We Blamed AI for Killing Junior Jobs for 3 Years. We Were Wrong.
A 243-million-hire study just named the real culprit. It's been hiding in plain sight since 2020.
Everyone's been yelling that AI is replacing juniors. In 2019, 1 in 3 tech hires was someone under 25. In 2025, it's 1 in 5. Stanford pulled the ADP payroll records. Harvard followed. AI was in the accused box and the verdict looked settled.
It wasn't.
A study from LSE, Warwick, and the Oxford Ellison Institute (243 million new hires, 407 million job postings, 4 countries, 8 years of data) points to a different culprit, much less visible and far more damaging: remote work. When Lambert and Schindler add that variable to the model, the AI coefficient collapses to statistical zero. Remote is the culprit. AI is not.
TLDR: The Lambert-Schindler study across 243M hires breaks the "AI = junior killer" narrative: when you control for remote work, the AI coefficient drops to statistical zero. What destroyed junior ROI in remote isn't a tech problem, and that changes everything about what you can actually do about it.
What's actually killing junior hiring is less headline-friendly and much harder to fix: remote broke the osmosis learning circuit that made a junior worth hiring in 18 months. Companies did the math. Junior ROI in remote is negative. They stopped hiring. No anti-junior ideology anywhere in that calculation. It's pure economics.
1 in 3. Then 1 in 5.
The shift started quietly. Between 2019 and 2025, the share of entry-level workers under 25 in new tech hires dropped from roughly 1 in 3 to 1 in 5. This wasn't a small correction. It's a structural rewrite of who gets a first job.
Lambert and Schindler track this across the US, UK, Canada, and Australia simultaneously. Same pattern, different markets, same timeline. The decline accelerates after 2020 and doesn't recover. By 2024-2025, the share of job postings not requiring prior experience also dropped by roughly 3 percentage points. That sounds small. It isn't: 407 million postings over 8 years, 3 points across all of them is a generation of entry-level positions that quietly disappeared from the board.
Ask any hiring manager who was posting entry-level roles in 2018 and doing the same in 2024. The conversation didn't change overnight. It shifted in small increments, each one individually defensible: "we need someone who can hit the ground running," "we're a small team and can't absorb the ramp-up time," "the role has evolved." Nobody announced a policy change. The policy emerged from countless individual decisions that all pointed the same direction.
The scale of the study matters here. 243 million new hires is larger than most longitudinal labor datasets by an order of magnitude. Stanford's ADP study was a sample. Harvard's was firm-level. Lambert and Schindler are working with something closer to the full picture. And the full picture says the culprit you think you know is wrong.
The Story Everyone Believed
To be fair to Stanford and Harvard: they weren't inventing the signal.
Stanford researchers using ADP payroll records found a 13% relative employment decline for workers aged 22-25 in the most AI-exposed roles since late 2022. A Harvard study found a 7.7% junior headcount reduction in companies adopting AI across 6 quarters from early 2023. Both sets of numbers hit the press with the force of confirmation. Everyone already suspected it. Now there was data.
The narrative was clean: AI handles the routine cognitive work that entry-level jobs used to provide. Companies see the automation potential, stop hiring for those roles, shift headcount toward experienced people who can direct and manage the tools. Juniors, whose value came from doing the routine work, find themselves competing for a smaller pool of real entry points. Economically coherent, fits the observed pattern, and both studies were methodologically solid for what they were measuring. The Stanford paper used actual payroll records, not surveys. The Harvard paper tracked firm-level headcount changes longitudinally. This wasn't sloppy research.
The problem isn't that the press exaggerated. The problem is that good researchers made a classic methodological error on a topic the whole world was primed to accept as true.
What the Model Was Missing
Omitted variable bias.
AI exposure and remote work exposure are not independent variables. The jobs most exposed to AI (routine cognitive tasks: code triage, data entry, customer queue management, document processing) are almost exactly the jobs most compatible with remote work. Build a model that only includes AI exposure and you're not controlling for the fact that remote work exploded at the same time and hit the same job categories.
The Lambert-Schindler fix is methodologically obvious in retrospect: add WFH exposure to the model. When they do, the AI coefficient "attenuates sharply and is often statistically indistinguishable from zero." WFH holds. AI doesn't. That's not a small adjustment. That's a different cause entirely.
The NY Fed data makes the boundary precise. The employment gap between workers under 28 and those 29 and above (roughly 1 percentage point of unemployment) exists almost exclusively in sectors where remote work is structurally common. In sectors where remote isn't possible (manufacturing, healthcare delivery, trades), the gap is near zero. AI exposure doesn't predict the gap. WFH does.
There's something I keep thinking about that has nothing to do with any of this. My kid asked me last weekend why some games save automatically and some don't. The ones that don't, you lose everything when the power cuts. The junior job market post-2020 feels exactly like a game that quietly removed autosave and never patched it back in. No recovery checkpoint. You just restart from the character creation screen.
The studies that missed this weren't wrong about their numbers. They were wrong about what those numbers were actually measuring.
Remote Cut the Learning Circuit
The mechanism isn't abstract. NY Fed economist Natalia Emanuel, working with Emma Harrington at UVA and Amanda Pallais at Harvard, tracked software development teams through the remote transition. Senior developers' code quality was essentially unchanged when they went remote. Junior developers' code quality dropped measurably. Code churn went up, bugs increased. Same physical distance from colleagues, radically different outcomes by experience level.
They replicated the pattern in customer support teams. Juniors going remote: longer resolution times, more calls per issue. Seniors going remote: barely moved.
The explanation is osmotic learning. A senior developer has already internalized the patterns. They know when a function is getting too complex before the linter fires. They can read a pull request and feel the architecture debt without running the code. They carry thousands of hours of ambient feedback baked into how they write and debug. None of that is transmittable through documentation or async code review: it was absorbed over years of physical proximity, overhearing a conversation about a production incident, watching someone fix a race condition in real time, and being close enough to ask the right question at the exact moment it makes sense. That circuit requires presence, not constant presence, but enough sustained proximity that the ambient signal stays continuous. Remote work doesn't just relocate the learning. It cuts the wire.
Research published in The Quarterly Journal of Economics confirms this at scale: in-person onboarding raises later productivity and reduces attrition even for employees who subsequently return to remote work. The gains are largest for younger workers. Proximity to colleagues is what drives feedback for juniors, not mentorship programs or structured review cycles.
Companies aren't running an anti-junior policy. They're running a P&L calculation. A junior in-person contributes meaningfully within 12-18 months. A junior in remote is a longer, more uncertain bet with measurably higher error rates on the way there. When every open role is remote by default, the expected value of hiring junior drops below the threshold worth taking. So they hire senior.
Blocked From the Top Too
The junior squeeze isn't only about entries drying up. The ladder is blocked from both ends.
According to Placer.ai foot-traffic data cited in a16z's June 2026 analysis, office visits are currently at 70% of pre-pandemic levels. Plateaued. Return-to-office mandates moved the needle briefly in 2023, then stalled. Office vacancy rates are above 14%, the highest since the 2008 financial crisis. Which means even companies that want the learning circuit back are operating at 70% of the physical conditions that made it work.
Senior employees aren't cycling out either. In law, finance, consulting, and media, tenure has been trending longer since 2023. The career pathways that used to open when seniors moved on, got promoted, started companies, or retired are moving slower. Fewer seats opening at the top means fewer positions cascading down through the middle and into entry-level. The double lock is structural.
The political economy of this sustains itself almost automatically. Companies avoid the uncomfortable conversation about whether their remote-by-default setup is actually compatible with developing talent. Politicians have a clean technological scapegoat instead of a messy management problem. Stanford and Harvard published results on a real signal that turned out to be non-causal, and the press ran a narrative that confirmed what everyone already believed. Nobody needed to be cynical for this to become the consensus. The mechanism was simpler: a real signal, a non-causal interpretation, and a narrative the entire world was already primed to accept.
For juniors navigating this, the algorithmic layer is its own game and it runs long before a human reads anything. If you want to understand how to game the algorithmic recruiting layer, the tactics are more concrete than most people realize.
2 Ways Out of a Broken Ladder
There are 2 exits. They're not for the same person.
The institutional exit is real. IBM in 2026 tripled its US entry-level hiring by explicitly redesigning junior roles around AI augmentation rather than treating AI as a replacement. The framing matters: instead of "we need fewer juniors because AI does their work," the bet was "we need juniors who work alongside AI from day one, and that's a trainable skill." The share of graduate postings in AI-exposed roles started recovering in early 2026 after 2 years of decline. The companies that crack the remote mentorship problem win the junior talent war while everyone else sits out. If you're on the hiring side, this is the 1 decision that compounds over the next 5 years.
I think this recovers faster than the pessimists say. Could be I'm reading early signals too optimistically. But the IBM counter-example is real, and it suggests the problem is soluble at the company level without waiting for a macro shift in the labor market.
The individual exit doesn't wait for institutional reform. For everyone looking at the queue and doing the math, the ladder being structurally blocked is actually clarifying. If the entry path is closed and the timeline is indefinite, the question stops being "how do I get onto the ladder" and becomes "why am I looking for a ladder in the first place."
The vibe-coder isn't someone who gave up on the market. That's someone who looked at the economics, understood that the return on waiting is negative, and decided to build the thing instead of waiting for permission to join the team that builds it. In 2026, that path is concrete in a way it wasn't 4 years ago. 1 shipped product beats 50 applications into the algorithmic void.
If you're starting down that path, CLIs beat MCP for production AI agents is worth understanding before you hit the production wall. And for the method itself, Vibe Coding, For Real is the 8-step Blueprint for builders who've hit the demo wall and want to ship something that actually runs.
We spent 3 years blaming an algorithm for what was actually a management policy. Convenient scapegoat: AI doesn't vote, doesn't give interviews, and doesn't defend itself in front of a committee.
The ladder is broken. The real lock isn't technological, it's managerial and economic. The companies that solve it, by explicitly redesigning junior roles around AI augmentation, are picking up the best entry-level talent while everyone else debates whether the market is "normalizing."
IBM did it. The signal is there.
You can wait for your potential employer to read the same study. Or you can stop applying and start vibe coding like a pro. C'est la vie.
Sources
- Lambert, C. & Schindler, S. (2026). The Broken Ladder: AI, Remote Work, and Early-Career Hiring. LSE/Warwick/Oxford Ellison Institute. SSRN
- Sternstein, M. (2026, June 5). Charts of the Week: RTO Stalled. a16z. a16z.news
- Remote work, not AI, is killing job prospects for the youth. The Register, June 2, 2026. theregister.com
- The real reason junior hiring is collapsing may not be AI. HCA Magazine, June 2026. hcamag.com
This post may contain affiliate links. If you click them, I might earn a small commission (costs you nothing, and helps me keep shipping quality articles every day for your reading pleasure).
Remote work—not AI—collapsed junior hiring economics. When you control for it, the AI signal disappears. The welcome kit's demo-vs-product checklist shows you how to spot when your AI hiring decisions are actually driven by economics, not technology.