Wheelchair
or Gym.
Jensen Huang says humans will soon write just 1% of the internet. Two months ago I wrote that the bots had already passed us in traffic. So the number didn’t shock me — the question behind it did. Today I can still tell AI writing from human writing. My experience tells me that skill is expiring. And once the internet stops being ours, the real question isn’t about the internet at all. It’s whether the human brain follows it down.
How soon is “soon”
Nobody measures who wrote the internet directly, so triangulate. Automated traffic crossed roughly half of everything around 2023–24 — the bots passed us while we were arguing about whether they would. Estimates for AI-generated content have run as high as 90% “within a few years” — and those estimates predate agents that write entire websites on their own. Huang’s 1% isn’t a wild forecast. It’s an extrapolation of a curve we’re already on: machine output scales without limit, human output doesn’t. My honest read is functionally within three to five years — not because humans stop writing, but because we stop being statistically visible.
The minority dialect
Here’s the operator confession that makes it concrete: we already write for machine readers. Every article this site publishes ships with an entry in a file called llms.txt — a version formatted specifically for AI to read, because we know who’s actually doing the reading now. The 1% future isn’t a prediction we’re waiting on; operators are quietly building for it today. Human writing is becoming the internet’s minority dialect — still spoken, no longer the language of record.
And the skill of telling the difference? It won’t die because people get worse at it. It dies because the signal itself is disappearing — detection tools keep losing, watermarks keep failing, and every model generation closes the gap a little more.
The Flynn effect ran on machines
Now the deeper question. Through the 20th century, measured IQ rose about three points a decade — the Flynn effect. The best explanation is uncomfortable for anyone afraid of automation: machines took the muscle work, and pushed humans into symbolic work. School instead of field. Abstraction instead of labour. The biggest intelligence gain in recorded history happened because technology took jobs.
Then, from the 1990s, the curve turned. Nordic conscript data shows the reversal clearly, and the pattern points to environment, not genes — the same families, declining scores. My generation blamed the PC. I’d sharpen that: it wasn’t the computer as a tool. It was the screen as a pacifier — consumption replacing construction, scrolling replacing reading.
Wheelchair or gym
So AI genuinely can go either way, and the fork is not in the technology. It’s in the mode of use.
AI as wheelchair: it thinks, you scroll. The cognitive-offloading research is not reassuring — GPS users lose navigation ability, calculator generations lose arithmetic, and writing is far closer to thinking than either. Hand the machine your thinking and the reverse Flynn effect gets a turbocharger.
AI as gym: the machine takes the grunt work — the drafting, the formatting, the hundred mechanical steps — and you climb to the layer it can’t own: judgment, taste, deciding what’s true and what matters. That is exactly the mechanism that produced the original Flynn rise. The first industrial revolution freed human muscle, and humanity chose mass education. The choice worked.
The honest problem: the wheelchair is the default. It takes no effort, and the tools are optimised for engagement, not for your development. The gym has to be chosen — deliberately, daily, against the grain of every app’s design. A national program won’t make that choice for anyone; it happens one user at a time.
Keeping the reps
What I actually do, and what I tell my team: let AI produce the first draft of everything — then the decision stays human, and the decision is the rep. Read the output critically instead of forwarding it. Write the hard paragraph yourself when it matters. Ask the model to argue against you, not for you. And watch one signal in yourself honestly: when AI gives you an answer, do you check it — or just feel relieved? The day you stop checking is the day the wheelchair arrived.
Published by IMA AI — August 2026. Written by someone who already publishes for machine readers — the llms.txt in our repo is the receipt — and still writes the hard paragraphs himself.