On deskilling, mis-skilling, and never-skilling

I remember the book Future Shock (1970) being on the bookshelves in our house growing up, part of my dad's Master's level studies into educational management.

In it, Alvin Toffler argued that education should teach people how to "learn, unlearn and relearn". A few years later, Harry Braverman published Labor and Monopoly Capital, which gave us the "deskilling" thesis.

I find it interesting that in some situations we talk about "unlearning" and see it as positive, and in other situations discuss "deskilling" and see it as negative.

If I stop doing long division now I've got a calculator, have I "unlearned" something or been "deskilled"? It depends, I suppose, on who's telling the story. Calls for unlearning lead to being booked in keynote slots at conferences, while discussions around deskilling generate moral panics.

---

I've written before about pejorative terms, pointing out that "AI slop" is sloppy, and suggesting that so-called "AI psychosis" isn't much better. Also, in an essay for FIELD STATION, Tom Watson and I argued that what some call "cognitive surrender" looks closer to "an abdication of judgement". In general, I think that when people aren't sure about the strength of their argument, they reach for a loaded label.

The MIT paper "Your Brain on ChatGPT" is a good example of this. As I pointed out at the time, the title channelled the famous 1980s "This Is Your Brain On Drugs" advert in the US, with the paper's headline findings resting on 18 people writing 20-minute essays. Nearly a year and a half later it's still a preprint, and a comment posted to arXiv in January suggests it might be worth reading its supposed results quite conservatively.

---

"Deskilling" is more than a label, though. Harry Braverman's point wasn't that losing a skill is bad in itself. It was about who gets to decide. If you decide not to do something any more, that's one thing; if management automate part of your job without your consent, that's something different entirely.

So here's my (slightly clunky) working definition: unlearning is deskilling you choose, of a skill that no longer provides any value to you.

The second part of that, the bit about "value", is where it gets interesting. Lisanne Bainbridge's 1983 paper "Ironies of Automation" showed that automating routine work leaves people with the rare, hard cases. Ironically, at the same time, it takes away the practice that kept them good enough to handle those cases in the first place.

The classic example here is pilots and their use of autopilot. The skill of flying an aeroplane hasn't gone out of date; it's hidden until the moment you need it.

There's now better evidence in this area than that provided by the MIT study. Endoscopists are medical professionals who put cameras on tubes into various parts of your body. A study in Poland showed that, once they got used to AI-assisted colonoscopy, their adenoma detection rate without the AI fell from 28.4% to 22.4% (Budzyń et al., 2025). While it was an observational study, and the overall cancer detection levels didn't seem to drop, it's nevertheless an interesting finding.

Meanwhile, a field experiment with nearly 1,000 high school maths students in Turkey (Bastani et al., 2025) found that unrestricted access to GPT-4 made practice sessions go better. Once access was taken away, however, those students did worse in exams than students who'd never had it.

A version set up to tutor the maths students rather than simply hand over answers mostly avoided the negative consequences. And, for me, that's the compelling finding. The design of the tool matters as much as whether students use it at all; the problem is often a badly designed task, not the student.

---

Interestingly, medical educators are now separating deskilling (losing a skill you had) from mis-skilling (picking up the AI's errors) and never-skilling (never building the foundations at all) (Abdulnour et al., 2025).

Never-skilling is the AI-era version of Toffler's question: if there's lots we'll no longer need to learn, how do we know which bits? You usually need a skill to judge whether the AI has done its job well – which is why people are usually wowed by examples of AI use outside of their domain of expertise, but not so much within it.

A survey of knowledge workers by Microsoft and Carnegie Mellon (Lee et al., 2025) found that people who trusted AI answers more reported less critical thinking. People who trusted themselves more reported more.

Charles Leifer, whose post I shared earlier this year, gets right to the heart of it. For him the real problem is AI's "inability to distinguish between difficulty worth keeping, and difficulty worth removing" (source). I'd add that this isn't unique to AI: people aren't always good at telling the difference either, especially when someone else is making the call for them.

So, in conclusion , I'd argue that AI isn't making us stupid, and not all deskilling is bad. Some skills rightly deserve to go. The useful questions are, as ever, the boring ones: is this skill still valuable, and who gets to decide when it isn't?

---

Image: Asemic sand writing via Are.na