A Short History of Outsourcing the Brain: The Cabbie, Socrates, and a Machine That Thinks for You

(0) A Brain That Holds Twenty-Five Thousand Streets but Can't Remember a Single Figure

Let me show you a very strange brain.

It belongs to a London taxi driver. Inside his head sit twenty-five thousand streets and over a thousand landmarks; with his eyes closed he can find his way from any point in the city to any other, no navigation required.

Scientists slid him into an MRI scanner and found that the part of his brain responsible for spatial memory—the posterior half of the hippocampus—was noticeably larger than average.

Sounds like a superhero, right?

Hold on.

That same group of scientists also found: this cabbie couldn't remember a complex figure they'd just shown him. On that test, his score was lower than an ordinary person's.

A brain that can hold all of London, but can't remember a single drawing.

This isn't a bug. It's the entire secret this essay is about:

The brain never simply "grows" or "shrinks"—it strikes a bargain. Whatever you want to load more of, you have to clear space for.

For over two thousand years, humanity has been using tools of all kinds to "make room"—writing, the printing press, the calculator, GPS. Every single time, we've made the same bargain.

And now we hold a tool unlike any before it. It doesn't remember things for you; it thinks for you.

In this bargain, what exactly are we trading away, and for what?

I want to walk you through this slightly frightening story from the angle of its "intellectual lineage."

(1) Maguire

Eleanor Maguire was a neuroscientist at University College London.

In 2000, she did something that would later be written into every psychology textbook: she stuffed 16 London cabbies into an MRI scanner and compared them against 50 ordinary people.

The result was earth-shattering—the cabbies had noticeably more gray matter in the posterior hippocampus, and the longer they'd been driving, the larger it had grown; correspondingly, their anterior hippocampus was somewhat smaller.

This was the first time humans had directly seen it: the adult brain, subjected to long, intense cognitive training, will grow new structure.

But there was a fatal loophole: could it be that people whose hippocampus was already large are simply the ones more likely to pass the cab-license exam? In other words, did "driving grow the brain," or did "large-brained people become drivers"?

In 2011, Maguire's team drove home the final blow.

To get a London cab license you have to pass a devilish exam called "The Knowledge"—it takes three to four years on average to master, with an extremely high washout rate. She recruited a group of people currently studying for it, scanned their brains once before they began, and scanned them again three or four years later.

The result was clean and decisive:

• Those who passed really did grow a larger posterior hippocampus.
• Those who failed, along with the ordinary control group, didn't budge an inch.

Before the exam, the brains of all three groups showed no difference whatsoever.

The self-selection loophole was sealed shut. It was the training that grew the brain.

But Maguire's honesty lay in this: she also reported the cost of that "bargain." The cabbies who passed showed worse delayed recall of complex figures—spatial expertise didn't come for free.

(Let me clear up a detail that's often muddled: "the anterior hippocampus was somewhat smaller" was a finding from the cross-sectional comparison in 2000; the longitudinal follow-up in 2011 only observed the posterior growing, and conjectured that the anterior might later shrink along with it—it did not directly measure any shrinkage at the time. Don't conflate the two.)

Remember this word: bargain. The whole story that follows is nothing but this.

(2) Socrates

You think "will tools make us dumber" is a new question?

Rewind twenty-four hundred years.

Around 370 BCE, in the Phaedrus, Plato put an Egyptian myth into the mouth of Socrates.

The god of invention, Theuth, went to present his gifts, claiming he had invented writing, which "will make the Egyptians wiser and improve their memory."

The king, Thamus, poured cold water on him then and there:

"This invention of yours will breed forgetfulness. People will rely on the written characters and cease to exercise their own memory. What you have given is not a remedy for memory, but merely a tool for 'reminding.' They will appear to be full of learning while in truth knowing nothing."

The worry of a king twenty-four hundred years ago is exactly the worry of today: people will stop exercising their inner faculties because they depend on external tools.

And what's the finest irony?

Socrates never wrote a word, yet Plato wrote this passage down. A man immortalized through writing, using writing to denounce writing.

Later the French philosopher Derrida wrote an entire essay on this matter, Plato's Pharmacy. He seized on the Greek word Plato used—pharmakon—which means both "remedy" and "poison" at once.

Here is the second word to remember in this essay: every cognitive technology is, by its very nature, a drug that both cures and sickens. So it was with writing, so with the printing press, so with GPS, and so too with AI.

The question was never "is it poisonous," but "how do you use it."

(3) Bohbot and Spiers

Pull the lens back to the modern day. After writing and the printing press, humanity invented something that truly "navigates for the brain"—GPS.

Would the London cabbie's brain, thanks to GPS, regress back to ordinary?

Two teams of scientists closed in on this question from two directions.

The first team, Spiers (Hugo Spiers), 2017, Nature Communications. He had people walk through London's Soho while their brains were scanned. What he found: when people worked out the route with their own minds, activity in the hippocampus and prefrontal cortex spiked at every junction; but once they followed the GPS voice, those two regions barely lit up at all.

GPS isn't helping you navigate. GPS is hitting the mute button on the very part of the brain that does the navigating.

The second team, Bohbot (Véronique Bohbot), 2020. Her tracking found that people who used GPS heavily over the long term showed a steeper decline in spatial memory years later. And it wasn't a case of "the directionally hopeless were the ones using GPS"—she ruled that causal direction out.

Sounds like "GPS degrades the brain" is nailed shut, doesn't it?

Not so fast—here we have to be honest. Bohbot's longitudinal sample was only about a dozen people; she herself warned that "the sample is too small, it may be a spurious correlation, don't draw strong conclusions," and she did no brain imaging at all—what she measured was behavior, not whether the brain had shrunk. Another pre-registered study even found that in ordinary people, navigation ability and hippocampus size are not correlated at all.

So the headline "GPS shrinks your brain" has, to this day, never been directly confirmed by any single study. It's a reasonable inference drawn from a solid chain of evidence, but the last nail hasn't been driven in.

Still, in 2024 there was a finding worth savoring.

A Harvard team dug through the US CDC's mortality data and found: the rate of death from Alzheimer's among taxi drivers was only about 1%, among ambulance drivers under 1%, while for the general population it was 3.9%. More telling still—bus drivers, pilots, and ship captains, who work fixed routes, did not get this protection.

It's "navigating" itself, not "driving," that protects the brain.

But here we have to apply the same standard of skepticism as before—we can't relax just because the conclusion is pleasing. This is a cross-sectional observational study; it can't prove causation, and the researchers themselves only said it "suggests a possibility." And there's an unavoidable confounder: the average age at death for taxi and ambulance drivers is itself lower (ambulance drivers around 64, taxi drivers around 68), while Alzheimer's is a highly age-related disease—if someone dies earlier of some other cause, their proportion of "death from Alzheimer's" is naturally suppressed. The study did adjust for age at death, but this layer of confounding can't be fully scrubbed away.

The researchers said it plainly too: as drivers come to rely more and more on GPS, this advantage (if it truly exists) may vanish in future generations.

The third word: use it or lose it. The juggling experiments proved long ago that the gray matter you build up will revert to its original state after just three months of not practicing. Ability is a matter of use. Whatever you outsource, retreats.

(4) That Fourth-Grader

By now you may be panicking: so is all outsourcing a bad thing? Should we just stop using any tools at all?

No. The evidence says exactly the opposite.

In 1986 there was a meta-analysis synthesizing 79 studies, specifically to see whether the calculator had ruined children's arithmetic.

The conclusion disappointed everyone in the "tools rot the brain" camp: in the vast majority of cases, calculators used alongside instruction either had no effect, or actually made pencil-and-paper calculation better and made children like math more.

But—note this but—fourth grade in elementary school was the sole exception. Persistent reliance on a calculator at that age did indeed drag down the development of foundational skills.

This one point is the finding in the entire outsourcing story that most deserves to be carved into the wall:

When a tool is integrated into learning, it's a gain; when a tool replaces learning, that's the loss. And the most dangerous moment is the stretch of time when the brain is in the middle of growing the skill.

An adult cabbie using GPS is merely under-exercising a muscle he's already built.

But a child who has never once found his own way may never grow that muscle at all.

Remember this fourth-grader. When we get to AI, he'll come back.

(5) Kosmyna

In the summer of 2025, a paper detonated across the whole internet.

The team of Kosmyna (Nataliya Kosmyna) at the MIT Media Lab put EEG caps on a group of people and had them write essays, splitting them into three groups: those using ChatGPT, those using a search engine, and those relying on their brains alone.

The result: "the ChatGPT group's brain connectivity was the lowest across all frequency bands," and afterward, when they were asked to recall the essay they had just written without AI, many couldn't remember a single line.

The team gave this phenomenon a frightening name: "cognitive debt."

The media climaxed instantly, headlines marching in lockstep: "MIT Proves: ChatGPT Makes You Dumber."

Hold on.

First, the paper is a preprint that has not been peer-reviewed.

Second, it had 54 people total, and the key "write again with the AI removed" stage was down to just 18 people—six per group.

Third, what it measured was the mental effort of writing one essay in the moment, not any lasting damage to intelligence.

Fourth, and most ironic of all—the author herself publicly begged reporters not to use words like "stupid," "dumb," "brain rot," "harm," or "damage." She said the paper proved nothing of the sort about LLMs making people stupid.

Independent scholars twisted the knife further: the brain-only group's connectivity growing stronger and stronger was very likely just because they practiced the same task three times over (a familiarity effect), while the AI group did their no-AI task only once. This is not a fair comparison at all.

You see—the pharmakon is back. Socrates' bucket of cold water was poured, unaltered, into the trending topics of 2025. A real but tiny signal was blown up into a panic of the century.

AI right now does indeed make you use your brain less. But between "using it less once" and "permanently getting dumber" lies an entire, still-unproven chasm.

(6) Toner-Rodgers

If (5) was "an exaggerated panic," then this chapter is "a fabricated hope."

In late 2024, an MIT PhD student, Toner-Rodgers (Aidan Toner-Rodgers), threw out a paper that dazzled the whole field of economics.

He claimed to have tracked 1,018 materials scientists after they adopted AI, and arrived at an extremely seductive conclusion:

The output of the top scientists nearly doubled, while the bottom third made almost no progress. Because the experts knew how to use professional judgment to filter the options the AI gave them, while the novices were led astray by a pile of false positives.

This was the perfect smoking gun for "AI amplifies the elite and abandons the mediocre." Two top-tier economists even endorsed it. Everyone who believed in the "pyramid" had found their Bible.

Then, in May 2025, MIT killed it themselves.

The institution issued a statement: it had "no confidence in the paper's data provenance, reliability, or validity." The two endorsing heavyweights withdrew their support, and the paper was pulled from the preprint platform. A subsequent investigation found that this guy had fabricated the data-use agreements, that the companies named said they had never run the experiment at all, and that he had even registered a domain impersonating the Corning company to shore up the lie (later ruled against in arbitration).

He is no longer at MIT.

The most seductive cornerstone of the whole "AI selectively amplifies the elite" story is fake.

The lesson this gave me matters more than the conclusion itself: the more you long to believe a conclusion, the harder you must check its source. A story that flatters everyone's biases is often too good to be true—because it was fake to begin with.

After that cornerstone collapsed, what real thing about the "pyramid" do we still hold?

(7) Brynjolfsson

The real thing that remains points in exactly the opposite direction.

Stanford's Brynjolfsson (Erik Brynjolfsson) ran a hardcore study in 2023: 5,179 customer-service agents, given an AI assistant.

The result:

• Average efficiency +14%.
• Novices and low-skilled workers, +34%.
• Top veterans, nearly 0.

What did the AI do? It took the tacit experience of the very best people, packaged and distilled it, and handed it to the greenest ones. A newcomer of two months, on the strength of AI, produced work at the level of a six-month veteran.

The BCG experiment with 758 consultants, the experiments with hundreds of developers using Copilot—all point the same way: the bottom rises the most, and the gap is compressed upward.

This is called the "leveler." It is the opposite of the "pyramid."

So you see, the evidence here splits in two:

• The fabricated paper said "AI widens the gap."
• A pile of real studies say "AI narrows the gap."

So that original intuition—"top-tier thinkers vs. a base that has given up thinking"—is it right after all?

The answer is held in the hands of the next person.

(8) Mollick and the 19% Illusion

That BCG consultant study was led, in part, by Mollick (Ethan Mollick) of the Wharton School. He gave AI's capability boundary a brilliant name: the "jagged frontier."

The meaning: AI's capability boundary is uneven—inside it AI is a god, outside it a demon, and you often don't know which side you're standing on.

The very same study hides two diametrically opposite stories:

• Within AI's capabilities: the quality of bottom-tier consultants soared 43%, everyone happy.
• Beyond AI's capabilities: the people using AI saw their accuracy drop from 84% to 60–70%, dragged confidently into wrong answers by the AI. And everyone's solutions grew more and more alike—diversity collapsed.

That second half is the true scholarly form of the opening worry about "a base that has given up thinking." It isn't wage divergence; it's cognitive homogenization: a huge crowd of people stop producing their own judgment and all converge onto that "looks pretty good" median answer the model spits out.

There's an even more cutting study, from an independent outfit called METR. In early 2025 they had 16 senior open-source developers use AI to edit codebases they knew inside out.

The result: they were 19% slower.

More devastating still—these people had beforehand expected to be 24% faster, and afterward still believed they'd been 20% faster.

A 40-percentage-point cognitive illusion: people use AI, their efficiency drops, yet they're convinced they're taking off.

(Let me stamp a timestamp on this number: it's a snapshot of that batch of tools from early 2025. METR itself softened its stance in a February 2026 follow-up—for the returning cohort, with newer tools swapped in, the estimate had flipped to something like "18% faster," though with wide error bars. So don't treat this line as eternal truth; it's more of an "early snapshot." But the 40-percentage-point cognitive illusion is far more durable than the perishable number of "faster or slower.")

"AI helps everyone equally"—false.

"AI only helps the elite"—also false (that paper was fabricated).

The string of experiments from 2023 to early 2025 points consistently in one direction: at the task level AI is a leveler, but it levels a vast swath of people up onto a confident, mediocre plateau.

(9) Terence Tao

And what of the apex of the pyramid? Has anyone really been amplified by AI into a god?

Yes. But not on the strength of that fake paper.

Terence Tao—one of the greatest living mathematicians, a Fields Medalist. In November 2025, he and his collaborators threw an AI system at 67 mathematical problems, and on roughly 20 of them, the AI's results surpassed the best in the existing literature (on 8 it did worse than the literature).

But listen to what he said:

"I would strongly caution against using it without the ability to independently verify the AI's output… it will amplify hallucination, sycophancy, and the ungrounded."

This is the secret of the apex.

Tao uses AI not because he can't be bothered to think, but because he is strong enough to judge when the AI is talking nonsense. In his hands AI is a magnifying glass—what it magnifies is his already-formidable judgment.

AlphaFold won the 2024 Nobel Prize in Chemistry and serves two million scientists; AI has reached gold-medal level at the Mathematical Olympiad. The amplification at the top is real.

But note that decisive watershed: Tao and the consultant led astray by AI in (8) are using the same tool. The only difference lies in one thing—he can catch the AI's confident mistakes, and the other cannot.

This, right here, is the hard currency of mental power in the new era.

(10) The Rebels

The story is told. Now I'll bring everyone back into the same room to answer the question from the opening:

Will AI create top-tier thinkers at the peak of a pyramid, and a base that has given up thinking?

My answer is: the intuition is half right, but the terrain is drawn wrong.

The real terrain isn't a pyramid; it's a plateau, capped by a single thin spike:

• The spike is the Terence Taos—those who treat AI as a verifiable co-pilot, retain independent judgment, and are amplified into something unprecedented. On this point, the existing evidence strongly agrees.
• The plateau is the leveled majority—AI has raised their floor, but also flattened their ceiling, producing a heap of "looks decent, but has stopped thinking independently" homogeneous content.

The real danger isn't that the base gets left behind; it's that the middle caves in into a plateau. It isn't that a few climb too high, but that the majority are lifted to a confident, mediocre plane—and then stop there, never climbing higher.

So, in these twenty-four hundred years of "outsourcing the brain," in this final bargain, what exactly are we trading, and for what?

String together what those few people told us, and the answer surfaces:

• Maguire's cabbie tells us: the brain is a bargain, not an addition or subtraction.
• Socrates tells us: every tool is a pharmakon; poison or remedy depends entirely on how you use it.
• Bohbot's GPS tells us: use it or lose it—whatever you outsource retreats.
• That fourth-grader tells us: an adult can outsource without harm, but outsource in the years of growing the skill, and that part of the brain may never grow at all.
• Terence Tao tells us: what's valuable in the new era is no longer "generating the answer," but "knowing when the answer is wrong."

So the truth of this bargain is:

In the past we outsourced memory to writing and the printing press—but you still had to reason for yourself.

Now we are beginning to outsource reasoning itself to the machine.

Last time we outsourced, the mental effort we saved went into thinking. This time, if you save away the thinking too, there is no next stop.

The whole proposition of education is thereby compressed into a single sentence: how to keep AI from short-circuiting that "necessary cognitive struggle." Because it is precisely the struggle that grows the skill. AI's greatest temptation is exactly this—to skip that stretch of road for you, the most uncomfortable one, and the one that builds the most muscle.

Will the diplomas of elite schools be devalued? Yes, but that's the surface. What's truly being devalued is "knowing a lot." What's truly appreciating in value is "the person who, amid a field of confident, homogeneous answers, can still say, 'No—this part is wrong.'"

That London cabbie's hippocampus, in fact, foretold all of this long ago—

Ability is a matter of use. The road you walk with your own feet grows into your brain; the road you hand to the machine to walk, the machine remembers, and you do not.

The machine can carry you to any place.

But it cannot, on your behalf, grow the road.

All data in this essay come from peer-reviewed research or original institutional reports, with the key conclusions verified point by point; where the evidence is early or single-point, its timestamp and limitations are noted in the text.