The First Profession to Forget How
Developers don’t want to work without AI anymore. The research says it might be making them worse — and they can’t feel it. Education should be paying very close attention.
ViZionary HoriZons
May 31, 2026
The First Profession to Forget How
Dr. Rockmore
In February 2026, a research lab called METR set out to do something that should have been routine. Repeat a study.
The year before, they’d run a clean experiment. Sixteen experienced open-source developers, each averaging five years on mature codebases, completed 246 real tasks, randomly assigned to allow or disallow AI tools. Tools had moved fast since then, so METR wanted to run it again and measure the gains.
They couldn’t. The developers wouldn’t participate, because they did not wish to work without AI — even briefly, even for the study.
On its face, that’s a success story. People don’t volunteer to give up tools that make them better. Nobody asks to trade GPS back for a paper atlas.
Except here’s what the original study actually found. Developers using AI took 19% longer to finish their tasks. Before they started, they predicted AI would save them 24% of their time. After it was over — with the data sitting in front of them — they still believed it had saved them 20%.
Read that again, because it’s the whole story.
The tool slowed them down. They felt faster. And the feeling didn’t just survive contact with the evidence — it barely moved.
That’s not a productivity story. That’s a perception story. And perception is my entire field.
When METR went back in February 2026, they admitted the new numbers were unreliable for a reason that should make every teacher sit up. The developers who found AI most valuable had effectively removed themselves from any condition that required working without it. The dependency had become invisible to the people inside it. They couldn’t run the experiment because the subjects could no longer locate the part of the work that was theirs.
If you want more like this, subscribe — this is what I build here every week.
I’ve watched a smaller version of this in my own classroom, and I suspect every teacher reading this has too.
A student generates five clean pages in ten minutes. The prose is tighter than a lot of adult writing. The formatting is professional. By every visible signal, learning happened. Then you take the machine away and ask for the same thing built independently, and what comes back is unrecognizable. The output was real. The understanding was never there.
What unsettles me isn’t the gap. It’s that the student often can’t see the gap any better than the developers could. They felt productive the whole time. The fluency of the tool got mistaken for fluency of the mind. That’s the same disconnect METR measured, just earlier in a life, where the stakes for a developing brain are higher.
For most of history, our tools amplified the body. The tractor amplified muscle. The forklift amplified strength. You never confused the forklift’s lift with your own, because the boundary was obvious — you could feel where your strength ended and the machine’s began.
AI amplifies cognition. And cognition doesn’t come with that boundary. When a tool drafts, suggests, organizes, and explains, it stops feeling like a tool and starts feeling like thinking. The line between what I can do and what the machine did for me dissolves, and we lose the one signal that used to tell us whether we were actually getting better.
This is the distinction I keep coming back to, and it’s the spine of everything I write here: amplification and construction are not the same process. Amplification raises your output today. Construction builds the capacity you’ll still have tomorrow when the tool is gone, slow, or wrong. A calculator makes arithmetic faster without making you better at mathematics. AI can make your output explode without a single new neuron of understanding being laid down underneath it.
And the software industry is now generating hard evidence that the costs are real, not theoretical. Independent researchers at Singapore Management University reported in April that AI output introduces long-term maintenance debt into real projects. The pattern has a shape: ship faster now, pay later in bugs that are harder to find than your own mistakes would have been.
The market is starting to notice. Amazon shut down an internal leaderboard that ranked employees by AI token usage after people gamed it by burning tokens to look productive, and Uber spent its entire 2026 AI budget in four months, with its own COO acknowledging it hadn’t produced a measurable jump in output. Token consumption turned out to be a proxy for activity, not value. Motion mistaken for progress. The whole economy is making the developers’ error at scale.
I want to be precise about what I’m claiming, because this is where most AI commentary gets lazy. AI is not the villain here. I use it every day, including in the building of this very essay, with my own disclosure stamped at the bottom. The problem was never the tool. The danger is not intimacy with tools. The danger is surrendered agency — handing over not just the labor but the judgment, and then losing the ability to tell that you’ve done it.
This is also why I don’t believe AI created the crisis we’re walking into. It’s accelerating one that was already underway. Between 2015 and 2025, reading scores fell across the large majority of U.S. districts, and the decline cut clean across income, geography, and race. It wasn’t lockdowns. Attention was already eroding, reading stamina was already collapsing, executive function was already thinning out under a decade of frictionless screens. The pandemic didn’t start that fire. It was the stress test that revealed how much had already burned. AI now arrives as the most frictionless cognitive offload device ever built, handed to a generation whose construction capacity was already running low.
So the carpenter image holds, but I’d sharpen it. The hammer doesn’t know which wall is load-bearing. It amplifies the carpenter’s skill and contains none of his judgment. The danger isn’t that we picked up a powerful hammer. It’s that we’re forgetting we were ever supposed to know which wall holds the house up — and the hammer feels so good in the hand that we don’t notice the forgetting.
Human beings are the CEO. AI is the amplification architecture. A CEO who can no longer evaluate the work coming up from below hasn’t been freed. They’ve been quietly replaced, and the org chart still has their name on it.
The developers couldn’t feel themselves getting slower. That’s the line I can’t shake. Not that they got slower — that they couldn’t feel it. The most measured, technical, data-literate workforce we have looked straight at a 19% slowdown and read it as a gain.
If they can’t feel it, the teacher mid-lesson can’t reliably feel it either. And the fourteen-year-old, three pages into something a machine wrote in his voice, has no chance of feeling it at all.
The first profession to experience cognitive dependency at scale may be software engineering. The first institution forced to deal with its consequences may be education. That’s the assignment for everyone building the next decade of this work. The question of the AI era was never how much output we could produce. We’ve answered that one. The real question is quieter and far more dangerous, because by the time it matters most, we may have lost the instrument that measures it.
Did learning still happen — and would we even know if it didn’t?
If you’re new here, subscribe so you don’t lose this thread.
https://drrockmore.substack.com/p/welcome-to-dr-rockmores-vizionary
Drafted with AI assistance; final voice, claims, and edits by Dr. Rockmore.
© 2026 Dr. Clay “Dr. Rockmore” Stidham / ViZionary HoriZons, LLC. All rights reserved. ICI© and all related original frameworks are protected intellectual property. Unauthorized use, reproduction, or adaptation is prohibited.
Tags:
AI Literacy, Education, Human Agency, Cognitive Architecture, Software Development, Artificial Intelligence, ICI, Brain-Based Living, Future of Work, ViZionary HoriZons