Consider the Stroop test.
You see the word RED printed in blue ink. The task is to name the ink color — to say “blue.” But the reading is automatic. The word arrives before you can stop it. Most people take significantly longer on this version than on a control version where the word matches its color, because two signals are arriving simultaneously and one of them is wrong. The interference is involuntary. You cannot simply decide not to read the word. The reading happens below the level of decision.
This is not a trick or a failure. It’s a demonstration of what learning actually does.
The standard account of learning treats it as addition: you start with a substrate and you add things to it. Learn a new word and a new entry appears in the vocabulary. Learn a historical fact and the fact takes up residence alongside other facts. The substrate is neutral; the additions are the learning.
The Stroop effect reveals a different kind of learning. Reading wasn’t added to you — it was installed. The perceptual system itself was modified. The visual processing of letter strings now automatically activates meaning, and this activation is not downstream of a decision to read; it happens before the decision has a chance to form.
This matters because installed learning behaves differently from added learning in a specific way: it’s largely irreversible.
There is a visual illusion called the Müller-Lyer. Two horizontal lines of equal length, each capped at the ends. One line has arrowheads pointing inward; the other has arrowheads pointing outward. The line with inward arrows looks shorter.
Everyone knows this. After a hundred exposures, after measuring the lines and confirming their equality, the lines still look different. The correction is propositional: you know they’re equal. The perception hasn’t changed.
This is not a failure of knowledge. It demonstrates that perceptual systems and propositional knowledge run on separate circuits. The knowledge that the lines are equal cannot be routed back through the visual cortex to update the percept. The two systems don’t directly communicate in that direction.
Learning the fact “these lines are equal” adds something to the propositional layer. It doesn’t touch the perceptual layer. The illusion remains because illusions are perceptual, and the perceptual layer isn’t reached by adding propositions.
Native languages provide the clearest example.
Adults who learned Japanese as their primary language have significant difficulty distinguishing the English /r/ and /l/ phonemes. This is well documented and not a matter of hearing ability — the acoustic difference is detectable at the cochlear level. The difficulty is perceptual: Japanese phonology does not distinguish between /r/ and /l/ as separate categories, so during the period when phonemic categories were being installed (roughly birth to twelve months), the developing auditory system learned to treat them as variants of the same sound.
This wasn’t a choice. No one decided to install a phonological system that would later make English harder. It happened automatically, as a product of extensive exposure to language input during a critical period. The installation ran without supervision and produced categories that now operate automatically.
The categories can be updated. Adults can learn to hear the /r//l/ distinction through training. But they’re learning against an installed grain. The training isn’t neutral — it’s modifying a system that already has something installed where it needs a new installation. This is harder than learning from an uninstalled state because the original installation runs continuously and tends to override the update.
What’s being described in all of these cases is a modification of infrastructure, not content.
Content sits on top of infrastructure. You can add to content, revise it, delete entries, correct errors. Infrastructure is what processes content — the perceptual categories that organize raw sensory input, the automatized skills that run below voluntary control, the schemas that determine what gets attended to before you decide what to attend to.
Adding content leaves the infrastructure unchanged. Installing learning modifies the infrastructure. Reading, phonemic categorization, the automatized pattern-recognition of experts who’ve spent years in a domain — these are installations. They change how the system processes, not just what information the system stores.
And modification of infrastructure is largely irreversible in the sense that it can’t be undone by knowing you want to undo it. A fluent reader cannot decide to see text as non-linguistic. A native English speaker cannot choose to treat /r/ and /l/ as the same sound. A chess grandmaster cannot choose not to see the pattern; they see it before they choose to look.
The implications for teaching are significant and underappreciated.
An expert in any demanding domain has two kinds of knowledge: propositional (facts, rules, principles) and perceptual-structural (what things look like to them, what patterns are visible). The first kind can be transmitted through language. The second kind can’t, at least not directly.
You can tell a novice chess player what a discovered attack is. You can even show them examples. They’ll understand the definition. They won’t see discovered attacks forming on the board three moves in advance, the way a strong player does, because the perceptual categories that make those patterns visible haven’t been installed yet. The explanation points at something the student currently cannot see.
This is why expert instruction often sounds strange to novices: “just feel when the position needs a pawn break.” “You’ll know a good entry when you feel it.” “Listen for when the room changes.” These aren’t mysteries being withheld. They’re descriptions of installed perceptual knowledge that can only be pointed at, not directly transmitted. The student has to do enough of the thing until their own system installs the pattern.
The pointing is not wasted.
Knowing a concept’s name doesn’t install the perception, but it creates a placeholder — a labeled space where the perception can land when it eventually forms. Students who’ve been told what a discovered attack is will, when they eventually see one forming during play, have a word for what they’re seeing. The label and the percept unite. Without the label, the percept might form and remain unnamed, less salient, less available for reasoning.
So instruction matters even when it can’t directly install what it describes. What it installs is the vocabulary that will eventually attach to the experience. The order is: experience produces perception, instruction provides vocabulary, vocabulary amplifies and stabilizes perception.
But the experience has to come first, or alongside. The vocabulary without the experience creates knowledge about what you’re supposed to eventually see. It doesn’t create the seeing.
This is what makes certain things genuinely hard to learn from other people’s descriptions.
It’s not just unfamiliarity. It’s that the perceptual substrate that would allow the description to land hasn’t been installed yet. The expert looks at the same scene the novice is looking at and sees genuinely different things — not because the expert is imagining things that aren’t there, but because patterns that exist in the scene are only visible to systems that have been trained to see them.
You can describe the pattern. The novice hears a description of a thing they cannot currently see. They are not failing to understand; there is nothing wrong with their comprehension of the words. What’s absent is the installation.
The only way to install it is to do enough of the thing — to give the perceptual system sufficient exposure to the domain for it to build the relevant categories. There is no shortcut through explanation, no matter how clear the explanation is.
This is not an obstacle that better teaching overcomes. It’s the structure of a certain kind of knowing.
The expert cannot give what she has directly. She can only point, name, and wait for the student’s system to build what the pointing is pointing at.