The Wrong Diagnosis

The story most people have absorbed about young graduates and AI goes like this: generative AI is eating entry-level knowledge work, and the first victims are the twenty-something analysts, paralegals, and junior coders who used to do the tasks that AI now handles more cheaply. A Federal Reserve Bank of New York analysis published this week says that story is mostly wrong about the cause, if not about the consequence. The mechanism compressing opportunity for young college graduates is not AI displacement. It is remote work.

The NY Fed's Liberty Street Economics team looked at the rise in unemployment among college graduates under 29. The rate moved from 3.1 percent (the 2017 to 2019 average) to 3.7 percent (the 2022 to 2025 average), a 0.6 percentage point increase that is modest in absolute terms but striking for a group that is normally well-insulated from joblessness. Their back-of-envelope decomposition attributes roughly 64 percent of that rise to the structural shift toward remote and hybrid work.

The mechanism is straightforward once named. Entry-level workers depend on proximity in ways that established workers do not. They need visibility to get assigned to visible projects. They need managers nearby to pick up on what they don't know yet. They need informal sponsorship to turn competence into advancement. Remote work compressed all of that, and it compressed it most for the workers who needed it most. The workers with established networks, professional reputations, and institutional standing kept their footing. The entrants without those things found themselves structurally disadvantaged in ways that the job listing or the acceptance letter did not warn them about.

The workers who most needed proximity lost it precisely because the workers with power over workplace structure no longer needed it themselves.

This is a distributional fact, not a technological one. The shift to remote work was decided by incumbents. The costs of that shift landed on entrants.

Who this falls hardest on. The distributional shape of the remote-work harm is not random. First-generation college graduates arrive without the professional networks that smooth entry for graduates from established families. Graduates from less prestigious institutions can't trade on institutional name recognition in the way that a degree from a target school once opened doors. Workers from underrepresented backgrounds in knowledge industries have fewer informal champions to compensate when formal sponsorship structures weaken. Remote work didn't create these inequalities, but it amplified them, by removing the daily visibility through which junior workers without pre-existing advantages had historically managed to make themselves legible to the people making promotion and retention decisions.

The narrative that got ahead of the data. The competing account comes from Erik Brynjolfsson and collaborators at Stanford, whose "Canaries in the Coal Mine" framing argued that young, educated workers in occupations with high AI exposure were experiencing relative employment deterioration. That framing drove a significant press and policy cycle. It is not wrong that AI-exposed occupations have seen structural shifts, and the Brynjolfsson correlation is real. But the NY Fed timing argument is important: the rise in young-graduate unemployment began before generative AI tools were in widespread workplace deployment. ChatGPT launched in November 2022. The labour market signal predates it.

There is a related pattern in corporate communications that makes this worse. Major tech firms in 2023 and 2024 repeatedly cited AI efficiency as justification for headcount reductions, even in cases where the restructuring decisions predated meaningful AI deployment. The firms got credit for being forward-looking and cost-disciplined; workers got let go. The AI-as-cause framing was useful for management communications whether or not it was accurate. The result is an evidentiary baseline that has been partially constructed by parties with an interest in a particular narrative.

The back-of-envelope caveat. The NY Fed analysis is explicit about what it is. The 64 percent figure is a decomposition, not a regression-based causal identification. The authors are not claiming to have ruled out AI; they are claiming that a parsimonious reading of the timing and the data points toward remote work as the primary driver. That is worth taking seriously. It is also worth holding with some care, because the 64 percent number will circulate in coverage with the same confident treatment that the Brynjolfsson findings received. The media cycle for labour analysis has a strong tendency to flatten "here is the best reading of the available evidence" into "here is what caused this." The NY Fed authors seem aware of this risk; their framing is cautious. Coverage of their finding may not be.

The sequence problem. Both analyses can be partially right, and that is actually the more uncomfortable possibility for policy. Remote work explains the documented past. AI may explain the near future. The tools that Brynjolfsson's occupational-exposure analysis was tracking are now in wider deployment than they were when the unemployment signal first appeared. The window for addressing the remote-work mechanism — through changes to onboarding norms, to entry-level hiring structures, to hybrid-work policy for junior employees — is open now. If legislators and firms spend the next three years debating AI displacement, that window closes, and the actual transition to AI-driven labour substitution will arrive with no managed response already in place.

Who was asked. I find it notable that this entire debate, in both its academic and policy dimensions, is conducted almost entirely among people who are not the young graduates absorbing the costs. Neither the remote-work transition nor the wave of AI tooling deployment was decided with entry-level workers' input. The question of which employer adopted which working arrangement was decided by the people who already had the jobs. The question of which workflows get automated is decided by the firms deploying the tools. The workers who enter a labour market shaped by these decisions are not parties to them.

That is not an argument against remote work or against AI adoption. It is a statement about whose experience is treated as evidence and whose experience is treated as outcome. The NY Fed analysis is a useful corrective to an AI-displacement narrative that outran its data. What neither paper addresses is that whether the cause is remote work or AI, the people least consulted about the structure of the labour market they are entering are the ones paying the highest price when that structure shifts.

The attribution question matters for policy. It matters more that neither answer, so far, has included the people most affected in the conversation about what to do.

Footnotes

¶