Skill atrophy is the loss of a capability you used to have, because something else now does it for you. It is not a new phenomenon. What is new is the speed, and the fact that the thing doing it for you produces output good enough that you cannot tell.
The question people actually want answered is narrower than the headlines: does using AI make me worse at my job, or only feel like it might? The evidence is now good enough to answer that, and the answer has a condition attached.
The finding that matters most
Bastani and colleagues ran a randomised controlled trial with roughly 1,000 students, published in PNAS in 2025. Students got one of three conditions: no AI, an unguarded GPT-4 tutor, or a guardrailed tutor that withheld answers and pushed students to reason.
While they had access, both AI groups outperformed the control. That is the part everyone quotes.
Then the tools were taken away and everyone sat a closed-book exam. The unguarded-AI group scored 17% below the no-AI control. They had learned less than students who never touched the tool, while feeling like they were learning more.
The guardrailed group scored the same as the control. Same model, same students, same material. The only difference was whether the tool let them skip the thinking.
This is why "is AI bad for skills" is the wrong question. The tool is not the variable. The interaction pattern is.
The finding that shows where it bites
Dell'Acqua, Mollick and colleagues ran 758 BCG consultants through realistic tasks with and without GPT-4. On tasks inside the model's competence, AI users were about 25% faster and produced better work.
On tasks outside it, the ones where the model is confidently wrong, AI users were 19 percentage points less accurate than the control.
The people who did best on the hard tasks were not the ones who used AI most or least. They were the ones who could tell which kind of task they were on. That discrimination is a skill. It is also exactly the skill that decays when you stop doing the work yourself, which is the trap: the ability to catch AI's errors is maintained by doing the thing AI is doing for you.
The finding that suggests a mechanism
Kosmyna and colleagues at MIT ran an EEG study on essay writing. Participants wrote with an LLM, with a search engine, or unaided. Neural connectivity scaled down with external support: the LLM group showed the weakest engagement.
The interesting part came in a later session, when the LLM group was asked to write without it. Their engagement stayed low. The authors describe this as accumulated cognitive debt.
Treat this one as suggestive rather than settled. EEG connectivity is not job performance, the samples are small, and it has had less replication than the other two. Nature's 2026 review of the early literature, Is AI ruining our skills?, reaches a similar verdict: the direction is consistent, the magnitudes are not yet nailed down.
What the evidence does not say
It does not say AI use is bad. Both the PNAS and BCG trials found real gains, and the guardrailed condition in Bastani kept the gains without the cost.
It does not say you should use AI less. Nothing in this literature supports a dosage model where fewer prompts equals more skill. The BCG consultants who used AI heavily on frontier tasks did well.
It does not establish that professional skill decays the way student learning does. Every study above is either a student population or a single-session lab task. The extrapolation to a twenty-year career is an inference, not a result. It is a reasonable inference. It is not proven. Confidence: medium.
And it does not say the effect is uniform. The atrophy risk concentrates where you are still learning. Nobody's expert judgement erodes because AI formatted a table for them.
The variable that decides it
Across all three studies, the same thing separates the good outcome from the bad one: whether you did the cognitive work before you saw the answer.
Bastani's guardrails withheld the answer until the student reasoned. Buçinca's earlier CSCW work on cognitive forcing functions found the same: making people commit to a judgement before revealing the AI's recommendation reduced over-reliance substantially, at a measurable cost in speed, which most participants disliked.
That last clause is the whole problem. The intervention that preserves skill is the one that feels worse in the moment. Nobody adopts it voluntarily, on every task, forever. Willpower is not a strategy.
What to actually do
The useful move is not to use AI less. It is to be deliberate about which category a task falls into, before you start.
Ask first, then look. Before you prompt, write your own answer in one sentence. It does not have to be right. It has to exist. The gap between your sentence and the model's output is where learning happens, and if you never wrote the sentence there is no gap to notice.
Sort tasks, not tools. Some work is mechanical and you should hand it over without guilt. Some is in your expert domain and AI should accelerate you. Some is the thing you are trying to get good at, and there AI should slow you down on purpose. The categories are per-task, not per-person, and they change as you improve.
Notice when the AI stops disagreeing with you. If nothing it produces surprises you, either you have mastered the domain or you have stopped reading critically. Those feel identical from the inside.
Protect the skills you would be sad to lose. Not all of them. Pick the few that make you who you are professionally, and do those by hand often enough to keep the calluses.
This is the design behind TAOS. It rates expertise per domain, then varies its behaviour by task: accelerating where you are strong, forcing you to think first where you are growing, and refusing to quietly do the work you said you wanted to keep. We ran it with ten knowledge workers and published what happened, including the parts that did not work.
The honest summary of the literature is this. AI does not de-skill you. Skipping the thinking de-skills you, and unguarded AI makes skipping the thinking the path of least resistance. Change the path.