The Broken Ladder: Remote Work, Not AI, Is What's Sidelining New Gradu – Good Machine

TL;DR

The New York Fed published research on 1 June 2026 estimating that remote work explains 64% of the rise in unemployment among young college graduates since the pandemic — and that generative AI does not.
Unemployment for college grads under 29 rose from 3.1% (2017–19) to 3.7% (2022–25). For recent grads aged 22–27, it hit 5.6% in March 2026, up from 3.6% in March 2019.
The entire increase is concentrated in "remotable" jobs — software engineering, financial analysis, white-collar roles. In jobs that need a body in a room (nursing, mechanical engineering), the generational gap never widened.
The mechanism is mundane and important: junior employees learn by sitting near someone who knows what they're doing. Distributed teams quietly stopped budgeting for that.
A parallel LSE / Oxford working paper covering 243 million hires across the US, UK, Canada and Australia reaches the same conclusion. When AI exposure and remote-work exposure are tested in the same model, the AI effect collapses toward zero. The remote-work effect stays.
Most of the AI-is-eating-junior-jobs commentary you've been reading for two years is, on the current evidence, looking at the wrong variable.

Stop. This isn't the story you've been told.

For two years, the dominant narrative about entry-level white-collar hiring has been: AI is doing the work that new graduates used to do, so new graduates are not getting hired.

It's a tidy story. It fits the mood. It also appears to be mostly wrong.

On Monday, three economists — Natalia Emanuel at the New York Fed, Emma Harrington at the University of Virginia, and Amanda Pallais at Harvard — published a piece on the Fed's Liberty Street Economics blog with a number in it. Sixty-four percent. That, they argue, is the share of the recent rise in youth unemployment that can be explained by the four-fold expansion of remote work since the pandemic.

The same paper finds that when you control for AI exposure across occupations, the remote-work effect remains. When you control for remote-work exposure, the AI effect mostly vanishes.

This is not a small footnote. It is a category error inside the most-discussed career story of the decade.

The number, and what's underneath it

The headline statistic is the unemployment rate for college graduates younger than 29: 3.1% in 2017–19, 3.7% in 2022–25. A 20% relative rise. For the narrower 22–27 cohort the Fed tracks separately, 5.6% in March 2026, up from 3.6% in March 2019.

Here is where it gets interesting. The Fed economists split every occupation into "remotable" (software engineering, financial analysis, marketing analytics) and "non-remotable" (nursing, mechanical engineering, skilled trades). They then looked at the age gap in unemployment within each bucket.

In non-remotable jobs, the age gap spiked in 2020 — pandemic recession — then normalised. Nothing structural.

In remotable jobs, the gap opened and stayed open. Young workers' unemployment in those occupations rose by nearly a full percentage point. Older workers' unemployment in the same occupations marginally fell. The two lines parted ways and never came back together.

The aggregate rise in young-graduate unemployment, the Fed concludes, can be traced almost entirely to remotable occupations.

Why this happens — the mechanism most analysis skipped

The same authors previously studied software engineers at a Fortune 500 tech company. The relevant finding, in their NBER paper The Power of Proximity to Coworkers:

Before the pandemic, engineers who sat in the same building as their teammates received 22% more comments on their code than engineers on distributed teams in the same firm. Same company, same role, same seniority. Just closer.
Junior engineers were the primary beneficiaries. The people with the most to learn were the ones a desk away from someone who could explain why.
When everyone went remote in 2020, that feedback gap largely vanished — because everyone's feedback dropped.
Engineers who'd been trained alongside colleagues were roughly twice as likely to later move to higher-paying firms. The proximity was building human capital that distance quietly erodes.

Then, the firm-level fact that closes the loop. As that same company embraced remote work, it stopped hiring new graduates and shifted to hiring engineers roughly a decade older on average. Later, when the company introduced an aggressive return-to-office policy, it resumed hiring new graduates.

Not a hypothesis. A specific firm, observable behaviour, two policy regimes.

The wider corroboration

This is not one paper. A working paper from researchers at the London School of Economics and the University of Oxford, published last month, scanned 243 million new hires and 407 million job postings across the US, UK, Canada and Australia between 2017 and 2025.

What they found:

Entry-level hiring is down between 14% and 29% across those four economies.
Senior hiring is up between 5% and 21%.
AI exposure and remote-work exposure each, on their own, predict the fall in junior hiring.
When both are put into the same regression, the AI coefficient "attenuates sharply and often becomes statistically indistinguishable from zero" (the authors' words). Remote-work exposure remains a strong, robust predictor across every specification.

Two independent research teams, different data sources, different methodologies, same answer. That is the bar.

Who benefits from the AI story

If the evidence is this clear, why has the AI-replacing-juniors narrative dominated?

Because it suits almost everyone who has a microphone.

It flatters senior engineers — they can't replace me, they're going after the kids. It flatters founders raising AI rounds — look at the labour market we're already disrupting. It flatters journalists — the AI story writes itself, the remote-work story sounds like a return-to-office op-ed. And it conveniently absolves the people who designed the post-pandemic distributed-work playbook of any responsibility for what happened to the people downstream of it.

The remote-work story is more uncomfortable. It implies that a policy a lot of mid-career workers fought for — and still want — has a generational cost. That cost is being paid by people who weren't in the room when the decision was made.

This is the part the analysis usually softens. We won't. The Fed authors put it plainly: "remote work has weakened employers' incentives to hire young workers by impeding on-the-job training." Read that twice.

Cross-layer implications

A few connections that the entry-level-hiring framing tends to miss:

Career compounding. The earlier NBER paper found in-person trainees were ~2× more likely to move to higher-paying firms later. The damage is not the 12 months of unemployment after graduation. It is the missing 10 years of feedback-shaped skill compounding that begins in those 12 months. Lost early-career mentorship doesn't show up in the unemployment statistic. It shows up in the wage curve a decade later.

Sectoral selection. If remotable white-collar jobs are quietly closing to graduates, expect a measurable shift in where 22-year-olds land — toward roles that require physical presence (healthcare, skilled trades, in-person ops). This is already visible in apprenticeship application data in the UK and Australia. It is not a tragedy. It is a reallocation. But it is happening under a narrative ("AI is taking your job") that misdescribes the actual lever.

The RTO debate, re-cast. The Trump administration's 2025 return-to-office push for federal workers was framed (and resisted) as a culture-war move. The Fed's research does not vindicate the politics, but it does change what's at stake. A blanket rejection of RTO is also a rejection of the structure that historically transmitted early-career skill. The honest answer is more granular than either side has wanted to admit. (Note for the calendar: on 3 June the same administration relaxed those rules for federal employees in 11 World Cup host cities — a small but telling reminder that none of this is settled.)

Geography. Remote work expanded fourfold; the cities that lost the most office foot traffic are the cities whose entry-level hiring contracted hardest. The early-career market is increasingly bifurcated by employer policy, not by zip code.

What the evidence says you should actually do

Addressed here to new and recent graduates, and the people advising them. Not to employers. Employers have their own paper to write.

Re-rank your employer shortlist by onboarding architecture, not brand. The variable that matters now is whether the employer has budgeted real money for developing juniors. Signals: a named graduate programme, structured rotations, dedicated mentor pairings, cohort cohorts of new hires arriving together, in-person or structured-hybrid onboarding for at least the first 6–12 months. A company that says "we're fully distributed and you'll figure it out" is, on this evidence, a worse early-career bet than a less prestigious employer that runs a real graduate scheme.
Read job descriptions for the mentorship signal. Language about pairing, code review cadence, weekly 1:1s with named senior engineers, in-person collaboration days, structured ramp plans — these are the green flags. The phrase "fast-paced async environment" is, for a new graduate, often a euphemism for "you will not be taught anything".
If you're already employed remotely as a junior, manufacture the proximity. The Fed mechanism is feedback frequency, not desk location. Ask for weekly code reviews by name, request to be invited to senior debugging sessions, schedule recurring 30-minute walkthroughs of work in progress. The structural problem is that distributed teams don't spontaneously generate this. Don't wait for spontaneity.
Don't over-rotate to "I'll just learn AI." AI fluency is necessary table stakes; on the evidence it is not the binding constraint on getting hired. The constraint is that distributed-team employers don't want to train you at all. Becoming better at prompt-engineering doesn't fix that.
For policy-makers and university careers services. The intervention that maps to the evidence is structured apprenticeship and graduate-programme subsidy — paying down the cost-of-training disincentive that remote teams have absorbed. Generic "AI upskilling" grants do not address the mechanism the research identifies.
The blunt one. If you have a choice between a hot remote-first AI start-up and a less glamorous employer with a real graduate programme, the data on this page says the second is more likely to give you a career. The first is more likely to give you a LinkedIn post.

What we still don't know

Hybrid as solution. The Fed paper compares remotable to non-remotable occupations; it does not finely distinguish 5-day-in-office from 2-day-in-office from 4-day-in-office regimes. We don't yet know the dose-response curve for in-person days versus junior skill development.
AI's lag. Both research teams stress that AI effects could rise. The current finding is that as of 2024–25 data, AI doesn't explain the gap. By 2027 data it might.
International transferability. The LSE/Oxford paper extends beyond the US, but the granular firm-level evidence is from a single Fortune 500. How well the mechanism generalises across industries — finance, consulting, marketing, design — is a live research question.
Selection effects. Some share of remotable-job graduates may be self-selecting into remote-first companies that were always less likely to invest in junior development. Disentangling that is unfinished work.

What would change the analysis: hybrid-specific data, longer post-2024 timelines, and firm-level evidence outside tech.

Bottom Line

The dominant story about why entry-level white-collar hiring has collapsed is mostly wrong, and the correct story is harder to look at. Remote work — the privilege mid-career workers fought to keep — has quietly raised the cost of training the next cohort, and rational firms have responded by hiring fewer of them. The graduate strategy that follows from the evidence is small, specific, and unsexy: pick the employer that still budgets for mentorship, and forget the AI panic.

Sources

Tier 1. Emanuel, N., Harrington, E., & Pallais, A. (1 June 2026). "Remote Work Leaves Younger Workers Sidelined." Federal Reserve Bank of New York, Liberty Street Economics.
Tier 1. Federal Reserve Bank of New York Media Advisory (28 May 2026). "New York Fed to Release Research on the Role of Remote Work…"
Tier 1. CNBC (1 June 2026). "Remote work is worsening youth unemployment, New York Fed finds."
Tier 1. NPR (1 June 2026). "Remote work — not AI — has sidelined recent college graduates."
Tier 1. AP via Bozeman Daily Chronicle (1 June 2026). "Young and unemployed? Remote work, not AI, may be the problem, study finds."
Tier 1. Fortune (2 June 2026). "New York Fed research is just one of many saying remote work is behind the real Gen Z hiring nightmare."
Tier 1. Bloomberg Law (3 June 2026). "Federal Agencies Can Permit Telework During FIFA World Cup."
Tier 2. Emanuel, Harrington, Pallais. "The Power of Proximity to Coworkers." NBER Working Paper (firm-level evidence cited in NY Fed post and Fortune).
Tier 2. Lambert & Schindler (May 2026). LSE / University of Oxford working paper on AI vs. remote-work exposure across 243M hires and 407M postings in US/UK/Canada/Australia.
Tier 2. Business Insider (2 June 2026). "AI Isn't Taking Gen Z's Jobs. Blame Remote Work, Says the Fed."
Tier 2. The Register (2 June 2026). "Remote work — not AI — is killing job prospects for the youth."
Tier 2. TheStreet (3 June 2026). "New York Fed reveals troubling truth for young workers."