First-graders write b where they meant d, and children learning Chinese characters write 日 (rì, "sun") with its parts flipped left to right. Adults file this under carelessness. But carelessness should not be this tidy: the errors are almost always left-right flips, hardly ever upside-down ones.
Three things this time: where in the brain this skill actually sits, why writing looks the way it does, and why one and the same reading disorder shows different symptoms in English, Italian, Chinese and Japanese. The starting point is a number: writing is about six thousand years old, and it was invented independently only a handful of times. That is nowhere near enough time for evolution to do anything.
The experiment Stroop ran in 1935 is still in use: print the word GREEN in red ink and ask what colour the ink is. You stumble — you are slower than if the word had been TABLE. Reverse it, ask people to read the word and ignore the ink, and there is no cost at all. A word's meaning barges in on its own; you have no switch for it. The colour is the part that takes effort.
And yet this automatic skill has no dedicated equipment. Eyes and ears have tens of millions of years of evolutionary history behind them. Writing has six thousand, invented independently few enough times to count on one hand, and before the nineteenth century most people could not read. No "reading gene" had time to be selected.
Brain imaging nevertheless finds a remarkably stable spot: the ventral surface where the left occipital and temporal lobes meet lights up at the sight of script. Bolger and colleagues pooled twenty-odd studies of alphabetic reading with nine of Chinese in 2005; the activation peaks for Chinese characters and for Latin words sit about six millimetres apart — at the resolution of brain imaging, essentially the same point. Readers of Hebrew and of Japanese peak nearby too.
Dehaene's explanation is called neuronal recycling: reading owns no land of its own, so it requisitions the patch of visual cortex that already did object and shape recognition — a patch that was always good at pulling a stable identity out of a tangle of lines, whatever the angle or size.
The most direct test comes from adults who learned to read late. Dehaene and colleagues reported a group of Brazilian and Portuguese adults in Science in 2010: 10 who had never learned to read, 22 who learned as adults, 31 who learned as children. The more literate a person was, the more strongly that left-hemisphere patch responded to script — and the stronger the phonological response to heard speech became as well. At the same location, the response to faces dropped slightly, while the corresponding right-hemisphere region responded to faces more. Cortex is a finite plot of land; the new tenant moved in and the old one shuffled over.
The strength of the evidence should be stated plainly: this is mostly cross-sectional correlation — people of differing literacy compared with each other, not the same people randomly assigned to learn or not. But the adult-literate group matters especially: it shows the renovation is not something only childhood can do.
If reading is lodging in the object-recognition cortex, then writing cannot look like just anything. It has to look like something that patch can recognise.
Changizi and colleagues did something plodding and solid in 2006: they took apart the characters of more than a hundred writing systems (Chinese characters included, and non-linguistic signs such as road markings too) and counted how often each type of stroke junction occurs — the L, where two strokes meet at an angle; the T, where one stroke butts into the middle of another; the Y, where three strokes radiate from a point; the X, where two strokes cross. Then they counted the same junction types among object contours in photographs of natural scenes. The two distributions match. They also measured that characters in non-logographic systems average about three strokes, and that this does not change with how many characters the system has.
How to read that match: writing did not pick its readers, readers' eyes picked the writing, and the shapes that dropped out over a few thousand years were largely the ones that were tiring to look at. The strength of the evidence is a cross-cultural statistical fit, which is correlational — nobody has watched the selection happen.
The cost side of the deal is hiding in children's spelling errors. The visual system treats a left-right flip as the same thing by default: a cat facing left and a cat facing right are the same cat. For objects that default is free money. For letters it is a disaster, because b and d, p and q are exact mirror images. A child writing letters backwards is not being sloppy; the default has simply not been switched off yet, and part of the work of learning to read is switching it off.
How do we know it is switched off rather than never present? Adults who never learned to read still tend to judge a shape and its mirror image as the same. More interesting is Tamil, a script with no mirror-image letter pairs. Danziger and Pederson found years ago that Tamil-only readers do not make the distinction much; a finer 2020 study compared three groups — non-readers, Tamil-only readers, and readers of both Tamil and the Latin alphabet — and only the last group had fully broken mirror invariance. Whether the script you learned contains mirror pairs decides whether you ever have to train this skill.
Paulesu and colleagues put dyslexic readers from Italy, France and England side by side in Science in 2001. Italian spelling maps onto sound almost one-to-one: see the letters and you know how to say them. English and French offer no such deal. Behaviourally the result was as expected — the Italian group read fast and accurately, the English and French groups far less so.
Look one level down and it changes. Measured against their own national control groups, all three fell behind by the same amount; and on positron emission scans, dyslexic readers in all three countries showed reduced activation in the same left-hemisphere territory, peaking near the middle temporal gyrus. Biologically it looks like one thing, behaviourally it grows into three — and the difference comes from orthography, the set of rules linking written form to sound.
That has a very practical consequence. Dyslexia instruments largely grew up in the English tradition and mostly catch error rates; in languages with tidy spelling rules the difficulty shifts instead onto reading slowly. So one and the same threshold misses different numbers of people in different languages. An Italian national cross-sectional study reported a rate of about 3.5% among children aged 8 to 10, roughly half the figure reported in the United States, with close to two thirds of those cases never previously diagnosed.
The Chinese side is livelier. Siok and colleagues reported in Nature in 2004 that abnormalities in Chinese children with dyslexia showed up more in the left middle frontal gyrus, unlike the left temporoparietal and occipitotemporal pattern typical of alphabetic reading. But Hu and colleagues, in Brain in 2010, controlled for age, ability and task and found the brain-activity differences between the two writing systems to be small. The argument is not over; recent cross-linguistic pooled analyses still oscillate between "a shared core" and "differences the script brings". Samples are typically a few dozen people and the imaging evidence is correlational, so do not read either side as settled.
A blunter piece of evidence comes from Japanese, which uses two systems at once, kanji and kana. After brain injury they can break apart separately: some patients read kana but cannot read kanji, others the reverse. That is a double dissociation — each can fail on its own, so they cannot be running down a single road. Inside one brain, two scripts take routes that do not fully overlap.
One thing should be said clearly: no script here is more advanced than another. Each carries its own bill. Tidy spelling makes decoding easy to get started on and moves the cost elsewhere; Chinese characters are slow going early and pay off in being readable across dialects and across centuries. The difference is in what is hard, not in who ranks above whom.
What is known: Braille readers also break mirror invariance, and not only for Braille patterns. If what gets borrowed is not "vision" but the job of pulling one identity out of a set of stable local configurations, then it works for touch just as well. What observation would separate those two accounts — a channel being borrowed, or something more abstract being borrowed?
At least three candidates: competition for finite area, a shift in how attention is allocated, or plain differences in task difficulty. With only a cross-sectional comparison of people at different literacy levels in hand, what design would separate them — tracking a group of adults as they learn to read, or comparing the direction of change for faces and for objects within the same people?
Options for the next step include testing bi-literate children on both scripts at once, matching task difficulty strictly, or dropping localisation altogether in favour of behavioural core deficits such as phonological awareness. Which explanation does each step rule out, and which one can it never rule out?