Researchers at MIT Media Lab put EEG caps on fifty-four people and asked them to write essays. One group used ChatGPT. One group used a search engine. One group used nothing but the wet grey lump they were born with.

The brain-only group lit up like a switchboard. Widespread, distributed neural connectivity, the kind you see when a brain is doing something to itself rather than just moving information around. The search engine group came in somewhere in the middle. The ChatGPT group was the quietest room in the building.

Then came the part that should stop everyone.

When the researchers asked participants to quote a line from the essay they had submitted minutes earlier, the AI group struggled. Not "misremembered the phrasing" struggled. Could not reliably produce a sentence from a document with their name on it. The words had passed through them like a train through a station.

The researchers called it cognitive debt. I would have gone with something less polite.

I want to be careful here, because this is exactly the kind of study that gets screenshotted into oblivion by people who want a reason to be furious. It is small. It is still a preprint, more than a year after release. Fifty-four people is a seminar, not a civilisation. Hold it loosely.

But hold it next to something else.

In early 2024, two researchers at NTNU in Norway, Ruud van der Weel and Audrey van der Meer, wired up thirty-six university students with a 256-channel EEG array and had them write words two ways: by hand with a digital pen, and by keyboard. Handwriting produced widespread connectivity in the theta and alpha bands, the patterns associated with laying down memory. Typing did not. Same words. Same brains. Different architecture engaged.

Again, caveats, said out loud rather than buried in a footnote: these were adults, not children, writing single words rather than composing arguments, and a published commentary in the same journal pushes back on how far you can stretch the classroom implications. Fine. Noted.

Then a third one, and this is the one I keep thinking about.

Cognitive offloading is a real term and a very simple idea. It is what you do when you write a phone number down instead of memorising it. When you set an alarm instead of remembering. When you take a photo of where you parked. You are moving a mental task out of your skull and into the environment, and it is one of the smartest things our species does. Language is offloading. Writing is offloading. The entire history of human progress is, in one reading, a history of finding places to put thinking that are not our heads.

The question is only ever what you do with the capacity you free up.

Michael Gerlich, at SBS Swiss Business School, surveyed 666 people in the UK with a twenty-three-item questionnaire, then sat down for long interviews with fifty of them. The questions were the boring, useful kind. How often do you use AI to find information? How often do you rely on it to make a decision? Do you use devices to remember things for you, or to solve problems for you? Then a set of items drawn from the Halpern Critical Thinking Assessment, a standard instrument, about whether you check sources, weigh evidence, notice bias, sit with a problem before reaching for an answer.

The correlations were strong. AI use against critical thinking: minus 0.68. Cognitive offloading against critical thinking: minus 0.75. In social science those are big numbers, big enough that I would want them replicated before betting anything important on them.

But it was the interviews that got me, because people were diagnosing themselves without being prompted.

"I use AI for everything, from scheduling to finding information. It's become a part of how I think."

"The more I use AI, the less I feel the need to problem-solve on my own. It's like I'm losing my ability to think critically."

"I rarely reflect on the biases behind the AI recommendations; I tend to trust them outright."

That second one is a person watching it happen to themselves, out loud, in a research interview. Nobody made them say it.

The age pattern is the part that matters most for anyone raising or teaching a teenager. The youngest group, seventeen to twenty-five, used AI most, offloaded most, and scored lowest on critical thinking. The oldest group, forty-six and up, used it least and scored highest. Education was the buffer. One participant with a master's said they cross-check everything because they know it is not always accurate. Another, a high school graduate, said: "I don't have the time or skills to verify what AI says; I just trust it."

The caveats matter here too. This is correlational, so it cannot tell you which way the arrow points. Maybe heavy AI use erodes critical thinking. Maybe people with weaker critical thinking reach for AI more. Probably both, feeding each other. It is UK-only, recruited through social media, and parts of the measure are self-reported, which is a bit like asking people to rate their own driving.

Three studies, three methods, three sets of limitations, all pointing at the same unglamorous idea. It is not the tool. It is what the tool removes.

And we have got that backwards. We treat effort as the tax you pay to reach the outcome. Friction to be engineered away, the way we engineered away the crank handle and the fax machine. But in learning, effort is not the tax.

Effort is the mechanism.

The struggle to find the word is where the word gets stored. The hand dragging across the page is doing something the keyboard does not. The half hour you spent stuck, pulling at your hair, forehead on the desk, hating the question and quietly hating yourself, is not the price of the insight. It is the insight being manufactured. Remove it and you have not made learning faster. You have made a very convincing simulation of learning, which is a different product entirely, and one nobody would buy if it were labelled honestly.

I was born without a sense of smell.

I found out at a primary school science fair. There is a classic demonstration where you cube an apple and a white onion, blindfold someone, block their nose, and they cannot tell which is which, because without aroma you are left with sugar and crunch and they are close enough. I got the blindfold and the nose clip and called them correctly, over and over. The other kids decided I was cheating. A teacher held my nose herself to be sure. I was, briefly, an eight-year-old under investigation for onion fraud.

Years later, for reasons that made sense at the time, I spent a chunk of my twenties working as  a sommelier.

Here is what that actually involves when the primary instrument is missing. It is not that everyone else gets the wine handed to them complete. Nobody does. Structure, acid, tannin, weight, length, the way alcohol sits at the back of the throat: that is all palate, and everyone has to work for it. What the nose gives you is the naming. The blackcurrant, the wet stone, the barnyard funk, the oak. It gives you identification, fast and free, and identification is what wine culture rewards, so it is what most people lean on.

I did not have that lane. So I built the map out of everything else, slowly, badly, with an enormous amount of wasted effort, and over years the map filled in. My palate got sharper because it had to carry the whole load. Faults, in particular. Vintage. Temperature. The specific way a wine goes quiet after twenty minutes in the glass.

I am not telling you I was better than my colleagues. Some of them were extraordinary. But their route ran through a sense that does the work for you, and mine did not, and that turns out to matter on exactly the nights when things get strange.

Because here is what happens in a blind tasting. Serve everything at the same temperature, take away the visual, and a startling number of experienced drinkers cannot reliably tell you whether what they are holding is red or white. Give someone a head cold and watch their expertise evaporate. Their instrument was doing the work, and when it goes offline there is very little underneath.

I never had the instrument. So I had to build the thing underneath.

That is the whole argument in a glass. The people who get the answer for free are not learning the thing. They are receiving it. And most of the time that is completely fine, right up until the moment it is not.

The World Bank ran a randomised controlled trial in Edo State, Nigeria. Eight hundred students in their first year of senior secondary, so roughly fifteen and sixteen years old, Year 10 in Australian terms. Six weeks of after-school English with an AI tutor, twice a week, in computer labs, over June and July 2024.

Six weeks later they were 0.31 standard deviations ahead of the control group. The World Bank's own cost-effectiveness analysis puts that at the equivalent of one and a half to two years of business-as-usual schooling, which places it among the most cost-effective education interventions anybody has rigorously tested.

The girls started behind the boys and made the biggest gains, closing the gap. The lower performers gained more than the high performers, which is the opposite of what almost every education technology in history has managed. Many of these students had never used a computer before week one, so they were learning to operate a mouse and learning English at the same time. Months later, in final exams, in other subjects, the effect was still showing up.

So there it is. Hand fifteen-year-olds a machine that does the hard part and they cover close to two years of English in six weeks. Every word I have written above appears to be wrong.

And it would be, if you stopped at the result. Which is what almost everyone did, because the result is the bit that fits in a headline and the reason it worked is buried in the methodology, where nothing goes viral.

Teachers structured every session. They set the tasks, they shaped the use, they were in the room the whole time. The AI was not handed to these kids the way we hand it to our kids. It was aimed.

And here is the number that matters most and gets quoted least: that exact tool, GPT-4 through Microsoft Copilot, was free to essentially every teenager on the planet during the same period. Hundreds of millions of them had it in their pocket. They did not gain two years of anything.

The tool was not the miracle. The tool was ambient. What made the difference was a human being deciding where the effort went.

We are not in an argument about whether AI is good or bad for learning. That argument is beneath the evidence, and both sides of it are being fought by people selling something.

The real question is narrower and much more uncomfortable: when the machine removes the difficulty, did anyone decide what the difficulty was for?

Sometimes the difficulty is genuinely waste. Nobody ever became wiser through formatting a bibliography. Whoever finally automated that should be given a knighthood, a public holiday and a statue in a prominent square, and I will personally hold the door, lay down my coat and grovel at their feet. Harvard referencing stole weeks of my life that I am never getting back, and it taught me precisely nothing except a lasting hatred of the semicolon.

But sometimes the difficulty is the entire point, and it is disguised as inefficiency. The head-on-the-desk half hour. The terrible first draft. The essay you had to write three times. The wine you could not read without building the map yourself. Automate those and you have not saved anybody time. You have deleted the gym and kept the mirror.

We are about to run this experiment on an entire generation, at speed, with no control group. That is happening whether or not anyone writes a thoughtful newsletter about it.

So the only question left is whether the adults in the room, the parents, the teachers, the managers, the ones who set the tasks, are going to make that decision deliberately. Or whether we are going to let a product roadmap make it for us, and find out what we deleted in about fifteen years.

Nigeria says it can go well. MIT says it can go badly. The variable in both cases was a human being.

Which is either the most hopeful thing about this whole situation, or the most demanding. I think it is both, and I do not think we get to pick.

If this is the kind of question you are trying to make sense of, it is the work I do. My book Full Stack Human is about staying human in a world run by technology, and the difference between the tasks worth automating and the ones quietly building your people sits right at the centre of my keynotes and workshops. If that is what your team is wrestling with, come find me.

Stay human. The world needs it.

— Tāne Hunter
Founder, Future Crunch
Co-author, Full Stack Human

P.S. If you read this on your phone in about ninety seconds and got the gist without really taking any of it in, that is fine, that is most reading, and I do it too. But you might notice you cannot quote a single line of it back. Which is, awkwardly, the entire point, and I have just done to you what ChatGPT did to those fifty-four students. Sorry about that. The handwritten note you take about it will stick better than this paragraph did.