"I've become super reliant on Copilot. It's like magic... but now I blank out when I have to write anything from scratch."
Margot is a freelance developer, nine years in. Her delivery speed has never been higher and her clients have never been happier. Then last month she opened a blank editor for a small utility script — no assistant, just her — and sat there. The syntax was gone. Not rusty. Gone.
Call it the vibe-coder's hollowing: the skill atrophy of full AI delegation, where the abilities that made you good dissolve from disuse while your output climbs — a deskilling you chose, can feel happening, and can't easily justify reversing. Nobody took Margot's work away. What's draining out is her ability to do it unassisted, with her consent, one delegated task at a time.
The promotion and the hollowing arrive together
The strangest part of this bind is how well it pays. A senior web developer in the UK, ten years in, described going full AI workflow in a post that reads like the whole 2026 condition in one voice:
"I've gone full AI workflow in the last 3 months. Thousands of £/$ in tokens. Multiple cursor windows with multiple agents doing shit. I don't think I've coded an entire file or feature myself in that time, just tweaks or slight refactors. And I know what that sounds like - I'm a dirty vibe-coder..."
By every external measure, the workflow is a triumph:
"My output has gone through the roof, I've gotten a fairly large raise/promotion and crazy generous token budget. But what if Claude goes away next week? There's NO WAY I'd be able to output what I am currently..."
And underneath the raise, the promotion, and the praise, the closing line of the post:
"Despite getting praise for my work, I feel like I'm cheating..."
Hold that sequence. The reward structure and the identity are moving in opposite directions — the metrics say better than ever while the self says fraud. That's the hollowing's signature, and it's why the bind stays invisible from outside. Nothing on the dashboard is wrong.
The "what if Claude goes away" fear deserves its own plain paragraph, because it's a business risk wearing an anxiety costume. A freelancer who can only deliver at current speed with the tool has concentrated their whole capacity in a vendor relationship they don't control. Pricing changes, model deprecations, an outage the week of a deadline — the exposure is real. The hedge is the same skill maintenance the workflow keeps arguing you out of; the atrophy question and the continuity question have the same answer.
Is AI actually making you worse at your job?
At the specific skills you delegate — yes, the evidence points that way. A 2025 study from Microsoft Research and Carnegie Mellon University surveyed 319 knowledge workers across 936 real AI-assisted tasks and found that the more confident workers were in the AI, the less critical thinking they reported applying. The work shifts from doing the task to supervising the tool's version of it.
Supervision is real work. Margot catches the assistant's mistakes precisely because nine years of unassisted practice taught her what mistakes look like. The catch is where that leaves next year: reviewing spends judgment faster than it builds it, and the capital she's drawing down was banked in exactly the hours the new workflow deletes.
The UK developer says the quiet part outright: the model's errors get caught because "being at senior level I can typically spot and correct them." The seniority doing the catching was built by years of the unassisted work that now feels unjustifiable. This series asked where the next senior comes from in the pipeline paradox. The hollowing is the same question asked from inside one career: what happens to this senior's edge, five years in?
Keeping the muscle is now an indulgence
The obvious answer — just practice — is exactly what the workflow makes irrational. The developer above tried:
"I was previously giving myself some rules where I'd only use AI to do repetitive tasks or I'd do a certain amount of tasks myself (no AI) just to keep myself frosty. Now I just...can't. I know I'm almost wasting time if I do."
This is what separates the hollowing from every ordinary deskilling before it. The tool sits on your desk, inside your editor, whispering the arithmetic: the unassisted version of this task costs more and ships slower, and the client is paying for neither. Every hour of practice has a visible price now, and the price gets billed against your own deadline.
Meanwhile the market grades the atrophy as growth. Faster shipping, bigger scope, happier clients, a raise. Practice — the from-scratch kind, the kind that built the instincts — reads as self-indulgence, to be scheduled against paying work and defended against your own spreadsheet.
And the mechanism crosses every trade. A marketing freelancer, same hollowing, different craft:
"48% of marketing gig workers admit that AI helps them deliver projects more efficiently. But here's the dark side nobody talks about: I can't write a headline without checking it against Claude first. My instincts are atrophying."
The tell in that quote is the word checking. The skill hasn't vanished — the trust in it has. Atrophy begins as a confidence leak long before it's a capability loss: the instinct still fires, and you no longer ship what it says without a second opinion from the machine.
Why it feels like losing yourself
A skill you've spent a decade building stops being a tool you use and becomes part of who you are. This series named that fusion the commodity self: weld your identity to a capacity, and your sense of who you are inherits that capacity's market price. AI repriced the capacity, and the self got repriced with it.
The hollowing runs the same wound deeper. The developer's grief mourned a craft the market stopped paying for. The vibe-coder's version is stranger: the market pays better than ever, and the craft is dissolving anyway — from the inside this time, through disuse. What goes quiet is more than capability. The daily evidence that you are what you call yourself goes quiet with it.
Margot's blank editor was that moment. The title says developer; the hands, on their own, couldn't prove it. The gap between the two is where the fraud feeling lives — and it behaves differently from the classic version. Old imposter syndrome doubted a competence that existed. This one watches a competence leave, in real time, and the watching is accurate.
What you decide to keep
Nobody's business survives a return to 2022. No client will fund an artisanal hand-coding era, and the tripled output can't be handed back. The honest move the atrophy stories point toward is smaller and harder: triage.
Name the two or three capacities that carry everything else — the judgment that catches the model's confident errors, the architecture sense, the voice, the read on what a client actually needs. Keep those in deliberate, scheduled, unassisted use. Let the rest go the way of long division. The question that sorts them is blunt: which capacities can the machine have, and which ones can't you stay yourself without?
On that short list, practice stops being nostalgia and becomes maintenance of the asset the entire workflow depends on — the judgment the supervision spends. An hour a week of from-scratch work on those few skills is the tithe the raise should be paying.
The identity work is slower, and it's the work every displaced trade in this series has eventually faced: moving the self off the capacity and onto the judgment. One marketer who made the move:
"AI skills pay 45% more. I stopped competing as a copywriter and started consulting as an 'AI marketing strategist.' Same skills, different frame. My income is higher than before AI, but the identity transition was brutal."
Brutal, and survivable. Her instincts still exist; what changed is what she builds her name on.
Margot kept the blank-editor hour. Friday mornings, no assistant, small tools she could have delegated. The code is unremarkable. The proof is the point.
In Haven AI's research across 8,300+ freelancer quotes, the vibe-coder's hollowing is the newest variant of the commodity self: the freelancer whose AI workflow multiplies output while the underlying skills atrophy from disuse, until the identity built on those skills starts asking what's left. The ones who steady themselves stop defending every skill and choose the few that carry the rest — the judgment the machine's output depends on — and keep those in deliberate use while the frame they sell under changes.