It’s 11 p.m. and the proposal is due at nine tomorrow. You’re too tired to write it well, so you ask AI for a draft, and what comes back is better than anything you could produce right now. You tidy two sentences, paste it into the proposal, and move on. Behind the relief, a question forms: when did I last write one of these myself, start to finish? Could I still?
That question deserves a better answer than the two it usually gets. The dismissive answer says nobody mourns the slide rule, tools are tools, adapt. The catastrophizing answer says your skills are rotting and you should go back to doing everything by hand. The real answer is more specific than either, and considerably more useful, because it tells you what to protect and what to let go.
Is relying on AI making me worse at my job?
It can, but not in the way the fear suggests. Skills decay where you stop practising them and stop checking the results. They hold or improve where you keep exercising judgment. Whether AI makes you worse depends less on how much you use it than on which parts of the work you still do yourself, and whether you can still tell good output from plausible output.
Keep those two words in mind: which parts. The research below, three workplace studies covering support agents, consultants, and professional writers, will help you find those parts in your own job.
What the best workplace study so far found
Erik Brynjolfsson, Danielle Li, and Lindsey Raymond followed 5,179 customer-support agents after an AI assistant was rolled out, in a study published in the Quarterly Journal of Economics in 2025. Productivity rose about 14 percent on average. But the average hid the story: the newest, least-skilled agents improved around 34 percent, while the most experienced, most skilled agents gained almost nothing. The mechanism explains why. The AI had absorbed the tacit habits of the best performers, the phrasings and moves they had developed over years, and made them available to everyone else.
The AI wasn’t adding intelligence from nowhere. It was lending novices the accumulated judgment of experts. If you’re new, that’s genuinely great news: the floor just rose to meet you. A separate randomized study by Shakked Noy and Whitney Zhang, published in Science, found the same pattern on professional writing tasks: large time savings and quality gains, with the least-strong writers benefiting most.
But notice what it means if you’re the expert, or intend to become one. In the support-agent study, AI gave the experts roughly nothing, because everything it offers is a compressed version of what they already had. The value you personally add is precisely the part AI doesn’t hand you. Everyone gets the borrowed layer now. The question is what you have underneath it.
The borrowed-competence problem
If your output is good because AI supplies the good part, your work improved while you stayed the same. For a first-year analyst, that’s a gift. For the same analyst three years in, it can mean the underlying skill never got built at all, because the practice that was supposed to build it was done by AI instead.
And the underlying skill still matters, because AI is unevenly good in ways that are hard to see from the outside. In a field experiment with 758 BCG consultants, published in Organization Science, Fabrizio Dell’Acqua and colleagues found large gains on tasks that suited the AI. But on a task chosen to sit just beyond its competence, consultants using AI were 19 percentage points less likely to reach the right answer than those working alone. The output on those tasks didn’t look worse. It was exactly as polished and confident as the good output, and that’s the trap: when AI is wrong, it’s wrong smoothly, and the warning has to come from you.
So the two findings assemble into one picture. AI raises the floor by lending everyone borrowed competence, and it occasionally leads its users astray with total confidence. Both point at the same conclusion: the skill that matters most in an AI-heavy job is the ability to evaluate the work. And evaluation is learned by doing the work. You can’t develop the judgment to assess a draft, an analysis, or a plan if you never produce one yourself.
Which skills decay, and which are safe to lend out
One distinction cuts through most of the worry: execution versus judgment.
Execution is producing the thing: the first draft, the boilerplate, the format conversion, the routine analysis. Offloading execution is what tools are for, and the losses are usually livable. Many people describe barely managing mental arithmetic since calculators, or losing their sense of direction to GPS, and mostly it costs them nothing, because a machine is checking the sum and the route.
Judgment is knowing what the thing should accomplish, whether this version accomplishes it, which of its claims will stand up once someone acts on them, and what to do when the standard answer doesn’t fit. No machine is checking that for you. It’s the layer your title, your pay, and your reputation actually rest on, and it decays the same way execution does: through disuse. The dangerous pattern isn’t using AI a lot. It’s a slide that happens in two steps: first you’re approving work you could no longer produce yourself, then you’re approving work you can no longer even judge. The move from the first step to the second is the one that hurts you, and it’s easy to miss, because the documents keep looking fine either way.
So make the inventory explicit. Which two or three capabilities is your role actually built on? Not the whole job description. The ones where, if they eroded, you’d be replaceable by anyone with the same subscription. Those are the ones to protect on purpose.
How do I keep the benefit without the decay?
Four working rules, none of which require using AI less.
Think first on anything that matters. Before asking AI for a recommendation, a diagnosis, or a plan, write your own in a few sentences. It takes five minutes, and it keeps your judgment in the leading position, with AI as a challenger rather than a substitute. When the two disagree, you’ve found either a blind spot or a machine error, and working out which one is exactly the exercise that keeps you sharp.
Keep practising your core craft end to end. At some regular interval, take one task at the centre of your job and do the whole thing yourself, without AI, at full standard. Musicians still practise scales long after they stop needing them consciously, and this is the same kind of maintenance: it keeps the skill available for the day you need it unassisted.
Verify where the risk lives. When you review AI-assisted work, yours or anyone’s, check the numbers, the factual claims, and whatever a reader might act on. That’s standard advice for catching errors, but it’s also underrated as practice: every check you run is an exercise for the judgment you’re trying to keep.
Close the loop by explaining. When AI gives you something you plan to use, restate its reasoning in your own words, to a colleague or a notebook. Where the restatement stumbles is where your understanding was borrowed rather than owned, and now you know what to study.
Underneath all four sits the same principle: AI amplifies whatever clarity and rigour you bring to it. Stop bringing them, and there’s less to amplify every month. Keep building them, and the tools get better for you at exactly the rate they get better.
A four-week experiment
Vague fears don’t resolve. Measured ones do. So measure this one.
Once a week for the next four weeks, pick one real task from the core of your job. Before touching AI, do it cold for fifteen minutes: your draft, your estimate, your diagnosis, whatever the task calls for. Then run it your usual way, with AI, and put the two side by side.
Three things can happen. Your cold version holds up, and the fear quiets: AI has been saving you execution, not replacing your judgment. Or the gap is real but you can see exactly why the AI version is better, which means your evaluation is intact and you’ve located a skill worth some deliberate practice. Or – the useful worst case – you can’t confidently say which version is better. Even that is a finding rather than a catastrophe. It tells you precisely where to point your practice while the gap is still small.
Four data points beat a year of quiet worry. And if you’d rather look at what you find with a second set of eyes, a free intro call is an easy way to do that.