AI Isn't Taking Your Job. It's Hollowing Out the Work That Made You Good

The thesis, up front

AI is not, primarily, deleting jobs. AI is hollowing out necessary work -- the unglamorous, low-glory, foundational labor that organizations quietly run on and that careers quietly get built on. And because "necessary" is the one word nobody puts on a keynote slide, the damage is showing up everywhere except in the number we are all told to watch.

When I say "necessary," I mean the work that has never been anyone's dream job. The documentation nobody reads until 2 a.m. during an outage. The QA pass on a colleague's pull request. The first-draft slide deck. The tier-one support ticket. The 40-line SQL query that answers a question nobody will remember asking. The reconciliation, the transcription, the summary, the "hey, can you look at this before I hit send?" review. This is the connective tissue of every organization on earth.

It is also -- and this is the part the efficiency crowd keeps missing -- the classroom. Necessary work is where competence is manufactured. Junior lawyers learn judgment in document review. Junior developers learn architecture by fixing other people's bugs. Junior analysts learn what a credible market sizing looks like by building thirty bad ones and watching a senior person tear them apart. Strip out the necessary work and you do not simply remove tasks. You remove the only reliable place where tacit knowledge transfers from one generation of practitioners to the next.

That is the hollowing. Not mass unemployment. Empty space where the ladder used to be.

The Context: everything looks fine, which is the problem

Start with the numbers that get quoted. US unemployment has been sitting in the low 4% range, which by any historical standard is a healthy labor market. GDP keeps grinding upward. Corporate margins in technology are enviable. If you only read the headline figure, you would conclude that the AI transition is going fine and that everyone worried about it is a Luddite with a newsletter.

Now look at the plumbing. The hires rate in the US -- the share of employed people who started a new job in a given month -- has been running near multi-decade lows outside of the pandemic shock, per the Bureau of Labor Statistics' Job Openings and Labor Turnover Survey. We are in a low-hire, low-fire market: companies are not cutting aggressively, but they are also not bringing people in. That configuration -- frozen headcount, no blood in the streets -- is precisely the one that hides structural damage, because nothing dramatic enough happens to make a headline.

Meanwhile the IMF's January 2024 analysis estimated that roughly 40% of global employment is exposed to AI, rising to about 60% in advanced economies. The World Economic Forum's Future of Jobs Report 2025 projects 170 million new roles created and 92 million displaced by 2030, and -- more tellingly -- expects 39% of core skills to change. And the New York Fed's labor market tracker for recent college graduates has consistently shown roughly four in ten new graduates working in jobs that do not require their degree.

None of those facts, on its own, proves my thesis. Together, they describe a market where the top-line number stays calm while the entry points quietly seal shut.

Argument 1: The canaries are at the bottom of the mine, and they are already gasping

The single most important piece of research on this topic in the last two years came out of Stanford's Digital Economy Lab. Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen looked at what actually happened to employment in occupations heavily exposed to generative AI, and separated it by career stage. Their finding, published as "Canaries in the Coal Mine?", was that early-career workers in the most AI-exposed occupations saw roughly a 13% relative decline in employment -- even after controlling for firm-level shocks. Experienced workers in the same occupations did not show the same drop. The effect was concentrated in software development and customer service.

Read that again, because it is not the story anyone was telling. The story everyone was telling was that AI would augment workers and eliminate tasks. What the data shows is that AI is eliminating the rungs. The people who already climbed the ladder are fine. The people who needed the first three rungs to get anywhere are standing in an empty stairwell.

Why the bottom? Because that is where work is most codifiable. A junior task is, almost by definition, a task someone already figured out how to describe. If you can describe it well enough to hand it to a 23-year-old with a two-week ramp, you can describe it well enough to hand to a model. Senior work resists automation for a boring reason: it is mostly judgment under ambiguity, and ambiguity is the one input that does not fit in a context window.

Argument 2: Necessary work is invisible by design, so nobody is defending it

Here is why this hollowing proceeds without a fight. Necessary work has no owner. It has no sponsor. It never appears on a quarterly business review because its entire value proposition is that nothing bad happened.

Documentation is the purest example. Ask any engineering leader whether documentation matters and they will say yes with great enthusiasm. Ask them to name the person whose performance review depends on it and watch the room get quiet. Necessary work is structurally under-defended in every organization I have ever seen, which means it is structurally the easiest thing to cut when a tool arrives that appears to do it "well enough."

The trouble is that "well enough" compounds. A model-generated summary of a meeting is 90% accurate. Fine. Then it becomes the only record. Then decisions build on the summary rather than the meeting. Then a decision built on a summary built on an earlier summary turns out to be subtly, expensively wrong, and nobody can trace the error because the artifact that would have contained the nuance -- a human's messy, over-detailed meeting notes -- was never written.

That last figure is from David Autor's work on rebuilding middle-class jobs with AI. Autor is, notably, one of the more optimistic economists on this question -- and I will get to his argument in the counter-argument section, because it is the strongest objection to everything I have written so far and I do not want you to think I have not heard it.

Argument 3: The productivity numbers are a mirage, and the mirage is the point

The case for AI-driven efficiency rests on a handful of studies. Let me give you all of them, then give you the one that should worry you.

The optimistic evidence is real. A field experiment with Boston Consulting Group consultants, published as "Navigating the Jagged Technological Frontier" by Fabrizio Dell'Acqua and colleagues, found that consultants using AI completed 12.2% more tasks, worked 25.1% faster, and produced 40% higher quality output on tasks inside the model's capability. The GitHub Copilot randomized study by Peng and colleagues at Microsoft found a 55.8% speedup on a specific, bounded HTTP server implementation task.

Now the one that matters. METR ran a randomized controlled trial with experienced open-source developers working on their own mature repositories -- exactly the kind of work AI vendors claim to accelerate. The result: developers were about 19% slower with AI tools. And here is the part that should genuinely unsettle you: those same developers estimated they had been 20% faster. They could not feel the slowdown. They felt faster while being slower.

That gap -- between perceived and actual value -- is the mechanism of the hollowing in one finding. AI tools are outstanding at producing plausible artifacts quickly. They are unreliable at producing correct artifacts in complex systems. And humans, notoriously, are bad at distinguishing plausible from correct when they did not build the thing.\p>

So we get a workplace where output volume goes up, review burden goes up, and the number of people who can independently verify the output goes down. That is not productivity. That is a debt being rolled forward, and it is being rolled forward by exactly the people who once would have been building the judgment to catch it.

One more data point, because it is the quietest and most damning. A study with 319 knowledge workers and 936 real examples, published in CHI 2025, found that higher confidence in AI correlated with less critical thinking effort -- while the workers themselves reported that skipping that effort felt fine. Every shortcut you do not notice you are taking is a skill you are not building.

Argument 4: Our measurement instruments cannot see the thing we are arguing about

Unemployment rate counts people actively looking for work. It cannot count a 24-year-old who is technically employed as a "research associate" but who has not been asked to produce an original analysis in eleven months. It cannot count the 31-year-old designer whose role quietly became "prompt the model, then fix the model's output" -- a job that keeps them busy and pays them adequately while providing a fundamentally different set of skills than the job they were hired to grow into.

The WEF's 39% skills-change figure is the closest thing we have to a measurement, and it is a projection, not a count. The Anthropic Economic Index gives us a fascinating view of which tasks people actually use AI for -- with coding and writing dominating -- but task-level usage is not the same as career-level hollowing. And the layoff statistics keep getting muddied by what a lot of us now call AI-washing: attributing a headcount reduction to "AI efficiency" when the real driver was a failed product line or a rate-driven cost cut. That attribution error makes the problem both harder to see and easier to exploit.

If you want to know how exposed your own situation is -- not your job title, your situation -- this is the thing I would actually run. Use the free Career Pulse Score at Workings.me. It asks how future-proof your career is in terms of the work you can do unsupervised, not the official classification of your role. That distinction is the whole ballgame.

The Counter-Argument: this is what ATMs did to bank tellers, and it worked out

Let me state the strongest version of the opposing case without sandbagging it, because it is genuinely strong and it is the reason I do not think we are heading for mass unemployment.

Every major technology shock in modern history has been read as a jobs apocalypse and turned out to be a jobs reshuffle. Autor's statistic is the killer: about 60% of the jobs Americans held in 2018 were in occupational categories that did not exist in 1940. We did not predict those jobs. We did not need to. The economy generated them. ATMs were supposed to end bank tellers; teller employment grew for decades afterward, because the ATM lowered the cost of opening branches. Spreadsheets were supposed to end accountants; the accounting profession expanded, because the software made it cheap to ask more questions. In his 2024 work, Autor argues explicitly that AI could be the tool that finally rebuilds the middle class -- by giving ordinary workers access to the kind of technical judgement that used to be reserved for credentialed elites. He calls this the possibility of de-democratizing expertise, and he is not a naive technologist. He is one of the more careful empirical economists alive.

I hold my position anyway. Here is why.

First, the previous transformations destroyed tasks, not training grounds. The mechanical loom automated weaving; it did not automate the process by which a person learns to be a weaver in the first place. GenAI is different because the tasks it removes are disproportionately the ones that were doing double duty as instruction. When the ATM arrived, a teller still had to learn to handle an angry customer, balance a drawer, spot a fraudulent check. When the first draft arrives pre-written, an analyst no longer has to learn to structure an argument, only to check one.

Second, the new jobs require the old skills as prerequisites. This is the flaw in the "60% of jobs are new" argument. It is true that new categories appear. It is also true that the humans who fill them usually arrive with foundational competence that was built somewhere earlier. You cannot supervise an AI's system design if you have never designed a system. You cannot audit a model's financial explanation if you have never built a financial model. The 60% figure tells us that creation happens. It is silent on whether the on-ramps survive.

Third, the timeline of an apprenticeship is not elastic. Economists think in decades. A human being thinks in months. The abstract promise that a 24-year-old unemployed market researcher will be retrained into an "AI-human collaboration architect" by 2033 is cold comfort when the rent is due in October and the only jobs available are the ones they already know how to do, which no longer exist. Historically, the workers who eat the transition cost are specific people, and the workers who capture the eventual benefit are usually different, younger, or better credentialed people.

So I will concede the long arc and hold the line on the short one. The hollowing is real. It is concentrated at the bottom. And "it worked out eventually" has never once been a plan.

What this looks like inside one career

"I was hired in early 2022 as an associate product analyst at a big insurer. My entire job, honestly, was to build the decks and run the first-pass numbers that my director would eventually present. Data pulls, variance analysis, the slides. Four years of that was supposed to turn me into someone who could run a business case.

By late 2024 my director was generating a first draft in about eleven minutes with an AI tool and asking me to 'check it.' My role quietly became verification. I was not building models anymore, I was auditing someone else's -- and not even a person's. I could not have told you what a defensible assumption looks like versus a plausible one, because I had never built one from scratch and watched it fail.

When the reorganization came in 2025, I had four years of tenure and maybe eight months of actual transferable skill. Nobody warned me about that part. They warned me AI might take my job. They did not warn me that AI would hollow it out first and leave me holding the shell."

Marcus is not a statistic. He is what the statistic looks like from the inside. He was never laid off for being bad at his job. He was, in a precise sense, laid off for never being allowed to be bad at anything -- because being bad at something, safely, with someone senior catching it, is how every competent professional in history got competent.

That is the argument. Not that AI is bad. Not that the technology should be stopped. That we are removing the necessary work, keeping the output targets, and pretending the two are separable.

They are not. And the people paying for that error are the ones with the least leverage to object.

What I'd Tell My Best Friend

If you called me tonight and said, "I think my job is being hollowed out and I cannot tell if I am imagining it," here is what I would say, in order.

First, stop asking whether your job is safe. Ask whether your skills are compounding. Those are different questions and the second one is the only one you control. A safe job that teaches you nothing is a slow-motion layoff. Write down the three things you could do three years ago that you cannot do confidently today -- or the reverse. If the list is empty, you have your answer.

Second, keep a judgment log. This is the single highest-leverage habit I know for surviving a hollowing, and almost nobody does it. Once a week, write down one decision you made and why -- the assumption you tested, the alternative you rejected, the signal that would have changed your mind. In eighteen months you will have a corpus of your own reasoning. That corpus is the thing an AI cannot generate and a hiring manager cannot find anywhere else. It is also, quietly, the thing that makes AI output usable, because you will have trained yourself to notice when a plausible answer is actually a bad one.

Third, do the unnecessary version of the task at least once. When you are asked to check someone's summary, write your own first. When you are handed a generated analysis, rebuild one section from raw data by hand. When you get an AI-drafted email, rewrite the opening paragraph yourself. You are not being inefficient. You are purchasing the skill that lets you supervise the machine, and that skill has a very short shelf life if you never exercise it.

Fourth, negotiate for apprenticeship, not just compensation. This is the most counterintuitive piece of advice in this article and I will defend it. If you are early in your career, a role that pays 8% less but puts you in a room where a senior person reviews your actual work is worth more than the higher offer where you will be handed outputs to polish. You can recover 8%. You cannot recover three years of not learning.

• Ask in the interview: "What work will I do that a tool cannot do, and who will tell me when I have done it badly?"
• Ask: "Who reviewed your last three hires' work, and how often?"
• Ask: "What did the person in this role learn last year that they could not do the year before?"
• If the answers are vague, that is the answer.

Three scenarios, three different moves

Scenario A: You graduated in the last three years and the entry-level door is stuck

Your problem is not that you lack talent. It is that the market no longer pays for the training version of the job, only the finished version. So manufacture the training yourself. Pick one narrow domain and go one layer deeper than the tools go: not "I can use AI to analyze data," but "I know why this cohort analysis is misleading and here is the segment that proves it." Ship that publicly. A portfolio of three pieces of work where you caught something an AI would have missed will outperform thirty pieces where you demonstrated you can prompt well -- because everyone can prompt well now, and almost nobody can catch.

Also, take the unglamorous adjacent job. Support, ops, implementation, field work. Those roles still contain the messy, un-codifiable human contact that builds judgment fastest. The title is not the point. The reps are.

Scenario B: You are eight to twelve years in and your role has quietly become verification

This is the most dangerous position in the labor market right now, and it looks comfortable from the inside. You have seniority, a decent salary, and less and less original work. The risk: your market value is increasingly a function of institutional knowledge, which does not transfer, rather than capability, which does.

The move is to re-own an output. Not a task -- an output. Find the thing your organization produces that nobody currently owns end to end, and own it. Become the person who signs off on the number, the document, the release. Signing off is a real skill and it is exactly the skill the hollowed-out junior version of you would have needed to develop. If you cannot get it internally, build it externally: a consulting engagement, a board seat, a volunteer project where you are the accountable human.

Scenario C: You are genuinely senior and you are fine

Then your obligation changes. The hollowing is not a you-problem anymore, it is a stewardship problem. You are the person who decides whether the junior role on your team still contains any real work, or whether you have converted it into a review queue. Both look the same on a headcount plan. They are opposites on a ten-year horizon.

Two concrete things you can do this quarter. Give one junior person a project with a real decision in it, and let them make the call even if you would have made it differently -- then debrief the reasoning, not just the outcome. And protect one piece of necessary work on your team from automation, explicitly and on the record, because it is load-bearing for someone's development. Documentation. Manual QA on the critical path. The slow version of the customer call. Name it in the planning doc so it cannot be quietly cut.

What I want you to think differently about

The public debate about AI and work is framed as a binary: will it take jobs, yes or no. That framing is a trap, and it is a trap that benefits the people who would rather you not look closely. The real question is not whether you will be employed. It is whether you will be developed. Employment and development have been bundled together for about a century -- you show up, you do the work, the work makes you better. AI is quietly unbundling them. You can absolutely keep the paycheck and lose the growth.

That is what I mean by hollowing. It is not a layoff. It is a job that still exists, still pays, still fills your calendar, and no longer teaches you anything. It is the worst kind of safe.

So here is the reframe. Stop auditing your job title against a list of AI-exposed occupations. Start auditing your week against one question: what did I do this week that I could not have done three years ago? If you can answer that consistently, you are not hollow. If you cannot, no amount of job security will save you, because the security is rented and the rent is going up.

If you want a structured starting point rather than a nagging feeling, run the Career Pulse Score at Workings.me. It is free, it takes a few minutes, and it will tell you how future-proof your career looks based on what you can actually do unsupervised -- which is the only definition that has ever survived a technology transition.

I am not pessimistic about AI. I am pessimistic about our willingness to notice what we are dismantling while the dashboard looks great. The rungs are coming out of the ladder. Somebody should say so out loud, and then somebody should start putting new ones back in -- starting with the one directly under their own feet.

Three things to do before Monday

• Audit your last two weeks. Highlight every task that was primarily producing a first draft, checking someone else's output, or summarizing something you did not attend. That highlighted column is your hollowing risk.
• Book one hour to do a task the slow way. Pick the highest-stakes recurring output in your role and rebuild one instance from raw inputs. Note what you learn that the tool never showed you.
• Have one uncomfortable conversation. With your manager, your mentor, or yourself: "What is the plan for me to build capability this year, not just deliver output?" If nobody has an answer, you now have information you did not have yesterday.

The necessary work was never glamorous. It was never anyone's ambition. It was just the thing that turned a person who had a job into a person who had a craft. We are automating it away faster than we are replacing it, and we are calling the result efficiency because the alternative is admitting we did not think it through.

Bring back the rungs. Start with yours.