This episode takes the deskilling trap from chapter six of the book (“The Collaboration Demands More of Me, Not Less”) and collides it with what the research is saying right now: a Nature news feature on early deskilling evidence across medicine and software engineering, the Lancet colonoscopy study behind it, and the survey showing 70 […] Below is the full script of this episode as an article.
I want to tell you about a group of doctors in Poland. Specialists. Endoscopists… the people who run colonoscopies and read what they’re seeing in real time. Every one of them had at least two thousand procedures behind them. These are not beginners. This is about as far from a beginner as medicine gets.
They get a new AI tool that watches the video feed with them and flags precancerous growths as they work. Good tool. It genuinely helps. But here’s the design choice that makes the study interesting: the tool is available on some days and not on others. So the researchers can watch what happens to the humans when the machine goes quiet.
Before the AI arrived, these specialists were finding at least one precancerous growth in about twenty-eight percent of procedures. After three months of working alongside the tool, on the days it was switched off… twenty-two percent.
Three months. That’s all it took for measurable skill to leak out of some of the most experienced hands in the field. And nobody felt it happening. That’s the part that should stop you. Not one of those physicians woke up and thought, I’m getting worse at this. The decline only showed up because somebody was measuring the off days.
The evidence
That study ran in The Lancet’s gastroenterology journal, and it’s the centerpiece of a Nature news feature from a few weeks ago asking, more or less… is AI ruining our skills? Their reading of the early evidence: the erosion is real, and it’s showing up in physicians and software engineers first because those are the fields somebody bothered to measure. The people doing the work can feel it coming, too. In one survey of American health care workers this year, seventy percent of nurses and seventy-seven percent of physicians said they’re worried about losing skills to AI reliance.
Now, boundaries first, because a scary finding deserves them. That colonoscopy result is one study. One country, one specialty, three months. One of the study’s own authors says it needs replication before anyone should treat it as settled. Fine. Hold it loosely.
But here’s the thing… we’ve known the shape of this pattern for over forty years. In nineteen eighty-three, a researcher named Lisanne Bainbridge published a paper about industrial automation with a phrase that has never stopped being relevant: the irony of automation. You automate because humans are unreliable. Then the automation removes the practice that kept the humans sharp… which makes them less reliable. The better the system, the faster the backup skills rot.
AI didn’t invent that trap. It industrialized it. It made the trap fluent, friendly, and available to every knowledge worker on earth for twenty bucks a month. And with AI there’s a twist Bainbridge didn’t have to deal with. When a pilot stops monitoring the instruments, eventually an alarm gets missed and everyone knows. When a professional stops engaging and just ships what the machine produces, the work still looks polished. The quality gap hides behind the polish. The degradation is invisible until someone with real domain expertise reads the work carefully… or until the tool goes down and there’s nothing underneath.
The effort is the point
This is the heart of chapter six of the book. The chapter title is the argument: the collaboration demands more of me, not less. Which is exactly backwards from how AI gets sold, right? Easier. Faster. Less effort. And for routine work… file conversion, boilerplate, the mechanical middle of the job… sure. Effort saved is the whole point.
But for work your judgment actually touches, the evidence keeps pointing one direction. The clearest version is an experiment at Wharton with nearly a thousand students. Two groups, same AI, same material. One group got open access. Ask anything, take the answers, move fast. The other group got a scaffolded version, built to make them do their own thinking first and use the AI as a tutor instead of an answer machine.
During practice, the easy path improved forty-eight percent. Sounds great, honestly. Then the hard path came in at a hundred and twenty-seven percent. More than double.
And then the researchers took the AI away. The easy-path group didn’t just lose their gains. They landed seventeen percent below where they’d started. Worse than if they had never touched the tool. The scaffolded group kept everything they’d gained. Zero deskilling.
Same tool. Same material. The only variable was whether the collaboration demanded engagement or let people coast.
I’ll give you the caveat the book gives, because it matters. Most of that scaffolding evidence comes from classrooms… students, controlled tasks. The claim that the same mechanism protects a working professional is an extension, supported by the direction of the evidence but thinner on direct data. Though the studies closest to real practice point the same way. There’s a clinical workflow study where physicians had to commit to their own diagnosis in writing before they saw what the AI thought. More effort, more time, more of that uncomfortable moment where you’re wrong on paper. Accuracy went up. The structure kept the human’s judgment in the loop, and the outcomes showed it.
Which is why the Nature piece landed for me the way it did. One of the colonoscopy study’s authors says there is no established solution against deskilling right now. And he’s right that nothing is proven at scale. But I don’t think the cupboard is empty. The early shape of a solution has been sitting in the engagement research the whole time: structure the collaboration so your judgment shows up before the machine’s does. The effort isn’t friction to eliminate. The effort is what keeps the skill alive.
In my own work
Let me show you what this looked like in my own work, because I got a forced lesson in it while finishing the book.
Before the manuscript went to print, I did something that felt a little masochistic. I took every argument in it… fifty-eight of them… and made two different AI models attack them. Not summarize. Not polish. Attack. Find the weak evidence. Name the prior research I was accidentally rebranding as new. Tell me where a claim wouldn’t survive a knowledgeable reader.
It was slower than not doing it. It was so much more effort than not doing it. Some arguments came through validated. Most got sharpened… scoped, qualified, rebuilt. And one of my favorites failed completely. The reviews said, in effect, this isn’t evidence, it’s a sample of one wearing a lab coat. They were right. I tore it down and rebuilt it as illustration instead of proof.
Here’s what I want you to notice about that. It’s the same class of tool everyone is worried will make us soft. Pointed one way, it will… it hands you fluent answers, you stop practicing, and three months later the off days tell the real story. Pointed the other way, that review put my judgment through more reps than any human editor ever has. Same machine. The difference was the structure of the collaboration, and who had to commit to a position first.
I run the same rule in the fractional marketing work I do through Auspicious now. Before a model touches a strategy question, I write my own call down. Then it argues. Then I rule. Slower on purpose.
What to do on Monday
So what do you do with this on Monday?
One researcher in that Nature piece said something quietly useful: just knowing deskilling exists should provoke some self-reflection about which skills you want to maintain, and which you’re willing to hand over. That’s the move. Make it a decision instead of a drift.
Pick the skills that are load-bearing for you… the judgment your clients or your company actually pays for. For those, work think-first. Commit to your own answer before you ask the machine. Every time. And once in a while, run a day with the tool off. Not as a purity test. As a diagnostic. The doctors in Poland didn’t know what three months had cost them until somebody measured the off days. You can measure your own.
Everything else? The boilerplate, the formatting, the mechanical middle… let it go. Deliberately. Deskilling in a skill you’ve consciously retired isn’t a trap. It’s delegation.
One idea from the book. One collision with what the research is saying right now. That’s the episode.
—
This show extends the book AI Empowered. Read the first chapter free at https://www.aiempoweredbook.com, or get the book on Amazon.
Aaron Douglas runs Auspicious, a fractional marketing practice.
Blog: https://auspicious.llc/insights/
X: https://x.com/aaron_douglas
LinkedIn: https://www.linkedin.com/in/aaronddouglas/
