Start here
The tool is down until Thursday. What stops?
Not a thought experiment. Tick the work that would halt, and leave the work that would carry on slower and uglier but would still get done. There is no score. The list is the answer.
Tick the ones that would stop.
How it turns up
Nobody noticed until the Tuesday it mattered.
The provider had an outage on a Monday morning. About ninety minutes. At Nina's company the work did not restart on Monday afternoon. It restarted on Wednesday.
Nina runs a team of eleven analysts. She said the strange part was not that people could not use the tool. It was that once it came back, three of them could not tell her whether the output they got was right. They had checked it. They had signed it off. And when she asked what they had actually looked at, the answer came slowly.
“They can all still do the work. I am no longer sure any of them could catch it if the work were wrong.”
Nina has been in the job nine years. She could catch it. So could the two people who joined before 2023. That is the whole list, and she only found that out because a server went down in another country.
Composite. Real details, more than one source.
What it is
It is not your hands. It is the picture in your head.
Everybody expects the wrong loss. The fear is that people will forget how to do the work. The research says they do not. They can still operate. What they lose is the sense of where the work stands and whether the answer in front of them is any good.
That is the worse version, because you cannot see it. Somebody who cannot do a thing at all is obvious within a week. Somebody who can still do it but can no longer judge it looks exactly like somebody who is fine, right up until the day they approve something they should have stopped.
In three lines
- Sixteen airline pilots flew a 747 simulator with the automation dialled up and down. Their instrument scan and their manual control were unimpaired.1
- What had degraded were the thinking tasks: holding a picture of where the aircraft was, knowing what came next, noticing when the situation had changed.1
- The regulator's answer was not more training. It was regular practice with everything switched off.2
You are probably here because
- Somebody signed off work they could not have produced themselves
- An outage cost you days rather than hours
- A new joiner asked how it works and nobody could say it without opening something
- You use a lot of this and would rather find the edge than meet it
What you take away
- The one study that found the opposite of what everyone assumes
- A ten-minute test that shows you where the edge is
- One practice, four days a year, borrowed from aviation
- What it is honestly not possible to prove yet
What the research found
Sixteen pilots, one simulator, and a surprise.
Stephen Casner and three colleagues put sixteen airline pilots in a Boeing 747-400 simulator. They flew routine flights and flights where something went wrong. The researchers turned the automation up and down on purpose, graded what the pilots did, and asked them along the way what they were thinking about.1
The expectation was that hand-flying would have decayed. It had not. The instrument scan was intact. The stick and rudder work was intact. What had decayed were the thinking parts: holding a picture of where the aircraft was without looking at the map, knowing what came next, noticing when the situation had changed. The researchers put it plainly. Cognitive skills fade faster than physical ones.
Can still operate the toolintact
Can still check the outputfaded
Can still say what happens nextfaded
Can still teach it to somebodyfaded
Two caveats belong here, because sixteen people is not many. It was a simulator rather than a line flight, and it was aviation rather than your office. The researchers measured pilots. Applying it to a company is this page's step, not theirs.
What makes the step worth taking is the response. In February 2022 the American aviation regulator published an advisory circular telling airlines to give pilots regular practice with everything switched off: flight director off, autopilot off, autothrottle off.2 A regulator does not write that for a problem nobody has.
Why it happens quietly
Every single choice along the way was the right one.
Nobody wakes up and decides to stop understanding their own work. It goes one sensible decision at a time. The model drafts it, so you edit instead of write. Editing is faster, so you do more of it. Doing more of it means you never sit with a blank page. Sitting with a blank page was the thing that built the judgement you are now using to edit.
The skill was never in the output. It was in the part you skipped.
Checking goes the same way. When the first fifty things you check turn out to be fine, the fifty-first gets less attention, and you will not notice the moment your attention changed. That is the mechanism behind automation bias, and this is what it leaves behind after it has run for a couple of years.
Where this bites hardest
Not in the people who never learned it. In the people who did, and stopped. Somebody who never wrote a proposal by hand does not know what is missing. Somebody who wrote two hundred of them and has not written one in eighteen months believes they still could, and is the person you rely on to catch the mistake.
The test
Ask somebody to explain it without opening anything.
Pick a piece of work your team does often and does well. Ask the person who does it to talk you through how it is made, start to finish, without opening the tool and without opening an example. Ten minutes.
If they can, the knowing is still there. If the explanation keeps reaching for something to point at, you have found the edge of it. That is not a performance problem and it should not be handled as one. It is a use-it-or-lose-it problem, and the person in front of you is usually relieved that somebody noticed.
Where it starts to bite
You need a lot of AI in the work before this is even possible.
01Campfire60%
02Wild West25%
03Blueprint10%
04Engine4%
05Ecosystem1%Level 5 · Ecosystem
On Campfire and Wild West there is nothing to atrophy, because the work has barely changed. On Blueprint the method is written down, and writing it down is itself a defence. It is on Engine and Ecosystem, where the machine carries the work end to end and everybody has stopped noticing that it does, that a capability can disappear without anybody feeling the loss.
Which is also why this one gets waved away. It arrives at exactly the point where everything finally seems to be working.
What to do
One day a quarter, with it switched off.
Airlines did not answer this with training. They answered it by putting hand-flying back into ordinary days. The same move works here and it costs you four days a year.
Pick one piece of real work per quarter. Not a drill, not a workshop. Real work with a real deadline, done without the tool. Somebody who has been here a long time does it alongside somebody who joined recently, and the pairing is the point rather than the output.
Three things fall out of that day and only one of them is the work. You find out who can still do it. The person who joined recently learns the reasoning instead of the shortcut. And you get an honest answer to a question you cannot ask any other way, which is what your company would actually manage on a bad week.
What not to do
Do not ban the tools for a week to make a point. It reads as punishment, people work around it, and you learn nothing except who is good at working around things. One day, chosen, with the reason said out loud.
What it costs to leave it
The bill arrives on your worst day.
An outage is the cheap version. You lose two days and everybody has a story about it afterwards. The expensive version has no outage in it at all.
Something goes out that should not have. It passed a person who looked at it, and that person is neither careless nor unqualified. They no longer had the picture that would have made the mistake visible. Nobody will call that skill atrophy afterwards. It will be called a mistake, and somebody will carry the blame for a capability the company let go of on purpose, one sensible decision at a time.
The second cost is slower and worse. If nobody practises the reasoning, nobody can teach it, and then you are not two years away from losing it. You are one retirement away.
Questions people ask
AI and skill atrophy, the questions people actually search for.
What is AI skill atrophy?
It is a capability that stops being practised because a machine does it every time. Nobody decides to give it up. The work still gets done, often faster, and the ability to do it without help fades in the background until something forces the question. It is the organisational version of a use-it-or-lose-it problem, and the loss is usually noticed by accident.
Does using AI actually make people worse at their jobs?
Not in the way most people expect. The best study of automation and skill, by Casner and colleagues in Human Factors in 2014, put sixteen airline pilots in a Boeing 747-400 simulator and varied how much automation they used. Their instrument scan and their manual control were unimpaired. What had degraded were the cognitive tasks: holding a picture of the situation, knowing what came next.
How do you know whether your company has lost a capability without noticing?
You look at how the work is arranged rather than at the people. emaho places companies on five AI Culture Levels, and skill atrophy only becomes possible at Level 4 and Level 5, where the machine carries work end to end and nobody is required to hold the reasoning any more. Below that the method is still in people's heads. The scan is six questions, three minutes and anonymous.
Which skills fade first when a machine does the work?
The thinking ones. Casner and colleagues concluded that cognitive skills degrade faster than psychomotor skills. Applied to office work that means the ability to operate the tool stays sharp while the ability to judge the output, explain the method and see what is missing goes first. That is the dangerous order, because judgement is exactly what a human is there for.
Why is skill atrophy hard to spot?
Because the person still performs. Somebody who cannot do a task at all is obvious within a week. Somebody who can still do it but can no longer tell whether it is right looks like somebody who is fine. The gap only shows up when the work is wrong and the review that should have caught it did not, and by then it gets called a mistake rather than a missing capability.
What can emaho do about skill atrophy specifically?
Two things, and neither is training. The first is a picture of where the reasoning still lives in your company, per person and per team. The second is designing the practice back into ordinary work: one real piece of work a quarter without the tool, an experienced person next to a recent joiner, and a habit that survives after the workshop is over. Start with your own profile, it is free.
How did aviation deal with this?
By putting practice back into ordinary days rather than into a classroom. In February 2022 the American aviation regulator published Advisory Circular 120-123, recommending that operators give pilots regular opportunities to fly manually with the flight director, autopilot and autothrottle switched off. The response to automation was not less automation. It was deliberate practice without it.
Is there evidence of skill atrophy inside companies using AI?
Not yet, and that is worth being honest about. There is a peer-reviewed theoretical model, published by Ganuthula in Human Behavior and Emerging Technologies in 2026, and there is decades of measurement in aviation. There is no field study measuring the effect inside companies. Anyone quoting a percentage on this is quoting a forecast or an anecdote.
Where does this sit in the wider set of AI challenges?
It is one of twenty-five. Skill atrophy is a late one, which is why it gets missed: it turns up when everything finally looks like it is working. The full set of 25 AI challenges runs from work nobody owns and control that quietly slips through to culture that shifts before anyone names it, each with the research behind it, the level where it starts to bite, and a test you can run this week.
How do you test whether your team has lost the skill?
Ask somebody to explain how a familiar piece of work is made, start to finish, without opening the tool and without opening an example. Ten minutes. If they can, the knowing is intact. If the explanation keeps reaching for something to point at, you have found the edge. It is not a performance test and it should not be run as one.
Does this mean we should use AI less?
No, and treating it that way tends to backfire. Banning the tools for a week reads as punishment, people work around it, and the only thing you learn is who is good at working around things. The aviation answer is more precise: keep the automation, and build in regular, deliberate practice without it so the underlying capability stays alive.
18
Oversight erodes
What happens to a check that keeps coming back fine, and how quickly people stop performing it properly.
Read this one 03No juniors to develop
The work juniors used to learn on is the work that went first, and what that does to your bench.
Read this one 14Own AI context
Everything about how your company works that a model would need and cannot find anywhere.
Read this oneAll 25 AI challenges, and what to do about each
The full set of what goes wrong inside a company once people start using AI: work that nobody owns, control that quietly slips, leadership decisions made on an out of date picture, and a culture that shifts before anyone names it. Every one with the research behind it, where it sits in the five levels, and a test you can run this week. One page, no email.
Your next step
Find out what your company could still do on a bad week.
This is not a training problem and it is not a tooling problem. It is a question about how the work is arranged, which is what the AI culture levels are for.
- Fifteen minutes, on their ownNo manager watching, nothing to pass or fail, and they see their own result first.
- A level per person, a picture per teamYou find out who still holds the reasoning and who has quietly stopped using it.
- One day a quarter, bookedReal work, tool switched off, an old hand next to a new one.
First profile free · no credit card · built by emaho
Paul Musters works with founders and CEOs of scale-ups and innovative SMEs on leadership, teams and AI-native ways of working. He built the AI culture levels and the set of twenty-five challenges this page belongs to.
Sources and numbers used on this page
- Stephen M. Casner, Richard W. Geven, Matthias P. Recker and Jonathan W. Schooler, The Retention of Manual Flying Skills in the Automated Cockpit, Human Factors, volume 56, December 2014, pages 1506 to 1516. Sixteen airline pilots flew routine and non-routine scenarios in a Boeing 747-400 simulator while the researchers varied the level of automation, graded performance and probed what the pilots were thinking. Source of: the finding that instrument scan and manual control were unimpaired while the cognitive tasks of manual flight were significantly affected, and the conclusion that cognitive skills degrade faster than psychomotor ones. Sixteen participants is a small sample, it was a simulator rather than a line flight, and it was aviation. The application to office work is this page's inference and not the researchers'.
- Federal Aviation Administration, Advisory Circular 120-123, February 2022. Recommends that operators give pilots regular opportunities to practise manual handling, including with the flight director, autopilot and autothrottle switched off. Source of: the regulatory response described above.
- Venkat Ram Reddy Ganuthula, The Paradox of Augmentation: A Theoretical Model of AI-Induced Skill Atrophy, Human Behavior and Emerging Technologies, 2026. A peer-reviewed theoretical model rather than a field measurement. Listed as the place where the AI-specific version of this argument is set out. The full text sat behind a paywall at the time of writing, so only the existence, the venue and the subject of the paper are relied on here.
- On what is missing. No field measurement of skill atrophy inside companies using AI was found while researching this page. Everything circulating on that point is anecdote or forecast, and that is said here rather than papered over. The scene involving Nina is a composite drawn from client conversations rather than one company. The bar widths in the figure are illustrative: the direction of the effect is measured, the sizes are not.
- emaho AI culture levels. The share of organisations per level is emaho's own estimate, calibrated against published adoption research rather than measured directly.