AI Skill Erosion: How to Use It Without Getting Worse at Your Job
You have probably had the thought, even if you didn’t say it out loud: am I getting worse at this?
You’re not alone in it. In a survey of 2,500 workers across ten countries, 50% said they rely too much on AI, 30% said they can’t function without it, and 39% believed the reliance was eroding their skills and making them less intelligent. Among Gen Z respondents that last figure was 46%.
Those are feelings, not measurements. But there’s a study that puts something more precise underneath them, and it points at a fix that isn’t “use AI less.”
The variable isn’t how much you use it
Researchers at Microsoft and Carnegie Mellon surveyed 319 knowledge workers about 936 real examples of using AI at work. The finding worth memorizing is this one, in their words:
“Higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking.”
Read that twice, because it reframes the problem. What predicts whether you think hard about a task isn’t how often you reach for AI. It’s how much you trust it relative to how much you trust yourself.
Which means two people can use AI for the identical task all week and only one of them gets worse at it. The one who stopped reading the output carefully is the one losing the skill, not the one who used it more.
What actually erodes, and what doesn’t
The mechanism is Letting something outside your head do the remembering or the reasoning: a calculator, a satnav, a colleague who always knows the answer. It isn’t inherently bad; it’s how you free up attention. It costs you when you offload something you needed to stay good at. , and it’s selective rather than general.
What decays is the thing you stop doing entirely. Not the thing you hand over and then check. That distinction matters more than it sounds, because it means the danger zone isn’t your hardest work. You scrutinise that anyway. It’s the routine, low-stakes end, where the output looks fine, nobody’s watching, and reading it properly feels like a waste of time.
Reporting on the same study notes that among the cognitive activities workers described, evaluation held up best: the effort to judge whether something is any good dropped least, while comprehension dropped most. That’s a useful signal about which muscle is still getting exercise, and which one to protect deliberately.
There’s also a genuine shift rather than a pure loss. The study describes critical thinking moving from gathering information to verifying it, and from solving the problem to integrating the answer into a real situation. That’s not nothing. Those are real skills, and they’re the ones the job now rewards.
Four habits that keep the skill
None of this requires using AI less. It requires using it in a way that keeps you in the loop on the things you actually want to stay good at.
1. Pick what you’re protecting, and be specific
You cannot defend every skill and you shouldn’t try. Name two or three that are genuinely load-bearing for your career: the analysis your judgement rests on, the writing that is recognizably yours, the domain knowledge people come to you for. Everything else is fair game to offload.
2. Draft first, then compare
For those two or three, write your own version before you look at the AI’s. Not the whole thing: five minutes and a rough shape is enough. Then ask for its version and see where they differ. Every difference is either something you missed or something it got wrong, and finding out which is the whole exercise. This is slower. That’s the point: you’re paying for retention.
3. Ask it to critique, not to produce
“Here’s my draft. What’s the weakest argument in it, and what would a skeptical reader push back on?” gets you a better result than “write this for me” and leaves the thinking where it was. It also sidesteps the confidence trap, because you’re using the tool as a second opinion rather than a first draft.
4. Watch your own trust
This is the actual warning light, and it has a name in the research on people working alongside machines: The tendency to accept what an automated system tells you over your own judgement, and to stop looking for the errors you’d catch in a person’s work. It’s been studied for decades in aviation and medicine, long before AI, and the pattern is always the same: the better the system usually is, the less carefully people check it. . Given that confidence in the tool predicts reduced thinking, the moment worth noticing is when you stop reading the output properly, not when you start using AI more, but when you start checking it less. If you can’t remember the last time you disagreed with something it produced, that’s the signal.
The honest counter-argument
Not every skill is worth keeping, and treating all erosion as loss is how you end up doing unnecessary work to preserve something nobody needs.
Nobody mourns their mental arithmetic. Almost nobody can still read a paper map, and the world has not obviously suffered. Plenty of what AI absorbs is genuinely tedious (formatting, first-draft boilerplate, looking things up) and the attention it frees is real.
The argument here isn’t protect everything. It’s that the choice should be deliberate rather than accidental. Right now, for most people, it’s accidental: the skills going are simply the ones that happened to be easiest to hand over, which has nothing to do with which ones were worth keeping.
The takeaway
Skill erosion isn’t caused by using AI a lot. It’s caused by trusting it more than you check it, on tasks you’ve quietly stopped doing yourself.
So the fix is small and specific. Name the two or three skills that actually matter for your work. On those, draft before you compare, and use AI to challenge your thinking rather than replace it. Everywhere else, offload freely. That was always the point.
And keep an eye on the one warning sign that predicts the rest: the day you stop reading the output carefully is the day the skill starts going.
Sources: the reliance figures are from GoTo and Workplace Intelligence’s “Pulse of Work in 2026” (opens in a new tab), 2,500 workers across ten countries (1,250 knowledge workers and 1,250 IT decision-makers), fielded November 2025 to January 2026. GoTo sells remote-work IT software, so it has an interest in how this gets framed. The critical-thinking research is Lee et al., “The Impact of Generative AI on Critical Thinking” (opens in a new tab) (Microsoft Research and Carnegie Mellon, CHI 2025), 319 knowledge workers and 936 first-hand examples. The quoted sentence is verbatim from its abstract, which is what I could verify directly; the per-activity effort breakdown comes from the full text via secondary coverage rather than my own reading of it. Worth noting Microsoft sells the tools this research is about, and published a finding that cuts against them anyway. The four habits are mine, not either study’s.
Related: Giving AI Context: The Skill That Actually Makes It Useful is the skill worth building rather than protecting. How to Check AI’s Work: You’re Looking for the Wrong Mistake is the practical version of “check it more than you trust it”, including what to look at.