Start with the version of this story that gets a headline. Earlier this year, Meta drew up a plan called Project OT — Organization Transformation — to make itself "AI native": slash up to 60% of the members of some teams, replace them with AI, and let a much smaller group of senior people supervise what was left. Reuters got the internal numbers. Pragmatic Engineer dated the planning to January, framed as prep for what would have been Meta's largest layoff ever. Then Meta backed off at the last minute, because its own internal data said the plan wasn't working: a 220% year-over-year jump in code changes to Meta's internal platforms and infrastructure, against only a 36% jump in changes that actually reached users. That's the tell. Activity went up. Output for actual humans barely moved. Someone at Meta looked at that ratio and pulled the plan.
This is the AI-replaces-workers story failing in exactly the way you'd expect a loud, all-at-once bet to fail: measurably, publicly, and fast enough that someone with authority to reverse it did. It's satisfying to read about, because it confirms a suspicion a lot of people in this industry already hold — that wholesale headcount replacement by AI agents is still mostly a slide-deck fantasy, and reality has a way of showing up in ugly internal metrics before it shows up in a press release. Fine. Believe that. It's supported.
Now here's the version of this story that doesn't get a Reuters investigation, because there's no single company to investigate and no single decision to walk back.
A Stanford Digital Economy Lab study — "Canaries in the Coal Mine? Six Facts About the Recent Employment Effects of Artificial Intelligence," by Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen — finds that employment among 22-to-25-year-olds in highly AI-exposed occupations is running about 19% below where it would be if it had kept pace with employment among similarly aged workers in less-exposed occupations. Nineteen percent isn't a rounding error. It's also not a story about mass firings. The researchers were explicit about the mechanism: the adjustment operates primarily through reduced hiring of young workers, not increased separations. Nobody's getting a severance package. Nobody's writing a Reuters-worthy internal memo about it. It happens in the ordinary, undramatic act of a hiring manager choosing not to open a junior req -- a decision nobody outside that team ever hears about, made for reasons that never rise to the level of a plan.
Put those two stories next to each other and the interesting question isn't "which one is true." Both are. The interesting question is which one is durable, and the honest answer is that the failed one is the one that was always going to get fixed, and the quiet one is the one nobody is incentivized to fix at all.
Project OT failed for a reason that's specific to what it was: a single, visible, centrally-planned bet, made by identifiable executives, measured against a baseline anyone could check. When the internal telemetry showed a company shoveling 220% more churn into its own codebase for 36% more of anything a user would notice, that gap was legible to the people who'd staked their credibility on the plan. Legible failure gets reversed. That's not a special virtue of Meta's leadership — it's just what happens when a bet is big enough, singular enough, and measured well enough that its failure shows up on somebody's dashboard before the damage compounds.
The entry-level hiring effect has none of those properties. It isn't one company's bet; it's the aggregate of thousands of individually reasonable decisions made by hiring managers who never coordinate with each other and never have to justify the decision to a Reuters reporter. Nobody at any single firm experiences "we quietly didn't hire three junior analysts this year" as a plan that can fail, because it was never framed as a plan — it was a series of small, locally rational choices that a tool made cheaper than a hire. There's no internal metric anywhere that reads "junior pipeline attrition: 19%, trending up" and lands on an executive's desk the way Meta's feature-shipping ratio did. The Stanford researchers had to build that number themselves, from ADP payroll data spanning many employers, because no single employer was tracking it, or had a reason to. Worth saying plainly: the researchers themselves note that the gap in their ADP sample runs larger than what national survey benchmarks show, so 19% may be on the high end of the true economy-wide effect. The direction of the effect isn't in question. The exact size might be.
That asymmetry — one failure mode that's visible and self-correcting, one that's invisible and self-reinforcing — is the actual story, and it's the reason the Meta reversal is not the reassuring data point it looks like at first read. If anything, it's a distraction from the version of this that's actually going to still be true well after this news cycle forgets Project OT.
Here's the mechanism worth sitting with. Entry-level jobs in knowledge work were never just jobs. They were the way the work itself trained people to eventually do the senior version of the job. The tedious, low-stakes, well-specified tasks that used to get handed to a first- or second-year analyst, engineer, or associate weren't just cheap labor — they were the reps. You learned the domain by doing the boring, well-specified parts of it under supervision, badly at first, until you were good enough to be trusted with the harder judgment calls. That pipeline only works if the boring, well-specified work still gets assigned to a human who's early in their career. Once that work is cheaper to route to a model, the assignment doesn't happen, and the reps don't happen either.
Silicon Canals' framing of the Stanford finding gets at exactly this: it's "raising an uncomfortable question about who becomes tomorrow's senior analyst, lawyer or coder." That's not a rhetorical flourish. It's the actual mechanism of the 19% gap. You don't get a senior analyst who skipped being a junior analyst. You get a gap in the pipeline that shows up years down the line, once the industry actually needs people qualified to do the job the AI tools are supposedly going to need humans supervising.
This is where the two stories connect in a way that should worry you more, not less. Project OT's failure mode was that Meta tried to remove the senior-supervising-AI layer too fast and broke things a human would have caught. But the entry-level hiring contraction is quietly removing the pool of people who would have grown into that supervising layer in the first place. Fail fast on the flashy version, and you still have a slow-motion problem eating the foundation underneath it — a foundation that, unlike Project OT, nobody has a dashboard for.
It would be tidy if this were purely a bottom-of-the-ladder problem, something senior people could watch from a safe distance. It isn't. Pragmatic Engineer's reporting on why engineering leaders — CTOs, VPEs, heads of engineering — are voluntarily leaving high-status roles points at a related strain further up: one departing leader described managing "AI psychosis" with founders and executive peers as having become very difficult. That's a different failure than Project OT's metrics gap, but it's adjacent to it. It's what happens when the people setting AI strategy at a company are working from unrealistic expectations about what the tooling can absorb, and the person actually running engineering is the one who has to keep saying no, repeatedly, until saying no stops being worth the job.
Read together, you get a ladder under pressure at both ends. At the bottom, the rungs are quietly disappearing because the work that used to justify hiring onto them is cheaper to route elsewhere. At the top, some of the people who climbed the ladder when it still had all its rungs are stepping off it because the job now includes managing expectations set by people who never had to climb it at all. The middle is where the actual work still happens, increasingly without a robust supply line feeding it from either direction.
The comfortable assumption is that markets self-correct: if companies under-hire juniors for long enough, the resulting shortage of mid-level and senior talent will eventually get expensive enough that hiring rebounds. Maybe. But that correction, if it comes, arrives on a lag measured in years, not quarters — long enough for an entire cohort to have made career decisions based on the door being shut. Project OT corrected in months, because the feedback loop was fast and the failure was legible to the person who could reverse it. The entry-level hiring gap has no equivalent feedback loop. Nobody currently benefiting from cheaper output notices the missing junior hire as a cost, because the bill doesn't come due to them — it comes due to whoever needs a senior person, years from now, who doesn't exist because the entry-level req never got opened this year.
That's the actual implication, and it's not a call to panic about AI taking jobs in general — the evidence here doesn't support that broader claim, and Project OT's failure is a real data point against the version of this where AI cleanly substitutes for skilled labor at scale. What the evidence supports is narrower and, in a way, more useful: the visible, executive-driven attempts to replace workers with AI keep running into their own limits and getting walked back, while the invisible, distributed decision to simply stop hiring the people who would have grown into those workers keeps running with nothing to stop it. If you're trying to guess which version of the AI-and-jobs story is still going to be true when this one has faded from the news, don't extrapolate from the one that made Reuters. Extrapolate from the one that didn't need to.
The Long Game is a free Sunday essay on AI, work, and the future of wealth. If this made you think, forward it to one person who'd appreciate it.
Build your own?
The Autonomous AI Newsletter Pipeline Guide: GumRoad or Dodo