The hardest myth to kill isn't "we only use 10% of our brains" — it's "a brain image counts as evidence".
You've probably met these claims: engineers are left-brained and artists right-brained; we only ever use 10% of the brain; some people are visual learners and need to be taught accordingly. They spread easily, they sound right, and they all carry a faint scientific accent. But the real target of this issue isn't those three claims — it's the machinery underneath them: why a statement that couldn't stand on its own suddenly becomes credible the moment the word "brain" is attached. And by the end the question is no longer "which one is false", but whether knowing the neuron layer counts as understanding at all — a problem neuroscience hasn't solved for itself either.
Start with the three most famous. The point isn't that they're wrong; it's where they go wrong — each one begins with a genuine finding and takes exactly one step too many.
Left brain, right brain. The true part: the hemispheres really do have processing preferences. In the vast majority of people (about 95% of right-handers) language leans left; faces, speech prosody and holistic spatial sense lean right. That lean is called lateralisation and it is thoroughly real. The false part is the next step: reading "this function leans slightly to one side" as "people come in left-brained and right-brained types". The raw material for that jump came from a very unusual group of patients — people whose bridge between the hemispheres was cut to control intractable epilepsy, the split-brain cases. In them, the two sides really do go their own way. But yours wasn't cut. Your bridge carries some 200 million fibres reconciling both sides every millisecond. When researchers went looking for "left-brained" and "right-brained" individuals in scans of over a thousand people, they found none: everybody uses both. (How that bridge works, and what happens when it's severed, corpus callosum)
The 10% claim. This one dies in a second, and not to neuroscience — to arithmetic. The brain is 2% of body mass and consumes about 20% of the body's energy. Evolution haggles over every spare gram of fat; it does not bankroll an organ idling nine tenths of the time. The clinic makes the same point daily: a stroke or injury that damages a very small patch almost always takes something with it — reading, recognising faces, control of one finger. There are no spare rooms in there.
Learning styles. The slipperiest of the three, because half of it is true: people do have preferences — you may genuinely prefer diagrams while someone else prefers hearing it explained. What's false is the second clause quietly attached to it: "so teaching each person in their preferred style makes them learn better." That's the meshing hypothesis, and decades of tests keep coming back the same way: preference doesn't predict which mode of instruction actually works for you. What decides that is the material — geometry wants a picture, pronunciation wants sound — largely regardless of who is sitting there.
Artificial neural networks fell into almost exactly the same trap. They're built from numerical units also called "neurons", and the natural early assumption was that one unit stands for one concept — find the "cat unit" and you've found cat. It turned out not to work that way: a single unit participates in a pile of unrelated features (the phenomenon is called superposition), and any one concept is smeared across thousands of units. "One slot, one thing" is the shape humans find most comfortable — and neither biological brains nor artificial networks grow that way.
Myths survive not on evidence but on looking like evidence. Several concrete mechanisms do that work.
First, the neuroscience accent itself scores points. In one experiment, non-expert readers got two versions of an explanation for the same psychological phenomenon: an ordinary psychological account, and the identical account plus one throwaway clause — "because the frontal lobe circuitry is activated at that moment". The extra clause adds no information whatsoever, yet readers found the second version more convincing. A bad explanation stops looking so bad once it's wearing neural vocabulary. (A related line of work asked whether adding a brain image makes it more persuasive still; that effect turned out much weaker than first reported and has repeatedly failed to replicate — in an issue about neuromyths, the effects being cited deserve their own replication footnote.)
Second, and harder: the scanner manufactures false positives by construction. A functional MRI carves the brain into tens of thousands of little cubes (voxels), and each cube gets its own statistical test. At the conventional threshold, roughly one voxel in twenty comes out "significant" on luck alone. Multiply that by tens of thousands and, uncorrected, you are guaranteed a handsome coloured blob.
This was demonstrated in 2009 in the most withering way available: a dead Atlantic salmon was placed in a scanner and shown photographs of people, and asked what emotion those people were experiencing. Without multiple-comparison correction, the fish's brain produced "significant activation". The authors weren't after the laugh (though they collected an Ig Nobel for it) — they were forcing the field to make correction the default.
A subtler trap is double dipping: pick out the voxels with the highest correlation across the whole brain, then compute a correlation from that same selected set, then report "r = 0.9". That's writing the exam after reading the answers; the number has to come out inflated. A 2009 paper that named names called these "voodoo correlations", and touched off a round of methodological housecleaning across social-cognitive imaging.
The same hole shows up in AI as leaderboard chasing: run dozens of random seeds and hyperparameter settings, then write up the best one. That is the identical piece of mathematics as "pick the brightest of fifty thousand voxels" — a good fraction of the extra tenths of a point on a benchmark is that salmon. And the defences are identical too: say in advance what you're going to look at, hold out data nobody has touched, and count the failed runs in the denominator.
Assume the correction is impeccable and the region genuinely activated. There's a deeper obstacle waiting, called reverse inference.
The forward step is easy: frighten someone and the amygdala usually activates — an experiment can measure that directly. The reverse step doesn't hold: seeing the amygdala active does not tell you the person is afraid. Far more than fear lights it up — anything novel, anything uncertain, anything intense enough to deserve extra attention, including happy things. To run backwards from "active" to "afraid" you'd need to know how readily that region responds to everything else — and many regions respond to almost everything. (One survey found Broca's area showing up in a substantial fraction of all imaging studies, whatever the task.)
There's another layer people skip: functional MRI doesn't measure firing at all. It measures changes in oxygenated blood flow — activity raises metabolic demand and blood follows. It's a proxy, lagging by several seconds, with hundreds of thousands of neurons packed into each voxel. And it doesn't distinguish excitation from inhibition: suppressing a region costs energy too, and looks just as bright.
Finally the old line, which has very specific weight here: correlation isn't causation. "Region Y activates during task X" says only that the two co-occur. To claim Y does X you have to intervene: disrupt it briefly with transcranial magnetic stimulation, look at what patients with damage there have lost, switch specific cells on and off with light in animals. Observation gives you suspects; convictions need intervention.
Interpretability research has a near-isomorphic practice called probing: train a small classifier on a model's internal activations, find that it can read out "this sentence is positive", and announce that the model represents sentiment there. That's the same logical error as "amygdala active = fear" — being decodable is not the same as being used. Which is why the field now has to add the causal step: edit, patch or ablate those activations and see whether the output changes. A lesson neuroscience took twenty years to learn, being relearned move for move.
Now suppose every methodological problem is solved. You have the complete wiring diagram and every spike from every cell. Is that understanding?
Two examples say no, painfully. The first: C. elegans, a 1 mm roundworm with 302 neurons, whose complete wiring diagram was published in 1986. Forty years on, nobody can predict from that diagram which way it will crawl next — because neuromodulators rewire the same wiring on the fly into different circuits. The connections are fixed; the working circuit is not.
The second is harsher. In 2017 two researchers ran neuroscience's entire standard toolkit on a 1980s microprocessor — a chip running Donkey Kong and Space Invaders, whose every transistor and every wire is fully known to us. They did "lesion studies", knocking out transistors one at a time, and duly found several whose removal stopped the game from running: in a brain you would announce the discovery of the "Donkey Kong neuron". They produced tuning curves, connectivity analyses, dimensionality reduction. The figures look great. Not one conclusion touched how the processor actually works. The tools weren't broken; something else was missing.
What's missing is levels. The vision scientist David Marr separated them long ago: the computational level (what problem is this system solving, and why is this the right solution), the algorithmic level (what representations, what steps), and the implementational level (how neurons, transmitters and ion channels make it happen). Each carries its own kind of explanation, and none substitutes for another. Staring at the implementation and climbing upward doesn't get you the top two — not for want of data, but because that's the wrong question.
And there's a converse: the same job can have many implementations. Timekeeping runs on gears, on quartz, on atomic transitions; memory runs on synaptic strength here and could run on something else elsewhere. So a one-to-one "this mental state = that neural activity" was never a reasonable thing to expect.
None of which says reduction is useless. Quite the opposite — everything in the previous thirty-odd issues came from it: dopamine prediction errors, glymphatic clearance running at night, the single receptor lock psychedelics fit. Reduction is the only tool that supplies mechanism for "why". Its limit is one sentence long: mechanism doesn't automatically become explanation. You still have to say what problem the machine is solving.
There's a ready-made control experiment here, and the result is uncomfortable: we hold every parameter of a large language model, can read and edit any of them, and can rerun it as often as we like — conditions no neuroscientist will ever get. And we still can't say how it does two-digit addition; the mechanisms teased out over the past few years were excavated by hand with a magnifying glass. What separates total implementation detail from understanding isn't volume of data — it's a vocabulary that spans levels, and both fields are short of the same thing.
"Not finding it among the parts doesn't mean it isn't there" is not a neuroscientific discovery. Several traditions hit this same wall long ago, each in its own terrain: