Deskilling: What We Unlearn When AI Always Has an Answer
A study in Nature Human Behaviour shows: the longer people interact with AI, the stronger their cognitive biases become, and they don’t notice. What that means for research, medicine, and anyone who works with AI today.
At some point about a year and a half ago, I realized I had delegated too much.
I had reached a point where I was using AI for practically every step of my work: literature searches, summaries, first drafts, code snippets, replies to difficult emails. It was fast. It worked. And then, one evening, I wanted to write a section myself and noticed that it felt harder than it used to.
That is not the story you usually read about AI. But it is a more honest one.
What the research shows
A recent study in Nature Human Behaviour examined what happens to cognitive GlossaryBiassystematic skew in data or outputs.Go to entry ↗ when people interact with AI over an extended period. The result is uncomfortable: the biases increase. And the people affected are not aware of it.
That is the truly unsettling part: not that the rate of bias rises, but that it happens blindly. Automation bias, the overweighting of AI outputs relative to one’s own judgment, is not corrected by experience. It is reinforced by habit.
Ironically, the more experience someone has with a system, the more they trust it, regardless of whether the trust is deserved.
What this means in practice
In science, I have observed this myself. The doctoral students I mentor use AI for tasks that used to be part of the learning process: formulating, structuring, contextualizing. The result arrives faster. It often sounds better. But when you ask why they interpreted a section the way they did, sometimes no answer comes. Because they didn’t interpret it; the model did.
In medicine, the same pattern is visible, with higher stakes. Physicians who use AI-assisted diagnostic systems begin to calibrate their own clinical assessment against the machine. When the model says it’s probably jet lag (as an early ChatGPT system did in a classic case of pulmonary embolism), it takes judgment to contradict it. That judgment comes from years of clinical experience. Anyone who doesn’t have it can’t apply it.
That is the paradox: AI is most dangerous where you can least do without it.
The skill curve, redrawn
A career in research and medicine used to look like this: many years of fundamental work, many mistakes, a lot of repetition. And then, at some point, the experience from which judgment emerges. The entry-level tasks were tedious, but they had a function: they trained you.
AI automates precisely these entry-level tasks first. Literature search, protocol drafts, data cleaning, first formulations. That is efficient. And it may deprive the next generation of the practice room that eventually qualifies them to judge the AI.
Put simply: someone who has never learned what a good statistical analysis plan looks like can’t judge whether the analysis plan ChatGPT generated is good. The judgment presupposes the competence that training produces. AI delivers the result, but not the training.
Two kinds of loss
There is a difference between short-term comfort and long-term competence.
The short-term comfort is real: AI does in minutes what used to take hours. The efficiency gains are measurable and significant. It would be wrong to play that down.
But there are two kinds of competence that can be lost along the way:
Methodological competence: the craft. Knowing how to run a clean literature search. How to explore a dataset. How to structure a paragraph that actually says something. This competence atrophies when it isn’t used, in experts too, not just in beginners.
Meta-competence: the ability to judge whether the result is good. This is the critical one, and it is the one AI can replace least but presupposes most. You have to be able to do something in order to judge whether the machine did it well.
What this does not mean
It would be wrong to conclude that using AI is itself the problem. That is not the message.
The message is: use AI as a tool that builds on competence, not as a substitute for it. Know, and be able to name, the difference between “I understand what the model did” and “I adopted the result.” And recognize that efficiency and competence only come together if you cultivate both.
I use AI every day now. For research, help with phrasing, coding assistance, ideation. But I have deliberately taken back tasks I used to delegate. Not because that is more efficient. But because my judgment lives on practice.
AI can’t think for me. It can show me what someone else would have thought. Knowing the difference is no small thing.
What to do now
In the short term: learn before you delegate. AI is not a teacher. It is a tool for the competent. Anyone who doesn’t yet have the fundamentals should acquire them another way first.
In the medium term: verify and understand outputs. Don’t just check whether the result looks plausible, but whether you can explain and defend it. That is a very different question.
In the longer term: institutions have to take this seriously. Universities, clinics, and research institutions that integrate AI into their workflows without asking what training the next generation will get risk a silent process of competence erosion that will only show up in ten years.
Anyone who asks the question now can still avert it.
Oscar is a postdoctoral researcher at the Else Kröner Fresenius Center for Digital Health (TU Dresden) and a co-founder of Denkmal.