What does it mean to understand in the AI era?

Around 300 BCE, King Ptolemy I of Egypt grew impatient with Euclid’s Elements — a dense, demanding collection of books that founded geometry — and asked its author whether there was some shorter path to mastering the subject than working through the whole thing. Euclid’s reply, recorded by the philosopher Proclus, has survived twenty-three centuries for a reason:

“There is no royal road to understanding.”¹

A king could command armies, levy taxes, and have a faster road built wherever he pleased, but he could not have understanding delivered. The work of grasping the material was his to do, and no amount of power or shortcut could do it for him. Twenty-three centuries later, the emergence of AI has made Euclid’s words even more relevant. Knowledge is still one of those rare, extremely valuable assets that cannot be bought; it can only be earned.

There is a particular feeling that arrives at the end of reading a well-written paragraph. The pieces click. You make the connections. You move on, with the sense that you understood.

Sometimes that feeling is earned. Often it isn’t. And telling the two apart has quietly become one of the more important skills nowadays — because for the first time in history, you can use the AI tools to summon a confident, articulate, well-organised explanation of nearly anything, instantly, for free. The explanation will feel like understanding. The question this post examines is whether it actually is.

To answer that, we need to be precise about a word we usually leave vague. What is understanding, actually — as a thing that happens in a mind? Not the dictionary definition, but the mechanism. Because once you can see the mechanism, two things become obvious: why genuine understanding is so valuable, and why the tools that hand you fluent explanations can either accelerate it or quietly hollow it out, depending entirely on how you use them.

Reading is not the transfer of knowledge

Here’s the intuitive, simplistic picture most of us have about reading: a text contains information; reading moves that information into your head; understanding is what you have when the transfer is complete. The book is a container, your mind is a container, and reading is the process of moving the information from one container to the other.

This picture is wrong, and the way it’s wrong is the whole story.

Consider a sentence that became famous in early machine translation for breaking parsers: “Time flies like an arrow.” You read it as a metaphor about time passing. But pair it with “fruit flies like a banana” and suddenly you realise the first sentence had other meanings all along. Maybe there is a species of insect, the “time fly”, with a peculiar fondness for projectiles. Maybe “time” is a command: go time some flies, the way you’d time a runner. And notice what the second sentence depends on: if you know that fruit flies are insects, you read it as “fruit flies are fond of bananas.” But if you’d never heard of a fruit fly, the only reading left is that a piece of fruit sails through the air the way a banana would. Same five words, but which meaning you acquire is decided entirely by what you already know. The words on the page didn’t change. What changed is which meaning your knowledge of the world allowed to emerge. This is exactly the phenomenon the cognitive psychologist Walter Kintsch spent his career modelling. Meaning isn’t read off the page; it is assembled by the reader.

That is the key move, and the container picture misses it entirely. The meaning was never in the text. It was constructed, in your head, out of a collaboration between the words on the page and the vast store of things you already know. Reading isn’t reception. It’s construction. The book is not a container of knowledge but a blueprint on how to construct it.

Walter Kintsch (1932–2023) was a German-American cognitive psychologist at the University of Colorado Boulder and one of the founding figures of the modern science of reading comprehension. He spent decades working out how that construction happens, and his account — the Construction–Integration model — remains the most influential model we have for describing what goes on when a person understands a text.²³ You don’t need the technical machinery, but you do need its central insight, because it reframes what it means to read and understand with AI.

The three layers of understanding

The central insight of the Construction–Integration model is that when you read something, your mind builds not one representation but three, stacked on top of each other.

The first is the surface — the literal words, in order. This layer is real but fleeting; it lasts seconds. You cannot recite the exact wording of the sentence before this one, even though you understood it perfectly. The words were a delivery vehicle, and once they delivered, they evaporated.

The second is the textbase — the network of bare facts the text states, stripped of their exact phrasing. Read “The snow was deep on the mountain. The skiers were lost, so they dug a snow cave, which provided them shelter,” and your textbase holds the stated propositions: snow, deep; skiers, lost; they dug a cave; the cave gave shelter. This is what the text said. It is a genuine representation, and crucially, you can have it without really understanding anything.

The third layer is where understanding actually lives. It’s called the situation model, and it’s the integrated mental picture you build by fusing the textbase with everything you already know. The text never tells you why a snow cave provides shelter. But you understand, because you know that compacted snow insulates, that wind lowers temperature faster than cold air, that a windbreak you can breathe in is the difference between surviving a cold night on a mountain and not. None of that is in the text. All of it is in your knowledge. The situation model is the difference between knowing what the text said and grasping the reality it describes.

Someone reads: “heavy things sink and light things float.” They can recite it; that’s a solid textbase. But do they understand it? You find out by asking something the rule never stated: a steel ship weighs thousands of tonnes, far more than a single pebble, so why does the pebble sink while the ship floats? The confident recitation tends to collapse right here. Answering requires a situation model. Sinking does not depend on “how heavy is it?” but on “how much heavier than the water it displaces”, which depends on a hull’s shape, not its mass alone. Reciting needs only the textbase. The gap between the two is the gap between remembering and understanding, and it can be enormous.

The distinction between the textbase and the situation model is worth holding side by side:

	Textbase	Situation model
What it is	The bare facts the text states	The text fused with your prior knowledge
In a word	What the text said	What the text is about
What it gives you	Recall — you can reproduce it	Generativity — you can use it, transfer it, explain it
How durable	Brittle; answers one question	Robust; reachable from many directions
Can you have it without understanding?	Yes	No — this is understanding
Can an AI build it for you?	Yes — it hands you another textbase	No — only you can, from your own knowledge

Why this is an important distinction in learning

Kintsch’s research could be summarized in the following phrase: the things that help you remember a text are not the same as the things that help you learn from it.⁴

You can build a perfectly serviceable textbase while never constructing the situation model that would let you use it. And the two come apart precisely when you need them most. A textbase is brittle: it can answer the question it was built for and little else. A situation model is generative: because it’s woven into your existing knowledge, you can reach the same idea from many directions, apply it to cases you’ve never seen, notice when it’s relevant somewhere unexpected, and explain it to someone else. That generativity is understanding. It’s also exactly what a textbase, however complete, cannot give you.

Two well-documented features of the mind make this distinction sharper still.

The first is what cognitive scientists call the illusion of explanatory depth: the robust, repeatedly demonstrated finding that people believe they understand things far more deeply than they actually do.⁵ Ask someone whether they understand how a bicycle works, or a zip, or why the seasons change, and they’ll say yes with confidence. Ask them to actually explain it, in detail, and the confidence vanishes. The understanding they were sure they had turns out to be just a textbase disguised as a situation model. The illusion is the default state. It only breaks when you try to produce the explanation yourself, and this is something to be aware of when using an AI tool that produces explanations for you.

The second is the role of prior knowledge. Because understanding is the fusion of text with what you already know, the more you know, the more you can understand. This is why an expert reads a paper in their field and sees the argument’s shape, the load-bearing assumption, the move in the third section that everything depends on, while a novice reading the identical words sees a flat list of facts. Kintsch and Anders Ericsson showed that experts effectively escape the normal limits of working memory for material in their domain, because they’ve built rich retrieval structures that let them hold and manipulate far more at once.⁶ The catch is unforgiving: those structures only work for knowledge you’ve actually built yourself. You cannot borrow them. You cannot download them. They are the slow accumulation that understanding both requires and produces.

The answering machine

Now place an AI assistant into this picture, and the tension comes into focus immediately.

What these tools are extraordinarily good at is producing the output of understanding — a clear, fluent, well-organised explanation, a tidy summary, a clean answer to your question. What they cannot do is perform the process of understanding on your behalf, because that process is the construction of a situation model in your head, out of your prior knowledge, and no external text, however good, can do that for you. It can only hand you another textbase.

Read an AI’s summary of a difficult chapter and you can acquire a serviceable textbase in a fraction of the time it would have taken to build one yourself. You’ll know what the chapter said. You may even be able to repeat it. But if you stop there, you have quietly skipped the one step that produces understanding: the effortful integration of that material with what you already know. You’ve acquired the symptom of understanding, a fluent account you can reproduce, without the thing itself. And because of the illusion of explanatory depth, you will very likely feel that you understand, right up until the moment you’re asked to explain the concepts in the chapter and discover that you cannot, because you have not built the generative situation model that will enable you to do so. The summary didn’t deepen your understanding. It substituted for it, and left a convincing replica in its place.

This isn’t a hypothetical worry; it shows up in controlled studies. When researchers had people learn a topic from an AI’s synthesis versus assembling their own understanding from sources, the AI group reliably ended up with superficial knowledge — they invested less, and what they could produce afterwards was shallower and more generic, even when the underlying facts were identical.⁷ In a randomised experiment with hundreds of secondary students, those who studied a passage by writing their own notes comprehended and retained more than those who studied with an AI assistant — even though the students preferred the assistant and found it more helpful.⁸ The preference is the trap. The easier path feels better and teaches less, and the feeling is precisely what makes the trade invisible.

But none of this is an argument against the tools and AI. It’s an argument against one particular way of using them, and just as much, against one particular way of building them. Most of these tools are designed to be helpful in the most immediate sense. Upload a document, open a chapter, get a tidy summary; hand over your reading, get a ready-made deck of flashcards, or a podcast, or a narrated video that walks you through it. That default, the one that hands you the output of understanding, is the one that hollows it out. And it’s no accident: the finished answer is what feels useful in the moment and keeps you coming back. But the way the current AI tools are built is not the way they have to be. A different design could turn the very same capabilities toward making you do the work that actually builds understanding.

Doing the work, or offloading it

The question to consider is the following: which cognitive work should the AI tool do for us?

Cognitive scientists have a name for handing a mental task to an external tool, cognitive offloading.⁹ It is not, in itself, good or bad. The mind has always leaned on external aids, from written notes to calculators, and the right offloading frees it for higher work. The decisive question is which work you hand off. Offload the work you would gain little understanding from anyway and you free capacity for deeper thinking; offload the work that forges the connections, and understanding is what you lose.

The harmful kind is easy to fall into with AI, because the tool is so good at producing the output of understanding. Ask it to make a cheatsheet for a chapter and it hands you a finished account you can read in seconds, but it has done the one thing that would have built your understanding, the effortful construction of a situation model out of your own knowledge. You are left with a textbase you did not build. And the evidence that this quietly erodes comprehension and critical thinking is mounting. Across a study of hundreds of participants, heavier reliance on AI tools predicted weaker critical thinking, with cognitive offloading identified as the mechanism responsible for the effect.¹⁰ The more you let the tool do the thinking, the less thinking you do.

The beneficial kind is the mirror image. Ask the same AI to draw the graph of a function, to search through hundreds of pages of text no person could hold in working memory at once, or to list all the main concepts of a book, and it offloads exactly the labor your mind genuinely cannot perform, handing back something you can now think with. Here the tool is not replacing your understanding but expanding your cognitive capabilities, doing the heavy lifting so that the construction work, the part that is yours, becomes feasible.

So the same instrument cuts both ways, and the research on what actually builds understanding points to how to stay on the right side of it. The findings are consistent, and they convey a clear point. Learning deepens as you move from passively receiving information, to actively manipulating it, to constructively generating something new from it, an explanation, a question, a connection, to interactively batting ideas back and forth with a responsive partner.¹¹ The deep end of that scale is exactly where understanding gets built, and it is defined by how much you are generating versus how much you are absorbing.

Map AI use onto the ICAP framework and the division between harmful and beneficial task offloading becomes clear. Reading the summary generated by the AI tutor instead of the text sits at the passive end: it generates, you absorb, and the very effort that would have constructed your situation model is removed.

But ask the AI to quiz you and explain your answers, and now you are interacting, which is a powerful learning operation. Ask it to take your explanation and find the hole in it. Hand it your own summary of the argument and have it show you what you missed. Make it ask you questions instead of just answering yours. Have it act as the confused student while you teach, because the moment you try to explain, the illusion of explanatory depth shatters and you see exactly where your understanding is real and where it is a textbase in disguise. Use it to highlight the main assumptions of the argument, the connection between the concepts of two chapters, the reason the author needed this lemma before that theorem. Every one of these keeps the construction work where it belongs: in your head. The tool is not climbing the mountain for you. It is a companion on the climb, pointing out a handhold, asking if you are sure about that route, catching you when you slip.

The same instrument, two opposite effects. Used to deliver finished understanding, it strands you in the shallows with just a convincing feeling and nothing under it. Used to provoke and test and partner your thinking, it pushes you toward the deep end faster than you could reach alone. The technology is identical. The only variable is whether it is doing the work or making you do it.

Euclid is still right

AI is genuinely a new, powerful tool. Tasks that were once possible only for a human, slow and cumbersome, are now instant. At the same time, we have inherited the accumulated knowledge of centuries, written down in books and papers of a depth and quality no chatbot can match. For the first time, these two things can meet. The deep, structured content of a good book, and an assistant that can help you work through it, question you on it, and clear the obstacles between you and the knowledge it conveys. That combination is a real opportunity, arguably one of the greatest tools for learning ever assembled.

But only if the assistant is doing the right work. Everything in the sections above was about the fault line that runs through this opportunity. Hand the AI the tasks that clear your path and it extends your reach; hand it the tasks that build understanding and it quietly does your thinking for you.

And that fault line rests on a truth none of these tools can change: understanding was never the explanation. It was, and remains, the situation model you build in your own mind — generative, woven into what you already know, yours. That has not become one bit easier to acquire. If anything, it now requires a new discipline to properly use these tools to enhance your understanding.

This is the area this series of posts will explore. What understanding really is, and how to tell the real thing from its illusion. Why summaries seduce, and what to do instead. How to turn an AI from an oracle that answers into an interactive companion. And how the right design, one that blends these new capabilities with the depth of good books, can build understanding rather than quietly replace it.

The premise underneath all of it is simple. The best of reading has always happened in your head — the questions, the connections, the model of the thing taking shape as you go. That part is still the point. The work of understanding is still yours to do. The right tools don’t do it for you. They just make sure you don’t have to do it alone.

---

This is the first in a series on understanding, comprehension, and learning with AI. Next we turn to the evidence, a survey of the research putting AI reading and learning tools to the test, and what it reveals about when they build understanding and when they erode it.

References

Footnotes

1. Paraphrased. Euclid’s actual words, as recorded by Proclus, were “there is no royal road to geometry.” But to the Greeks, geometry — in the rigorous, demonstrative form Euclid gave it in the Elements — stood as the very model of supreme knowledge; Plato’s Academy reputedly bore the inscription “let no one ignorant of geometry enter.” Read in that light, “no royal road to geometry” was, in effect, “no royal road to understanding,” which is how the line is rendered here. ↩

2. Walter Kintsch, Comprehension: A Paradigm for Cognition (Cambridge University Press, 1998). ↩

3. Walter Kintsch, “The role of knowledge in discourse comprehension: A construction–integration model,” Psychological Review 95, no. 2 (1988): 163–182. ↩

4. Walter Kintsch, “Text comprehension, memory, and learning,” American Psychologist 49, no. 4 (1994): 294–303. ↩

5. Leonid Rozenblit and Frank Keil, “The misunderstood limits of folk science: an illusion of explanatory depth,” Cognitive Science 26, no. 5 (2002): 521–562. ↩

6. K. Anders Ericsson and Walter Kintsch, “Long-term working memory,” Psychological Review 102, no. 2 (1995): 211–245. ↩

7. Shiri Melumad and Jin Ho Yun, “Experimental evidence of the effects of large language models versus web search on depth of learning,” PNAS Nexus 4, no. 10 (2025). ↩

8. P. Kreijkes et al., “Effects of LLM use and note-taking on reading comprehension and memory: A randomised experiment in secondary schools,” Computers & Education 243 (2026). ↩

9. Evan F. Risko and Sam J. Gilbert, “Cognitive Offloading,” Trends in Cognitive Sciences 20, no. 9 (2016): 676–688. ↩

10. Michael Gerlich, “AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking,” Societies 15, no. 1 (2025): 6. ↩

11. Michelene T. H. Chi and Ruth Wylie, “The ICAP Framework: Linking Cognitive Engagement to Active Learning Outcomes,” Educational Psychologist 49, no. 4 (2014): 219–243. ↩