AI Won't Take Your Job. Atrophy Will.

Everyone you work with is now one prompt away from a competent answer. A strategy memo, a migration plan, a working prototype — the floor has risen for all of it, for everyone, at the same time.

That levelling is genuinely great. But it quietly changes the question careers are built on. When everyone can produce the same output, what makes you distinguishable is no longer what you can produce — it's what you can think.

And here's the uncomfortable part: the default way most people use AI is eroding exactly that.

The real risk isn't replacement

The public debate is stuck on whether AI takes jobs. The nearer, quieter risk is what I'd call intellectual atrophy: the slow erosion of your ability to think deeply, critically, and independently — happening precisely while your output looks better than ever. Nothing breaks. Nothing alerts. The deliverables keep shipping. You just gradually stop being the one who could have produced them.

Thinking behaves like a muscle: it holds its strength only while it's loaded. Navigation is the cautionary tale — after enough years of turn-by-turn directions, many of us can no longer hold a map of our own city in our heads. Nobody decided to stop being able to navigate. They just stopped navigating.

The same trade is now on offer for every cognitive task: drafting, debugging, structuring an argument, weighing a decision. Each convenience is a small exchange — speed now for a rep of practice forgone. One exchange is harmless. A workflow made of them compounds, and what feels like simplifying your day is quietly thinning out the skill you'll be selling next year.

Polish stopped being a signal

It used to be safe to infer depth from polish. A well-structured document with sharp phrasing and tidy references implied days of contact with the material — you couldn't fake the surface without doing the work underneath. AI broke that link. Polish is free now, which means polish tells you nothing.

Where the difference still shows is under interrogation. Put two equally beautiful pieces of work side by side and start asking: why this option? What did you rule out? What breaks first under load? One person answers from a model in their head — they visited every branch the document summarises. The other person has nothing beneath the surface, because the thinking never happened anywhere.

Why copy-paste makes you worse

Learning research has a blunt finding usually called the generation effect: you retain what you generate — what you attempt, struggle with, and produce — far better than what you merely read and accept. The friction is not an inefficiency in learning. The friction is the learning.

Accepting an AI answer you never attempted skips exactly that step. You get the output without the encoding. Do it once and you've saved ten minutes. Make it your default and you've built a pipeline that produces deliverables while systematically starving the skill that's supposed to be behind them.

“It's easier to be an editor than an author” is true — but editing is only legitimate if you could have been the author. Editors improve drafts because they carry a model of what good looks like. So apply the test honestly: could I have produced this? If the answer keeps being no, you're not editing. You're forwarding.

I've written before that understanding is the bottleneck for teams, and that architecture can't be outsourced to an LLM. This is the personal version of the same law: the erosion doesn't arrive as one big decision, it arrives task by task.

Outsource the task, never the thinking

None of this argues for using AI less. I use it constantly, and you should too — for boilerplate, transformations, first drafts, searches, scaffolding. The line to hold is narrower and more important: AI executes, you comprehend and decide. Four practices keep that line intact.

1. Predict before you prompt

Before you ask, spend two minutes writing your own answer — a sketch, three bullets, a guess at the design. Then prompt, and diff the two. Where the AI beat you is precisely what you need to study; where you beat it is your edge, now made visible. The output is identical either way. The difference is that you turned the task into a training rep instead of a bypass.

2. Interrogate what you didn't write

For anything you accept: why does this work? What were the alternatives? What breaks first? If you can't answer, you're not done — and “the AI did it” is never an explanation. If you can't explain a piece of work, it isn't yours yet; it's just near you.

3. Keep manual reps

Athletes still lift, even though forklifts exist. Deliberately do a fraction of the work unaided: debug something before pasting the stack trace, write the occasional first draft yourself, estimate the number before you compute it. This isn't nostalgia for the hard way — it's maintenance on the instrument you think with.

4. Decide with AI, never through it

Let it generate options, surface evidence, argue against you — never make the final call where it matters. Its knowledge is an average of the internet: biased, conflicting, confidently wrong in places. Judgment is your job, and judgment only exists if you know your field well enough to sense when something is off. That feeling has a name — trained intuition — and it's exactly what atrophies first when you stop exercising it.

Crutch or ladder

Same tool, two postures. Used as a crutch — prompt, paste, move on — AI holds you up while your own footing weakens, and your output is, by construction, the same as everyone else's who typed a similar prompt. Routine, repeatable, replaceable. Used as a ladder, every output is a floor to stand on: a first draft you push past, a baseline you interrogate, a rung on the way to work the tool couldn't have produced alone.

The models will keep improving on the same schedule for everybody — you can't differentiate on access to them. Taste, judgment, and depth still compound privately, but only if you keep training them.

AI levels what we can all produce. What you can think is still yours to build — or to lose.