Your Brain Wasn't Built to Read

Here's a wild fact to drop at your next dinner party: reading is younger than agriculture. Writing systems only showed up around 5,000 years ago — which, in evolutionary terms, is roughly yesterday. Your brain had absolutely no time to evolve dedicated hardware for it.

That matters more than it sounds.

Speech? Hardwired. Babies raised in silence don't spontaneously learn to read, but they will babble — vocal language appears to be part of the standard-issue cognitive package. Reading is not. No child in human history has ever spontaneously learned to read without instruction. Not one.

This distinction is at the heart of something AI discourse almost entirely ignores: when an LLM "reads," it is doing something categorically different from what a seven-year-old struggling through a phonics worksheet is doing. And that difference is far more revealing than most people realize.

The Brain Remakes Itself — Slowly, Effortfully

Because the brain didn't evolve a reading organ, it has to borrow one. Neuroimaging research has shown that proficient readers develop a specialized region in the left occipital-temporal cortex — sometimes called the visual word form area — that becomes exquisitely tuned to the shapes of letters and words. In non-readers, that region is doing something else entirely. Learning to read literally repurposes existing neural real estate.

This repurposing doesn't happen passively. It requires months — typically years — of effortful instruction. And the most powerful predictor of reading success isn't vocabulary or general intelligence. It's phonological awareness: the ability to hear that "cat" is made of three sounds, that "bat" and "cat" rhyme, that you can strip the /k/ from "cat" and get "at."

That might sound trivial. It is not. For a child, recognizing that the written letter "c" maps onto a specific sound in your mouth — the back-of-throat click of a velar stop — is an act of translation that requires understanding spoken language has an internal structure you can consciously manipulate. It's deeply, irreducibly active.

This is why phonics instruction works, and why skipping it is a disaster for struggling readers. And it's why dyslexia — which affects roughly 1 in 5 children to varying degrees — is at its core a phonological processing difficulty. The letters look fine. The problem is in the sound-to-symbol mapping.

What LLMs Actually Do With Text

Here's where I want to push back gently on a framing I see everywhere: the idea that LLMs "read" and "understand" text.

LLMs receive text as tokens — numerical IDs for words or subwords. There is no phonological stage. No sound. No mouth. No laborious mapping of /b/ onto the shape "b." The model has never "sounded out" a word in its life. It went straight to the end-state representation — the mature, pattern-matched encoding of text — without any of the bootstrapping process that creates meaning in a human reader.

A fascinating 2025 study found that the brain's temporal hierarchy of language processing — fast phonological responses in auditory cortex, slower semantic and syntactic integration in frontal regions — maps remarkably well onto the layered architecture of large language models. Lower LLM layers correspond to the early, fast neural responses; higher layers match the slower, higher-order processing (Toneva et al., 2025). It's a striking result that has gotten a lot of attention.

But here's what gives me pause: that correspondence describes the endpoint, not the journey. It tells us that mature, proficient language processing in the brain and in a transformer share some computational structure. What it doesn't tell us is that they got there the same way — or that the LLM's representations mean the same thing.

The Grounding Problem, Wearing Reading Glasses

Dove et al. (2024) make what I think is an underappreciated point about LLMs and meaning. They introduce the concept of "symbol ungrounding" — arguing that while LLMs demonstrate impressive semantic behavior (analogical reasoning, common-sense inference, semantic similarity), they consistently fail when tasks require perceptual binding, embodied simulation, or affordance reasoning. Great at statistical patterns in language; they come apart when you need them to simulate what it feels like to interact with the physical world.

For reading, this matters in a subtle way. A child learning to read isn't just learning statistical co-occurrences between letters. They're connecting those letters to sounds that live in their throat and mouth, to words that mean things they've touched and tasted and been scared by. The symbol "fire" eventually connects — via a long, effortful developmental chain — to an actual felt sense of heat and danger. That chain starts with phonology: hearing the word before ever seeing it written.

LLMs have no such chain. Their symbols are ungrounded by design.

Active Work, Not Passive Intake

Friston et al. (2024) put it starkly in a paper drawing on Karl Friston's Active Inference framework. The core argument is deceptively simple: human learning is active. We don't absorb the world; we probe it. We minimize surprise through both perception and action. LLMs, trained on static text corpora, cannot do this — they're learning from a world that has already been pre-digested into language, with all the sensorimotor scaffolding stripped out.

Learning to read is one of the clearest examples of this distinction. Watching a child learn to read, you see constant action: lips moving while sounding out words, a finger tracing lines of text, the physical effort of holding sounds in working memory while decoding the next letter. This is not passive information intake. It's more like a workout.

And this active process is what creates the durable, transferable skill. A child who has genuinely mastered phonological decoding can read a word they've never seen before. They can handle new proper nouns, invented words in fantasy novels, text in a dialect they've barely encountered. The generalization is robust precisely because the underlying skill is abstract — it's about the relationship between sounds and symbols, not memorized patterns.

That's the thing about reading: its difficulty is a feature. The cognitive work required to build the skill is what makes the skill so powerful.

For Parents Watching a Kid Struggle

If any of this sounds familiar because you have a child in the thick of learning to read — hang in there. The research strongly supports explicit, systematic phonics instruction. The effortfulness isn't a sign something is wrong; it's the process working.

If your child is struggling significantly with phonological tasks — rhyming, blending sounds, identifying first and last sounds in words — it's worth raising with their teacher or a reading specialist. Early support for phonological processing difficulties makes an enormous difference. (A reading specialist or educational psychologist can help sort out whether more targeted intervention would help.)

The Takeaway

Friston et al. (2024) are careful to note that generative AI can serve as a valuable scaffold within well-designed active learning environments — as a tool that amplifies human exploration rather than replacing it. That framing feels right to me.

But when someone tells you an LLM "reads" text, a reasonable response is: sort of, and also not really, depending on what you mean by "reads." If you mean it processes sequential text and extracts structured representations — yes, impressively so, and the neural architecture analogy is real (Toneva et al., 2025). If you mean it went through the laborious, biologically expensive, phonologically grounded, embodied bootstrapping process that turns marks on a page into meaning — no. It absolutely did not.

A child taking three weeks to master the word "the" is doing something a trillion-parameter model has never done and, by design, never will.

That's not a criticism of LLMs. It's a description of a real difference. And if you want to understand intelligence — biological or artificial — the distinctions are exactly where you have to start.