Everyone's Getting Dumber and Calling It Efficiency

There’s a confession going around lately, always delivered in the same lowered voice, like someone admitting to a habit. A senior consultant can’t draft a client email without running it through Copilot first. A tenured academic opens ChatGPT before he opens a blank document, because the blank document scares him now. A product manager, twelve years into a career built on writing tight, persuasive memos, says the ten-page doc that used to take an afternoon now feels physically out of reach — not because the ideas are gone, but because the muscle that turns ideas into prose has gone slack.

None of these people are dumb. None are lazy. What’s happening is simpler and worse: each of them has, without ever quite deciding to, handed over a faculty they used to own. Dump a pile of unstructured thoughts into a machine, ask it to “make this sound professional,” and it does. The email goes out. And a sliver of the actual capability quietly dies.

This is not a hot take anymore. It’s a pattern that’s showing up across neuroscience, medicine, aviation, and workplace research, and it’s converging hard enough that “anecdotal” doesn’t cover it. This is not an argument for abstinence or Luddism — it’s an argument for a discipline that’s been forgotten: effortful use.

The oldest lesson, relearned

None of this is new, and that’s actually reassuring. Every technology that extends a human capacity also, used carelessly, lets the underlying capacity rot. This has a settled literature behind it.

Aviation figured it out decades ago. As cockpits automated through the 80s and 90s, pilots got worse at flying. By 2013, a panel convened by the U.S. Federal Aviation Administration warned that commercial pilots were leaning on automated systems so heavily that some had lost the manual skills to take over when those systems failed — they called it “automation addiction.”¹ A follow-up Department of Transportation Inspector General report found the FAA couldn’t even track how often pilots flew manually, and that pilots typically leave the aircraft on autopilot roughly ninety percent of the time.² Controlled studies backed this up: manual flying skills erode measurably from lack of practice, and recent hands-on flying time predicts fine-motor performance better than total career experience does.³ The NTSB pinned part of the blame for the 2013 Asiana crash at San Francisco on a crew that over-relied on automation it didn’t fully understand.⁴

The lesson aviation paid for in lives generalizes cleanly: a skill left unused is a skill being lost, and nobody notices until the automation fails and it’s time to perform.

Same story, different domain, with GPS. Habitual GPS users show measurably worse spatial memory than people who navigate unaided — a longitudinal study in Scientific Reports tracked this and found heavy GPS use preceded the decline in hippocampal-dependent spatial memory, not the reverse.⁵ Worth sitting with: the hippocampus, the exact structure that goes unused when following turn-by-turn directions, is among the first regions to atrophy in Alzheimer’s disease.⁶ The part of the brain being neglected here is the part that can least afford it.

The Google prologue

Before the large language model, there was the search engine — and the search engine already ran a smaller version of this experiment.

In 2011, psychologist Betsy Sparrow published a study in Science that’s only gotten more relevant with time. When people expect future access to information, they remember the information itself worse — and remember instead where to find it.⁷ Sparrow called this “transactive memory”: treating the internet the way you’d treat a colleague who happens to know everything, so the knowledge itself never gets held onto.⁸ A later meta-analysis in Frontiers in Public Health confirmed the “Google effect” is real and consistent — with a useful wrinkle: people with more existing knowledge were less susceptible to it than people who knew less.⁹ Expertise is partly self-protecting. The fuller the head going in, the less a tool can hollow it out.

This is the line that separates healthy offloading from the corrosive kind. Cognitive offloading isn’t inherently bad — a note, a reminder, a calculator, a map are all offloading, and all fine.¹⁰ The trouble starts when the offloaded thing is the process that learning, judgment, and synthesis actually depend on. There’s a real difference between using a calculator to skip long division and never learning what division is. One frees up capacity. The other closes the door on it. (There’s an earlier post here making the optimistic case for that first kind of offloading in design — this is the counterweight.)

The evidence turns direct

GPS and search gave the parable. The last eighteen months gave the direct measurement — and this is the part that’s hard to argue with.

First, medicine. A multicentre observational study in The Lancet Gastroenterology & Hepatology from August 2025 followed nineteen experienced endoscopists — each had performed more than two thousand colonoscopies — as their clinics introduced AI assistance for spotting precancerous polyps.¹¹ Then researchers checked how the same doctors performed without the AI. Three months before AI: 28.4% adenoma detection rate on unassisted colonoscopies. Three months after: 22.4% — a six-point absolute drop, twenty percent relative, across nearly all of them.¹² The authors called it the first real-world documentation of AI-driven “deskilling” in medicine, and admitted they hadn’t expected it from clinicians this experienced.¹³

These weren’t residents. These were among the most practiced people in the field. Three months of routine assistance was enough to blunt a life-or-death skill built over careers. If it happens to doctors catching cancer, it happens to anyone catching typos.

Second study lands even closer to home. In June 2025, Nataliya Kos’myna’s team at MIT Media Lab published “Your Brain on ChatGPT,” fitting fifty-four participants with EEG headsets and having them write essays under one of three conditions: ChatGPT, a search engine, or nothing but their own head.¹⁴ The neural results weren’t close. Brain-only writers had the strongest, most distributed connectivity. Search-engine users were in the middle. ChatGPT users had the weakest connectivity of the three — cognitive engagement scaled down in direct proportion to how much the tool did.¹⁵

Two more findings worth flagging. First, ownership and memory collapsed in the AI group: participants reported the lowest sense of ownership over their own essays and often couldn’t quote back a single line of what they’d just “written.”¹⁶ The work passed through them without ever lodging anywhere. Second — and this is the one that should worry anyone assuming the effect wears off — when the ChatGPT group was finally asked to write unaided in a fourth session, their brain activity stayed weak. The disengagement didn’t switch off when the tool was taken away. The authors called this accumulating deficit “cognitive debt.”¹⁷

The metaphor holds up under scrutiny. Debt feels good in the moment and gets expensive later. It compounds. And it comes due exactly when it’s least affordable — the moment the tool’s unavailable, or wrong, or the task needs the originality the tool can’t supply.

To be fair, the limits on this one matter: the MIT paper was a preprint, not yet peer-reviewed at the time it went viral; the sample was small and mostly Boston-area students; the task was a short essay, not the messy reality of real professional writing.¹⁸ Those caveats matter. They don’t erase a result pointing in the exact same direction as the medical, aviation, GPS, and memory research. Five independent bodies of evidence pointing the same way isn’t a coincidence worth waiting out.

What actually goes missing

Calling this a loss of “skill” undersells it. What’s actually eroding is something bigger, and naming it right matters for the fix.

The largest field study on this in real workplaces — a survey of 319 knowledge workers from Microsoft and Carnegie Mellon, presented at CHI 2025 — mapped this precisely.¹⁹ Across nearly every cognitive task, from comprehension to analysis to synthesis, workers said generative AI reduced the mental effort they put in.²⁰ That’s the whole pitch, obviously. But there’s a more uncomfortable correlation buried in there: the more a worker trusted the AI, the less critical thinking they applied to its output. Meanwhile, workers with more confidence in their own judgment engaged more critically — though it cost them more effort to do it.²¹

The researchers described thinking itself shifting shape under AI use: from information gathering to information verification, from problem-solving to response integration, from doing the task to supervising the task.²² Every one of those shifts sounds efficient. Every one also moves the hard cognitive work out of the person’s head and into the machine’s, leaving behind the thinner job of editor and approver. Here’s the trap: it’s not possible to competently supervise work that the skill to do has already atrophied out of you. The endoscopist who can’t spot the polyp anymore can’t catch the AI missing one either. The manager who can’t structure an argument anymore can’t notice when the machine’s argument is quietly hollow.

This is the real cost of dictating a mess of thoughts and asking the machine to “craft a response.” Structuring a thought is the thinking. Turning a pile of half-formed intuitions into an ordered paragraph isn’t clerical prep work that happens before the real thinking — it is the real thinking. Hand that off, and the important part gets delegated, not the drudgery, while the illusion of authorship stays behind. Same hollowing-out shows up in AI-laundered design feedback — confident, structured, and owned by absolutely nobody.

The counter-argument, taken seriously

Fair’s fair — there’s a real version of the opposing case, and it deserves a straight answer instead of a dismissal.

The case for full embrace goes like this: every generation panics about the tool that kills off some prior skill. Socrates, in the Phaedrus, worried writing would destroy memory — and in a narrow sense he wasn’t wrong, nobody memorizes epics anymore — but externalizing memory into text was the precondition for basically everything humans have built since. Nobody mourns lost skill at long division, or reading a library card catalogue, or folding a paper map. Skills go obsolete once the replacement tool is reliable enough that keeping the underlying ability sharp stops being worth the cost. Maybe unaided prose is just next on that list — a craft looked back on with mild nostalgia and zero real regret, the way penmanship is now. On this read, the “handicap” here is just transition discomfort, and the person who fluently orchestrates AI tools will simply out-produce the purist, every time.

There’s real weight to this, and the Microsoft/CMU researchers themselves lean toward integration over abstinence — pointing to AI tools designed to provoke reasoning instead of replacing it. One prototype in their work prompted users to articulate their own thinking before getting feedback, and it improved engagement instead of eroding it.²³ The tool isn’t the villain. The manner of use is.

But the analogy to writing and arithmetic breaks at a specific joint. Long division and map-reading are narrow skills. Structuring an argument, holding a problem in the head long enough to see its actual shape — these are general faculties, load-bearing under nearly all high-value cognitive work. Offload a narrow skill, and the general faculty gets freed up to do more. Offload the general faculty, and there’s nothing left underneath. The colonoscopy study is the tell: what atrophied wasn’t paperwork, it was perception — the actual core of the job. That’s the line. Automate the narrow stuff. Protect the general.

What to do about it

None of this is an argument to quit the tools. That’s unrealistic, and given how genuinely useful they are, unwise too. It’s an argument to deliberately reintroduce the effort that convenience quietly removed — treat cognitive fitness the way physical fitness gets treated, as something that needs resistance to stay intact. A few disciplines fall directly out of the evidence above.

Draft before delegating. The single most damaging habit is the one from the opening — dictating unstructured thoughts and asking the machine to structure them. Flip that order. Structure it badly, alone, first — then hand it to the tool for polish. The MIT finding was that the “brain-to-LLM” group, who engaged their own thinking before reaching for the tool, kept far more of the neural engagement and memory that the “LLM-first” group lost.²⁴ The order of operations isn’t a small detail here. It’s the entire game.

Keep an AI-free zone. The endoscopists lost their edge because every case became AI-assisted — no protected space existed to keep the raw skill alive. Pick categories of work — the email that actually matters, the argument being judged on, the memo with a name attached — and do the cognitive labor unaided there, specifically because it counts as practice. Aviation already taught this lesson: recent hands-on practice predicts performance better than total accumulated experience.³ Practice isn’t nostalgia. It’s maintenance.

Verify from knowledge, not from trust. The sharpest finding from the Microsoft/CMU study was that trust in the AI displaced critical thinking, while confidence in one’s own expertise restored it.²¹ The defense against blind acceptance is a stocked mind. The Google-effect meta-analysis makes the same point from the other direction: people who already know a lot get hollowed out the least.⁹ Keep learning things into the head, not just learning where to look them up later. Existing knowledge is the immune system that makes it possible to use AI without being used by it.

Watch for the ownership signal. If restating, unaided and in plain words, the substance of something just produced with AI isn’t possible — that thing wasn’t produced, it was passed through. The inability to quote back an essay was, in the MIT study, the clinical sign of shallow encoding.¹⁶ Treat that as a warning light. Real authorship leaves a trace in memory. Its absence is the tell that the work went around the mind instead of through it.

Coda

The people confiding this quiet dependency in hushed voices aren’t describing a character flaw. They’re describing a failure to notice, until it was well underway, that convenience is never free — that every faculty left unexercised is a faculty being forfeited. The machine didn’t steal anyone’s fluency. It was given away, one delegated email at a time, in exchange for a small daily saving of effort, with nobody watching the ledger where the debt was quietly piling up.

The good news: cognitive debt, unlike some debts, pays down. The faculty that atrophies from disuse comes back with use. But recovery needs the one thing these tools are engineered to spare everyone from — and the one thing this whole piece has been arguing for reclaiming: doing hard mental work when an easier option is sitting right there, glowing, offering to do it instead.

The blank document isn’t the enemy. Fear of it is the symptom. The cure is sitting with it, unassisted, more often than feels comfortable — because that discomfort is just the feeling of a mind still doing its own work.