Is AI making us dumber? AI skill atrophy explained

Published:

What happens to the skills we stop practicing, and why you won’t notice until it matters

Consider a small experiment: recite a phone number known by heart.

Chances are it belongs to a landline in a house moved out of years ago, learned before smartphones took over. Nothing picked up since has stuck, because nothing since has required it.

Nobody chose to give up memorising phone numbers. There was no decision, no weighing of pros and cons, no farewell to a skill. The habit faded, and then the ability faded with it.

That is what makes handing a skill to a machine so deceptive. It does not register as a loss. It feels like relief: one less tedious chore.

What is AI skill atrophy?

As explained here, every technology spares people some effort. That is its whole purpose. The real question is what happens when the effort stops altogether, and whether that is a step back or a step forward.

There is plenty of vocabulary for the economic side: automation, technological unemployment, reskilling. What is missing is a way to describe what happens inside a person who stops practising a skill because a machine does it, instantly and without complaint.

In a forthcoming book on AI’s changing role in society, this phenomenon is called AI skill atrophy: the gradual weakening of a human, especially intellectual, capability caused by repeatedly outsourcing its practice to artificial intelligence.

Astronauts offer a parallel. In microgravity their muscles have little to push against, so without deliberate exercise they waste away. An exoskeleton works similarly. Worn on Earth, it allows far heavier lifting than would otherwise be possible. Worn daily for a year, it leaves the wearer’s own muscles quietly adjusted to the arrangement. The work still gets done, but the body underneath grows weaker, right on schedule.

AI can be seen as an intellectual exoskeleton. It grants cognitive superpowers in the short term and makes possible things that could not have been imagined before. But it may also be wearing down the very muscles that built those abilities.

The loss is not intelligence. It is any capability handed over faster than it is practised: in work, in study, in judgement, and in the everyday knack of sensing that something is off.

When AI does the thinking

Not long ago, an Oxford PhD student submitted a research proposal for feedback. It was excellent. A video call followed to discuss it, and the questions began.

Even the basic ones proved hard to answer. The problem was not a misunderstanding of the subject. The student simply could not retrace how the ideas on the page had come about. The proposal was polished, but when the reasoning was drawn out of it, there was surprisingly little behind it.

The case is not isolated. Researchers are finding the same gap in AI-assisted work.

>>>  The first A.I. using imagination

In the widely cited MIT Media Lab study “Your Brain on ChatGPT”, 54 participants wrote essays with an AI assistant, with a search engine, or with no help, all while wearing EEG caps. Brain connectivity dropped as the level of assistance rose: strongest among those working alone, weakest among those using AI. The AI group also had trouble quoting from essays finished only minutes earlier.

The essays existed. The words existed. But less of the thinking had passed through the person who produced them.

Behind these results lies an ordinary mechanism: cognitive offloading. Humans have always done it. A shopping list is offloading. So is a calendar, a knot tied in a handkerchief, or a spouse who remembers everyone’s birthday.

Researchers at Microsoft and Carnegie Mellon surveyed 319 knowledge workers about 936 real uses of AI at work. The more workers trusted the AI, the less critical thinking they reported doing.

The finding is intuitive. When an output looks good, the urge to challenge it weakens. And when questioning, checking, and working through problems stop, so does the exercise of the skills needed to judge whether the output is any good. It is a vicious circle, and after two or three years of working with LLMs, it is a familiar one.

The question, then, is what the price of short-term superpowers might be if the underlying muscles are left to waste away.

The atrophy ladder

The pattern runs in five steps. Once they are visible, it becomes clear which rung a given skill occupies.

  1. Automation. A machine becomes capable of doing the task.
  2. Outsourcing. The task is handed over, for a perfectly good reason: a deadline, a client, a busy Tuesday.
  3. Under-exercise. The know-how remains, but the practice stops. This stage is invisible and can last for years.
  4. Atrophy. The ability degrades through disuse. It is discovered, if at all, the day the tool is unavailable.
  5. Dependency. The tool is no longer a convenience but a prerequisite. Without it, even starting seems impossible.

Most commentary leaps from step one to step five and declares the outcome inevitable. It is not. Step three is where control can still be retaken.

The ladder can be applied to any ordinary working week: drafting, finding sources, structuring an argument, reading something long enough to disagree with it, recalling what was on the mind before the chat window opened. Some of those skills will be let go gladly. Others will be wanted back, and their absence will go unnoticed until the afternoon they are needed.

Outsourcing thinking, understanding, and decisions

Andrej Karpathy, a co-founder of OpenAI, recently shared a neat formulation: thinking can be outsourced, but understanding cannot.

It is elegant, but it does not go far enough. Understanding can be outsourced. So can decisions, and anyone who has used the latest AI models has already done so.

>>>  Why AI tech leaders are asking to pump the brakes

Mark Turrell, author of Scaling and founder of the unDavos Summit, makes the point concrete. Asking Claude Code to build a professional website outsources more than the coding. It outsources thousands of small design choices. Where does the button go? How much space between sections? Should that heading be bigger? Which colour works better? Navigation on the left or on top?

The AI answers all of it. If the result looks good, the user is satisfied and moves on without spending any more mental effort on those choices. In Turrell’s view, that is how more and more daily work gets handed over: one decision at a time.

The muscles already abandoned

Before considering what AI might take, it is worth noting what is already gone.

Memorisation went to search engines. Navigation went to GPS and mapping apps. Mental arithmetic went to the calculator. Spelling went to autocorrect, so thoroughly that many people cannot say whether they can still spell. Now brainstorming, research, coding, translation, drawing, and structuring an argument are following. Even remembering a train of thought from twenty minutes earlier can be delegated, with tools like Goldfish that keep the thread of the reasoning and hand it back on request.

None of these handovers was a mistake. Most were clear improvements.

But they share a feature. Each was something done constantly, not because the practice was valued, but because there was no alternative. The practice was a by-product of necessity. Remove the necessity and the practice disappears, silently, with no one deciding anything.

Here it helps to split “skill” into three things, using the PhD student as an illustration.

Knowledge is knowing what a proposal needs: a premise, a theoretical foundation, a set of research questions. Application is assembling those parts into something coherent. Capacity is being able to do it at eleven at night before the deadline, when the material resists and the first three attempts fall apart.

AI leaves knowledge nearly untouched. That is why the MIT result is so striking: the participants had not forgotten how essays work, they just could not quote their own. The structure of a good literature review, the shape of a clean function, the way a paragraph should turn all remain familiar. What erodes is the layer beneath, and it erodes without disturbing the visible layer, which is exactly why it goes unnoticed.

The likely outcome is not a generation that knows less. It is a generation that knows how things are done but cannot do them without help. That is a different kind of loss, and a much harder one to capture in longitudinal studies.

How the brain decides what to keep

Two further stories come from London.

To drive a London black cab, candidates must pass an exam called the Knowledge: around twenty-five thousand streets, memorised without a map. It typically takes three to four years.

>>>  A "pain axis" inside AI models

A University College London team led by neuroscientist Eleanor Maguire scanned the brains of 79 trainees before they began and again years later, by which point 39 had qualified. Those who passed had grown more grey matter in the posterior hippocampus, the region tied to spatial memory. Those who dropped out had not. Four years of struggle had physically rebuilt the organ doing the struggling.

That is the famous half of the finding. The half rarely mentioned is that the drivers who qualified did worse than controls on certain other memory tasks, such as recalling complex visual information. Trainees who failed showed no such decline.

Their brains had not simply gained a capability. Over time, they had changed what they were good at.

The finding echoes Sherlock Holmes in A Study in Scarlet, where he tells Watson that the mind is like a small, empty attic, to be stocked only with furniture chosen deliberately. The skilled workman, he says, keeps only the tools that help his work, and, as he puts it, “for every addition of knowledge you forget something that you knew before.” (Chapter 2)

Holmes was onto something real: what is practised shapes what a person becomes good at, and what is no longer practised can fade because it is no longer needed. He was ruthless about what he let into his attic, which is why Watson was astonished to learn he knew nothing of basic astronomy, literature, or philosophy. Holmes was not trying to know everything. He was sharpening the skills his work demanded.

That is also the promise of AI. More of the lower-level work, such as remembering, searching, calculating, and summarising, can be handed off, leaving more time for judgement, reasoning, and deciding what actually matters.

The difficulty is that there is no villain here. Nobody is taking these abilities away. They are given up one reasonable convenience at a time, and there is never a moment of surrender to resist. There is only a Tuesday afternoon when work is behind, the shortcut is sitting right there, and refusing it would, that afternoon, be a slightly foolish choice.

A final thought: delegate on purpose, not by default.

The answer is not to give up the tools, but to stay aware of what is being handed over. Finding the way without a map now and then, writing a first draft unaided, or working through a problem before asking for the solution keeps a skill alive instead of letting it fade. Delegating is fine, as long as it remains a choice. This matters beyond the individual. A generation that inherits only delegated skills, and never practises an ability it could still use, may find itself without it on the day it is needed. A skill used occasionally is a skill preserved.

Related articles

Recent articles