Cognitive atrophy: The hidden cost of using AI the wrong way
AI amplifies your work habits, good and bad. On metacognition, curation, and five habits that build competence instead of eroding it.
We talk a lot about what AI can do for us. We talk remarkably little about what it does to us.
Most people who have used language models regularly over the past year recognise the feeling: you face a problem you would once have wrestled with, and the first reflex is no longer to think. It is to ask. The answer arrives in seconds, it is usually good enough, and you move on. Efficient? Undoubtedly. But something happens quietly when thinking is always outsourced: the muscle you do not use weakens. The phenomenon has a name, cognitive atrophy, and it deserves to be taken seriously, not as technology scepticism, but as a practical workplace problem.
Tool or sparring partner: the choice decides the outcome
There are two fundamental ways to use AI, and they have opposite effects on your own competence.
The first is the answer machine: you have a question, you get an answer, you use the answer. This is fast, and for trivial tasks it is exactly the right use. The problem arises when the answer machine also takes over the tasks that used to build your understanding: the reasoning, the trade-offs, the mistakes you learned from.
The second is the sparring partner: you use the model to challenge your own draft, ask for counterarguments, and test your understanding by explaining something back and having it picked apart. Here AI does the opposite of replacing thinking: it intensifies it. The difference is not in the technology, but in who does the cognitive work.
A simple test: if you could not defend the conclusion in a meeting without the screen in front of you, you have not learned anything. You have merely transported text.
Metacognition is the new core competence
If one skill separates those who grow with AI from those who wither with it, it is metacognition, the ability to think about your own thinking. In practice, it means being able to answer three questions before delegating a task:
What do I already know well enough to delegate safely? Routine tasks within your own field can be delegated without loss, because you are able to quality-assure the result. An experienced professional who lets AI write the first draft of a standard letter loses nothing: she would not have learned anything from writing it for the tenth time.
What am I currently learning, and should therefore do myself? This is where delegation is most dangerous. Anyone building a skill, such as a new procedure, a new field or a new language, who lets AI do the work pays with their own learning curve. Short-term gain, long-term debt.
What do I not know that I do not know? This is where AI as a sparring partner truly shines: using it to map your blind spots, not to hide them.
Curation beats consumption
Information overload is not new, but AI has industrialised the production of plausible text. Value therefore shifts from finding information to filtering it. Curation, the ability to leave things out, becomes a core professional skill on a par with source criticism.
In practice, this means building your own filters: a small number of sources you actually trust, a deliberate attitude to what you let into your working day, and a habit of asking "who wrote this, and why?" even when the answer comes from a model. Perhaps especially then.
The generalist rises
One of the more interesting shifts being discussed right now is what AI does to the value of deep specialisation. As models handle more and more of the specialised craft, the value of being able to connect fields increases: understanding enough about law, technology, economics and your own domain to ask the right questions across them.
This does not mean the specialist dies. It means the specialist who also masters adjacent fields through AI-supported learning gains a significant edge over the one digging ever deeper into a single shaft. Lifelong learning goes from platitude to competitive factor, not as courses you take, but as a way of working.
Five habits that build instead of erode
- Write the draft yourself first when the task lies within a field you are still developing. Use AI for critique, not production.
- Explain it back. Ask the model to quiz you on a topic you have just "learned". If you cannot answer without help, it was not learning.
- Ask for resistance. "What is the strongest argument against this?" is one of the most valuable prompts there is.
- Delegate deliberately, not reflexively. Take half a second before every question: is this something I should have thought through myself?
- Keep something analogue. Read long texts without summaries now and then. Stamina in thinking is also a skill that withers without use.
Conclusion
AI is neither a threat to thinking nor a guarantee of it. It is an amplifier, of good work habits and of bad ones. Those who treat it as an answer machine will find their competence gradually moving out of their heads and into the tool. Those who treat it as a demanding sparring partner will learn faster than any generation before them.
The difference is not technological. It is metacognitive. And it starts with a simple question the next time your fingers reach for the keyboard: who is doing the thinking here: me or the machine?