AI writes most of my code now. It writes most of the code for everyone I work with, too. By every promise made from a conference stage in the last two years, we should be shipping three times faster, our teams should be a third the size, and a healthy chunk of us should be updating our LinkedIn to "open to work" with that little green ring of doom.
Reader, I regret to inform you that nothing happened.
Not "not yet." Not "the curve is about to bend, trust the roadmap." It measurably did not happen — and the why is so much more interesting than either the hype or the panic. It's also the part nobody on either side wants to say out loud at parties: we automated the cheap part of software, and the cheap part was never the bottleneck.
Let me show you the math. Then I'll tell you a story about 1,000 tireless workers in 200 B.C. that explains why "cheaper labor" and "a richer economy" are absolutely not the same sentence — no matter how confidently someone says otherwise into a microphone.
The apocalypse you ordered vs. the one that shipped
The forecasts were apocalyptic and very specific. Mass displacement. White-collar extinction. A jobless decade with your name on it. So here's the awkward bit for Team Doom: two-plus years in, the aggregate data mostly isn't there.
Anthropic's own labor-market research — and yes, that's the company building the thing — found no systematic rise in unemployment for highly AI-exposed workers since late 2022. Yale's Budget Lab found that even among the unemployed, there's no clear upward trend in how AI-exposed their old jobs were. The Dallas Fed checked wages in exposed occupations and found they aren't uniformly falling — exactly what you'd expect if AI were augmenting people instead of replacing them.
I'll be honest later about the one place the doom genuinely is showing up (it's real, and if you're a junior, it's pointed straight at you). But "the machines took all the jobs" is not what the numbers say. So why not? Buckle up, this is the technical part.
Reason 1: A benchmark is not a job
Every "AI does the work of N engineers" claim is built on a category error wearing a nice suit. It measures AI doing tasks and then announces AI can do jobs. Those are different things, and the gap between them is the entire ballgame.
A model can pass a coding benchmark, one-shot a leetcode problem, and conjure a CRUD endpoint before you've finished your coffee. That's the first 70–80% of a piece of work, and it feels like magic because it slams the demo gap shut instantly. But shipping isn't a demo. A demo has to run once, for a friendly audience, on a good day. Production code has to run a million times, under load, with auth, against edge cases nobody wrote down, inside a system carrying ten years of accumulated context and tacit rules that live in exactly one senior engineer's head (and they're on vacation).
That last 20–30% — correctness, integration, the cursed legacy constraint, being the one who's accountable when it pages at 3 a.m. — is where most of the time and nearly all of the cost live. And it's precisely what AI is worst at, because it has no intent. It doesn't know what you actually meant. It will cheerfully hand you something plausible and wrong, and now a human has to read it, understand it, and decide whether to trust it. A job is not a stack of independent tasks you can hand off one at a time. It's a system of interdependent judgment — and you cannot autocomplete your way through a system.
Reason 2: Cost has almost nothing to do with outcome
This is the argument I think is the most underrated, so let me put it in bold and then yell about it gently.
The marginal cost of producing a line of code is collapsing toward zero. That part of the hype is genuinely true. But software projects were never gated by the cost of producing code. They're gated by things AI doesn't touch:
- Knowing what to build. The expensive failure mode in software is building the wrong thing efficiently. AI just helps you build the wrong thing faster, with great test coverage.
- Coordination. Brooks's Law did not get repealed. Communication overhead grows with the number of agents — human or robot — touching a system. More output can mean more to reconcile, not more done.
- Verification. Someone has to confirm the generated code is correct, safe, and maintainable. The cheaper generation gets, the more there is to babysit.
- Integration and maintenance. Code is a liability you maintain forever, not a trophy you produce once.
Drop the cost of the typing by 90% and the project still ships at roughly the same speed, with roughly the same success rate, because the binding constraint was somewhere else the whole time. We made the cheap part cheaper and then acted shocked the bill didn't move. Input cost and project outcome are barely correlated — and any senior who's watched a "we'll just throw more capacity at it!" project faceplant knows this in their bones. AI is the new "more capacity." The boulder doesn't care how many hands are on it if it's stuck against the wrong wall.
Reason 3: The productivity gains don't show up in the metrics
If individual developers are now so much faster, where is it in the team's numbers? Cycle time, deploy frequency, defect rate, time-to-ship — for most teams, these have barely twitched. This isn't a vibe. We finally have the receipt.
In July 2025, METR ran a randomized controlled trial — the clinical-trial gold standard, not a vendor survey with a hopeful headline. Sixteen experienced open-source developers, 246 real tasks pulled from their own mature repos (averaging a million-plus lines of code), each task randomly assigned to allow or forbid AI tools (mostly Cursor Pro with Claude). The result (full paper here):
Developers using AI took 19% longer. They had predicted it would make them 24% faster. And even after finishing slower, they still believed AI had sped them up by about 20%.
Sit with that one. The perception gap isn't a rounding error — it's a full-blown hallucination, and for once it's the humans doing it. The tool feels fast (you're typing less! it's doing stuff!) while the stopwatch quietly disagrees, because the time just relocates: less time writing, more time reading, prompting, reviewing, and politely correcting confident nonsense.
Two honest caveats, because they make the point stronger, not weaker. First, this was early-2025 tooling, and the slowdown is biggest for experts working on code they already know — where the AI is mostly friction, because they already had the answer in their head. Juniors on unfamiliar code often feel the magic for real. Second, METR's own February 2026 follow-up on newer tools couldn't get a clean signal, so I'm not going to pretend "19% slower" is a law of physics. But here's what's not in dispute: the one rigorous RCT we have found a slowdown, and nobody has produced rigorous evidence of the transformational team-level speedup the keynote promised. Even research out of the labs has found AI assistance can dent conceptual understanding and debugging while delivering no average efficiency gain.
The deepest reason, though, is just arithmetic. Call it Amdahl's Law for shipping software: if writing code is, say, 30% of your end-to-end delivery time, then making the coding infinitely fast improves the whole pipeline by at most 30%. The other 70% — planning, review, QA, deploys, waiting on dependencies, the meeting to schedule the meeting — is untouched. And it gets worse: speed up one stage and you don't remove the bottleneck, you relocate it. Generate code 5× faster and your review queue becomes the new traffic jam. More PRs, more surface area, more bugs to triage. Output went up. Outcome did not. Lines of code was always a vanity metric, like measuring a novel by the kilogram.
Meanwhile, executives are on earnings calls assuring analysts their best engineers haven't hand-written code in months — they just "generate and supervise." So: controlled experiments finding slowdowns on one side, CEOs declaring a revolution on the other. Both cannot be right. Only one of them has a stock price to defend.
The story nobody's modeling: your 1,000 tireless workers
Now the economics — and this is the part I genuinely can't stop thinking about. Forget the code for a second. Picture it.
It's 200 B.C. You own 1,000 workers. They're the dream workforce: they never eat, never sleep, never complain, never quit, never unionize, never "circle back." You pay them, and they do exactly what you say, all day, forever. Your villas, your roads, your pottery — output through the roof. You are, on paper, the most productive man in the city.
There's just one quirk. Your workers don't spend anything in your city. Every coin they earn, they ship across the sea to their real master, who lounges on another continent and could not care less about your local economy.
At first, glorious. You're out-producing everyone. But the baker down the street used to sell bread to the laborers you replaced. Those people have no wages now, so they don't buy bread. The baker can't pay the cobbler. The cobbler can't pay the tailor. The coins that used to circulate through the city — wage to purchase to wage to purchase — have quietly drained across the ocean and stopped coming back.
Then one morning you look up and realize: nobody in the city can afford to buy a single thing you make. Your customers were the workers. You out-produced the entire city and went broke doing it.
That's the whole problem, in a toga.
In a healthy economy, workers aren't just producers — they're the customers. A wage isn't a cost that vanishes into the void; it gets spent, and that spending is somebody else's revenue, which becomes somebody else's wage. Economists call it the circular flow of income, with a multiplier riding shotgun: one dollar of wages, spent and re-spent locally, supports several dollars of activity. That circulation is the economy. Henry Ford clocked this a century ago when he famously doubled his workers' pay — partly so they could actually afford to buy the cars they were building. Demand has to live somewhere.
An AI "worker" is the slave that ships its coins overseas. It's a phenomenal producer and a perfectly useless consumer. The "wage" you'd have paid a person — who'd have spent it at the local baker, the local everything — instead becomes a subscription or API fee that flows to a tiny, geographically concentrated handful of labs and their shareholders, where it's largely retained and plowed straight back into more compute. It does not recirculate as broad consumer demand at anything like the velocity a paycheck does. In the language of the fable: the gold sails across the ocean and forgets the way home.
So you can win the productivity game and still lose the economy, because productivity is not prosperity if the gains concentrate and the demand base quietly erodes. The firm that automates its entire workforce eventually discovers its market has no money — because, in aggregate, the workforce was the market. The snake gets peckish and eats its own tail.
(Is this the whole story? No — it's a deliberately sharp model, and I'll say so. The optimist's rebuttal is real: cheaper goods free up income to spend elsewhere, capital owners do eventually invest, and historically new sectors absorb displaced labor. I'll grant every word of it below. But the fable isolates the one dynamic the cheerful "AI will make everything cheaper!" takes love to skip: cheaper production with no offsetting income is just demand destruction wearing a productivity costume.)
What is actually happening (the part I owe you)
If I stopped here, I'd be doing the exact thing I just dunked on the hype-men for. So: the apocalypse isn't extinction, but something narrower and meaner is genuinely happening, and the data is very clear about who it's aimed at.
It's the entry-level squeeze. Workers aged 22–25 in the most AI-exposed roles have seen a measurable drop in employment, and hiring of young workers in exposed occupations has slowed. Goldman expects unemployment to tick up with AI as "the big story" in the labor market, with entry-level knowledge and content workers most exposed — while carefully hedging that it's "not a foregone conclusion."
This is the honest center of the whole thing. Aggregate employment can look perfectly stable while career mobility quietly collapses underneath it. The pattern isn't mass firing; it's task compression, hiring freezes, and a brutal narrowing of the on-ramp. Companies aren't replacing their seniors with AI. They're quietly not hiring juniors — because the junior's old job, those small well-defined starter tasks you learned the craft on, is exactly the 70% AI now does for pocket change. We are, in effect, sawing off the bottom of the ladder and then squinting up at it, wondering aloud where the next generation of seniors is supposed to come from.
The caveats, before you @ me
- It's early. Actual AI deployment is still a fraction of what's theoretically possible — fully automating even half of today's tasks could take ~20 years — and the research consensus puts the biggest labor effects in the 2027–2030 window. "No effect yet" is not "no effect ever." I'm telling you the apocalypse framing is wrong, not that you should kick back forever.
- History sides with the optimists, but it's not a receipt from the future. Over 85% of employment growth since 1940 came from technology-driven job creation; new tools have reliably created more work than they destroyed; the lump-of-labor fallacy is a fallacy for good reasons. "This time is different" has lost every previous bet — which doesn't mean it can't eventually win one.
- The slave fable is a model, not a proof. The demand-destruction dynamic is a real risk, not a guaranteed ending. Markets adapt, prices fall, income reallocates. The point isn't that collapse is inevitable; it's that "we made labor cheaper" doesn't automatically mean "we made everyone richer," and anyone who tells you it does is quietly skipping the hardest step.
The takeaway
The robots didn't take the jobs, because they only took the tasks — and tasks were never the bottleneck, the cost was never the constraint, and the productivity was never hiding in the typing. The apocalypse keeps not happening because it was the wrong shape: we were promised Terminator and we got a GDP report that won't move and a velocity chart that's a dead ringer for last year's.
The thing actually worth worrying about isn't dramatic enough for a headline. It's quieter, and it compounds: an economy getting better at producing and worse at circulating, and a generation of juniors locked out of the door every current senior strolled through.
The robots aren't coming for your job. They're quietly making sure the next person never gets one — and then they'll push a shopping cart through the empty mall, wondering why nobody's buying.
Top comments (0)