Research published this year has given a name to something employers have been circling for a while. Deskilling is what happens when an expert stops practising and gets worse. Never-skilling is what happens when a novice never gets good in the first place, and it is the more awkward problem, because the people it affects are the ones companies are already hiring fewer of.
The sharpest evidence comes from a randomised controlled trial run by Anthropic researchers Judy Hanwen Shen and Alex Tamkin, published in January.
They recruited 52 mostly junior software engineers, gave half of them an AI assistant, asked all of them to learn Trio, a Python library none of them knew, and then quizzed everyone on the concepts they had used minutes before.
The AI group averaged 50%. The hand-coding group averaged 67%. Anthropic describes the gap as the equivalent of nearly two letter grades, and it was statistically significant, with a p-value of 0.01.
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!
The speed benefit, which is the entire reason anyone reaches for the assistant, did not really materialise. The AI group finished about two minutes faster, a difference that failed to reach significance, partly because several participants spent up to 11 minutes composing queries, roughly a third of their allotted time.
They learned less, finished no faster, and came out worst on the thing that matters most when the machine is wrong. That thing is debugging, where the gap between the groups was widest. The control group, denied an assistant, hit errors and had to resolve them, which is a fair description of how debugging is learned. The AI group did not hit the errors.
Medicine has arrived at the same worry from a different direction. A Nature Medicine Perspective published in May, led by Duke-NUS Medical School with co-authors at Harvard, UCL, and King’s College London, coined it for trainees who lean on AI during their formative clinical years and never build the reasoning that safe, independent practice requires.
It adds a third category with even less attention on it: mis-skilling, the trainee who accepts an AI error uncritically and files it away as fact.
Those authors are careful in a way the coverage of them has not always been. Direct evidence from medical training, they write, is absent. The argument rests on learning theory and on early signals from non-clinical settings, which is to say from studies like Anthropic’s.
Their prescription is a three-phase framework: build competence without AI, then teach people to calibrate their scepticism, then introduce the tools under supervision.
How the tool is used matters more than whether it is used. In the Anthropic trial, the high scorers asked conceptual questions or requested explanations alongside the code. The low scorers delegated wholesale, or leaned on the assistant to debug for them.
Employers are already pricing this in. Gartner predicts that critical-thinking atrophy will push half of global organisations to require “AI-free” skills assessments through 2026, which is a polite way of saying that hiring managers no longer trust a portfolio.
Ford, meanwhile, has been rehiring engineers to fix what its AI systems got wrong, an expensive demonstration of what happens when the people who could have caught the error are no longer on the payroll.
The trial comes with real limits, and its authors say so. The sample was small, the quiz measured comprehension immediately rather than months later, and it used a sidebar assistant rather than an agentic coder. The researchers expect the impact of those to be more pronounced, not less.
It is worth noticing who ran it. Anthropic sells the assistant, and it has published a paper arguing that using the assistant carelessly makes you worse at your job. That is either unusual candour or the opening of a pitch for learning modes, and both readings can be true.
What the research does not say is that juniors should code by hand. What it says is that the shortcut and the skill are not the same road, and that the industry has spent two years assuming they were.
