matters isn't the exact figure, which gets argued over every twenty years. What matters is that the container is fixed, and that nobody is born with a bigger one for having started using TikTok earlier.

One life, the same walls, more and more to cram inside

We have a single life. Seventy-five, eighty, ninety years with luck. Inside that life fits whatever fits in a human head whose processing capacity hasn't changed in the last ten thousand years. What has changed, and a lot, is the amount of stuff expected to go inside.

Think about what a medieval peasant had to fit into his head. The seasons, the liturgical calendar, the names of four generations of his village, the catechism, the prices at three nearby markets. That's not nothing, but it fits in a human head without overflowing it. Jump to the nineteenth century: a daily paper, the cheap book, the telegraph. To the twentieth: radio, cinema, television, home video. To the nineties: internet, email, the web. To 2010: social media, smartphone, non-stop streaming. To 2026: generative AI producing text, image, video and voice faster than any human can consume them.

Each leap adds channels without removing the earlier ones. The average boomer still watches the evening news while checking WhatsApp and asking a conversational assistant how many calories there are in a banana. The digital native adds Discord, Twitch, Reddit, TikTok, Spotify, Notion. The kid starting school this September is going to grow up talking to AI systems before learning to read fluently.

The human receiver's cognitive capacity, meanwhile, is still four to seven items over twenty seconds. The same figure as the peasant. The same figure as Cervantes. The same figure as your grandmother.

Who decides what stays inside

And here comes the part almost nobody mentions. We don't prioritise what gets remembered. Nor how it gets remembered. The system does that on its own, and it does it badly, by criteria that look nothing like the ones the adult subject would state if you asked.

What sticks is what gets repeated, what carries emotional charge, what gets processed several times in different ways, what's connected to prior schemas. What isn't, falls away. Doesn't matter that it's important. Doesn't matter that you read it underlining every line. Doesn't matter that you'll need it tomorrow. The system doesn't consult you.

That's why a teenager remembers, with surgical detail, the lyrics to two hundred reggaeton songs and not the periodic table, even after putting more formal hours into the second. The songs have repetition, rhythm, emotional charge, a social context. The periodic table has a single ordered exposure and a test at the end. The container is no democracy. It rewards what fires more circuits, not what the syllabus decreed a priority. Multiply this by today's flow. The human receiver is being bombarded by channels that have spent twenty years of A/B testing optimising precisely the levers the cognitive system uses to fix memory: repetition, emotion, novelty, intermittent reward. The algorithms know better than you do what's going to stay in your head this afternoon. And what stays isn't what you'd choose if they handed you the controls.

The boomer falling behind. And the flip side.

Past a certain age it starts to show. Those of us around sixty or over admit without drama that it's harder to get new information in. Learning a piece of software from scratch, remembering passwords, holding three new faces in a meeting. It's not romantic decline: it's that the container, besides being fixed, has spent decades working with a very dense repertoire of prior schemas, and getting in information that hooks onto none of them takes more raw effort than it did at twenty.

Here the easy narrative collapses on its own. The comfortable reading would be: poor boomers, victims of accelerated change. Not exactly. The same boomer who takes half an hour to set up an app has a density of schemas a teenager won't have for another forty years, if they build them at all. He spots a liar by the cadence of three sentences. He can read an invoice. He catches a badly built argument on the first pass. His container isn't smaller; it's more selective and has more mileage. Less gets in per unit of time, yes, and it gets in better filtered because he's spent a lifetime building the filters.

The digital native has exactly the opposite problem. Plenty of inflow, few mature filters, schemas under construction, raw capacity intact. Not a victim by default and not the golden generation of tech marketing either. Someone who's getting four or five daily hours of stimuli engineered to fix whatever the business wants fixed, with a cognitive system that hasn't yet consolidated its defences. Who comes out of the film better off depends on which dimension you measure. There's no clear loser.

The kid who'll be born talking to the machine

Let's get to the interesting case. A child born in 2026. At three they'll have a conversational assistant that answers any question with infinite patience. At six AI models will tell them stories, able to bend the plot to their attention span. At twelve they'll dictate schoolwork to a machine that hands it back polished. At sixteen they'll sit university exams with assistants they'll have trained over years.

Their container, biologically, will be the same as yours. Four to seven items. Twenty seconds.

The reasonable question isn't whether they'll be smarter or more stupid. That question is lazy. The question is what is going to get fixed inside that container when all the processing on the outside is delegated to a layer they don't control. If they never have to hold a fifteen-step line of reasoning without help, they don't build the schema that automates those fifteen steps. If they never have to remember a phone number, they don't build the operative memory that holds sequences. If they never have to write a bad draft before the good one, they don't build the skill of crossing out.

No friction, no schema

This is called, in the jargon, cognitive offloading: handing an external system a mental operation you used to do yourself. It's not bad in itself. Libraries, notes and spreadsheets are cognitive offloading, and they haven't produced idiots en masse. The difference with generative AI is friction. Taking notes sustains the question, controls the flow, forces the container to fill and empty deliberately. Generative AI removes that friction almost entirely. The explanation arrives already structured, complete, in a teacherly tone, with no visible gaps. It gives you exactly the sensation of understanding without going through the process that produces understanding. And that difference isn't visible from inside: it shows up months later, when someone asks you to apply cold what you thought you'd learned and you discover you learned nothing.

The 2026 kid is going to have to negotiate when to lean on the external layer and when to process himself, knowing the second costs more and produces something more stable. Nobody is teaching him to make that negotiation. His parents don't know how. His teachers don't either.

Is the education system ready?

No. The short answer is no.

The long one has its nuances. Curricula are still written under the implicit assumption of the cliché that titles this article: that adding doesn't subtract, that the pupil's container is elastic, that the education legislator's job is to decide what to put in without having to decide what to leave out. This was already false with Miller in hand. The New South Wales Department of Education published in 2017 a guide aimed at teachers, Cognitive Load Theory: Research that teachers really need to understand, which translated the evidence into classroom design without sugaring it. Almost nobody outside the department itself read it. With generative AI in the classroom the question is already absurd. If the average pupil can generate three thousand coherent words on any topic in thirty seconds, what's needed isn't teaching him to produce those three thousand words, but to judge whether they're any good and to sustain a line of reasoning of his own when the machine isn't there. Those two competences require things the current syllabus doesn't protect: time without the machine, unproductive effort, deliberate friction, getting it wrong without help.

The education industry, meanwhile, is selling the opposite. Tablets in preschool, AI tutors in primary, adaptive platforms, gamified micro-learning. It all runs in the direction of lowering friction. Lowering friction works beautifully for making the pupil feel competent this afternoon. For making him competent ten years from now, without the tool, it's exactly the wrong road.

Educational classism, old and new

We arrive at the uncomfortable spot. Developed countries are going to have citizens with different cognitive capacities from the less developed ones. Is an educational classism going to appear?

Educational classism already exists. It's existed as long as schools have. It exists between countries, between regions of the same country, between neighbourhoods of the same city, between families on the same street. No sensible person denies it. The interesting question is what AI does to the slope that was already there.

The intuitive answer, the reassuring one, is that AI is a leveller. It puts an expert tutor within reach of anyone with a phone. The labourer's daughter accesses the same assistant as the notary's daughter. Sounds lovely.

The observable reality is dirtier. Families with high cultural capital are using AI to do more and better what they already did: they're teaching their children to use it as a critical tool, to distrust the output, to cross-check, to delegate the mechanical tasks and reserve the hard ones to train themselves on. Families with low cultural capital are using AI, when they use it, to have the kid's homework done for him. Same tool, opposite uses, divergent trajectories. The slope is turning into a step.

The slope between countries

Add the national dimension. Countries with solid education systems are integrating AI into the classroom with some sense: public debate, budget, teacher training. Countries without a solid education system are letting the kid discover ChatGPT on his own on his father's phone. Fifteen years from now that kid will compete with one from Helsinki who spent his adolescence learning when to lean on the machine and when not to. The same fixed container in the head. One has trained the schemas that make the most of it; the other has filled it with fifteen-second videos for a decade.

This isn't fixed by handing out tablets. It might not be fixed at all.

What stays inside

We have one life. We're filling it with channels that multiply and information that accelerates. We don't choose what stays inside us: the system does, by criteria optimised by third parties who aren't us. Those of us who came from before notice it's harder to get new things in, with nobody telling us that's the visible face of some very well-trained filters that have their own advantages. Those coming after us are going to grow up with an intermediate layer of AI between them and almost everything they learn, and what stays inside is going to depend on how much friction the system we build for them allows.

The question isn't whether AI augments our capacities. The question is what exactly it does with the fixed container we already had, and who decides what's left out when the flow exceeds the opening. Anyone offering you a comfortable answer right now is worth a second look.

Definiciones

Working memory: the mental space where the information being processed at a given moment is manipulated. It has a very limited capacity (between four and nine items depending on the task) and a short duration (around twenty seconds without active rehearsal). It isn't the same as short-term memory: working memory doesn't just retain, it also operates on what's retained.

Cognitive load (carga cognitiva): the total amount of mental resources a task demands of working memory. Sweller's Cognitive Load Theory distinguishes three components —intrinsic (the material's own difficulty), extraneous (added by how it's presented) and germane (the kind that builds lasting schemas)— which compete for the same container.

Chunking: grouping several loose items into a larger unit that the system processes as a single item. It lets you handle more information without enlarging the container, at the cost of demanding well-built prior schemas. A chess expert sees a real position as one unit; an amateur sees fifteen loose pieces.

Cognitive offloading: handing an external system —a book, a notebook, a spreadsheet, a generative AI— a mental operation the subject used to do himself. It's a neutral operation in itself; what makes it problematic or virtuous is the degree of friction it keeps and the kind of schemas the subject stops training.

Educational classism: structural inequality in access to cultural capital and to the material conditions that make learning possible. It isn't a new phenomenon, but the arrival of AI in the classroom and the home is stratifying it in new ways, depending on how each family and each education system mediates the use of the tool.

Referencias

Miller, G. A. (1956), The Magical Number Seven, Plus or Minus Two: Some Limits on Our Capacity for Processing Information, Psychological Review 63, pp. 81–97. The original experimental basis for the idea that working memory holds seven items, plus or minus two.

Cowan, N. (2001), The Magical Number 4 in Short-Term Memory: A Reconsideration of Mental Storage Capacity, Behavioral and Brain Sciences 24, pp. 87–114. A review of the post-Miller evidence; argues that effective capacity, without deliberate grouping, hovers around four items. DOI: 10.1017/S0140525X01003922.

Sweller, J. (1988), Cognitive Load During Problem Solving: Effects on Learning, Cognitive Science 12, pp. 257–285. The starting point of Cognitive Load Theory; takes the limits of working memory and builds on them a framework applicable to instructional design.

Kahneman, D. (2011), Thinking, Fast and Slow. The framework of System 1 and System 2, and of the idea that deliberative attention works like a battery draining over the course of the day.

NSW Department of Education (2017), Cognitive Load Theory: Research that teachers really need to understand. An institutional document applying the framework to classroom and curriculum design. Available at https://education.nsw.gov.au/content/dam/main-education/about-us/educational-data/cese/2017-cognitive-load-theory.pdf.

Para profundizar

Sweller, J., van Merriënboer, J. J. G. and Paas, F. (2019). Cognitive Architecture and Instructional Design: 20 Years Later. Educational Psychology Review 31. A two-decades-on review of Cognitive Load Theory. .

Paas, F. and van Merriënboer, J. J. G. (2020). Cognitive-Load Theory: Methods to Manage Working Memory Load in the Learning of Complex Tasks. Current Directions in Psychological Science 29. A synthesis applied to the contemporary design of tasks and interfaces. .

Carr, N. (2010). The Shallows. What the Internet Is Doing to Our Brains. Norton. An early reading of how digital environments reshape attention habits.

Newport, C. (2016). Deep Work; and (2019). Digital Minimalism. Grand Central. A practical application of cognitive limits to contemporary intellectual work and to the hygiene of information channels.

También te interesa

En otros sitios

#ai-memory #cognitive-debt #learning #attention #ai-dependency