16APR

Two federal studies split on AI jobs

4 min read

13:29UTC

The New York Fed found minimal sign AI is cutting hiring and said the decline in exposed roles began before ChatGPT shipped, 30 days after Stanford put the figure at 34 prevented hires per declared layoff.

← Back to Three federal surveys, one 34-to-1 gap Jump to analysis ↓

EconomicDeveloping

Key takeaway

Until a federal agency designates one canonical measure, any AI-jobs law sits on contested foundations.

The Federal Reserve Bank of New York published an occupation-level study of job-posting data on 14 May 2026 reporting minimal evidence that artificial intelligence (AI) has directly cut labour demand since ChatGPT shipped in December 2022. ¹ The paper, by NY Fed economists on the bank's Liberty Street Economics blog, found the relative fall in postings for AI-exposed roles began before that launch, and that junior and senior roles declined at similar rates. ²

That reading sits directly against the Stanford Digital Economy Lab, which 30 days earlier read the aggregate hiring rate and attributed roughly one million prevented US hires a year to AI, a figure 34 times the declared layoff count . Stanford concentrated the damage on the under-25s; the NY Fed found no such age cliff. A worker entering the market is now told by two credible institutions, a month apart, both that the entry-level cliff is real and that it does not exist.

Each team reached its answer by choosing a different dataset, not by disputing the underlying numbers. Stanford reads JOLTS (the Job Openings and Labor Turnover Survey, a monthly federal count of openings, hires and separations) and infers AI causation from a collapsing hiring rate. The NY Fed reads occupation-level postings and finds the decline predates the technology, which removes the timing evidence Stanford's causal story rests on. Less than 10% of workers sit in high-exposure occupations and 40% sit in jobs with no measured exposure at all, so most of the workforce is nowhere near the frontier the headlines describe. ³

There is no official answer to arbitrate between them. The Bureau of Labor Statistics (BLS) has published nothing on the question , and the Federal Reserve Board has already shown three federal surveys putting AI adoption at 18%, 41% and 78% for the same period . A government that cannot agree how many firms use AI has now been handed two opposite readings of whether it is costing jobs. Every layoff below has to be weighed against that vacuum.

Deep Analysis

In plain English

When people ask whether AI is taking jobs, two things can be measured: how many workers have been laid off and blamed AI (explicit layoffs), and whether companies are hiring fewer people than they used to (hiring suppression). The Federal Reserve Bank of New York, a branch of the US central bank, studied job adverts by occupation and found that roles exposed to AI had fewer postings than before, but the decline started before ChatGPT launched in December 2022. That timing is important: if hiring was already falling before AI tools became widely available, it suggests other causes (automation, economic cycles) may be driving the trend. Stanford University's Digital Economy Lab looked at a different dataset called JOLTS (Job Openings and Labor Turnover Survey), which the Bureau of Labor Statistics, the US government agency responsible for employment figures, publishes each month. Stanford calculated that the hiring rate is so depressed that AI is effectively blocking about one million new hires a year, even without explicit layoff announcements. Both studies are published by respected institutions and both could be correct about different things. The NY Fed says post-2022 AI is not the cause; Stanford says someone or something is blocking a million hires a year. The Bureau of Labor Statistics, which should be the official referee, has still published nothing on the question. Until it does, the policy debate about protecting workers from AI displacement has no agreed factual baseline.

Deep Analysis

Root Causes

The divergence traces to a structural gap in US federal labour statistics. The Bureau of Labor Statistics runs JOLTS as an aggregate survey, not an occupation-level or cause-attributed series.

Stanford's team applied a valid econometric technique to aggregate data and produced a displacement estimate; the NY Fed applied a different valid technique to occupation-level microdata and produced a contradictory one. Neither is wrong on its own terms; both reflect the absence of a mandated, cause-attributed federal AI workplace series.

The Federal Reserve Board's own survey-reconciliation paper, showing federal instruments returning 18%, 41% and 78% AI adoption for the same period , documents that the measurement infrastructure itself has not been calibrated. Any policy built on either study inherits that infrastructure failure as a design constraint. The BLS's April 2026 skip of its scheduled GenAI workplace paper left that calibration gap open at the moment it most mattered.

What could happen next?

Risk
Policy interventions designed on the NY Fed baseline (minimal displacement) will undershoot if Stanford's hiring-suppression figure is correct; the measurement gap creates a systemic underresponse risk.
Precedent
The BLS absence from this measurement debate sets a precedent where regional Fed banks and private universities set the de facto federal evidence base for major labour policy questions.

Source Landscape

This story draws on neutral-leaning sources

Primary parallel: The Bureau of Labor Statistics' measurement of personal-computer displacement in the 1980s and 1990s produced a nearly identical methodological split.

BLS occupation-level data showed clerical employment holding up through the PC adoption period; Federal Reserve Bank of Chicago economist Daniel Aaronson and colleagues later established using matched employer-employee data that PC adoption suppressed hiring by 15-20% in exposed occupations over the same window, a signal invisible in the occupational-aggregate series the BLS published at the time. The NY Fed vs Stanford divergence replicates that structure: two compatible datasets, two opposite readings, a missing third dataset that would settle the question.

Counter-parallel: The 2013 Frey-Osborne Oxford study projected 47% of US occupations at high risk of computerisation; BLS occupation-level outcome data over the following decade showed nothing close to that figure realise. That precedent suggests occupation-level exposure metrics systematically overestimate displacement and supports the NY Fed's occupational-posting approach as the more conservative and historically reliable measure.

Less than 10% of US workers sitting in high AI-exposure roles, with 40% in zero-exposure jobs, implies the AI displacement story is highly concentrated in a narrow slice of the labour market: knowledge-work occupations, technology roles, and administrative functions.

If Stanford's one-million-hire suppression figure is correct, it falls almost entirely on that 10%, translating to a 15-20% hiring reduction for exposed workers, which is consistent with the 16% employment decline in young AI-exposed workers documented by Stanford's own cohort analysis.

The measurement gap has direct fiscal consequences. US unemployment insurance eligibility depends on demonstrated layoff, not hiring suppression. A worker who was never hired generates no UI claim and no WARN Act filing, making the Stanford-hypothesised million suppressed hires structurally invisible to both the benefits system and the disclosure regime.

Until the BLS produces a cause-attributed series, federal spending models for AI workforce transition are operating without a verified denominator.

Consensus view: MIT economist David Autor and co-authors Brendan Price and Christina Patterson, in their 2024 labour-market review, argued that occupation-level job-posting data and aggregate hiring flows measure structurally different phenomena: postings capture employers' near-term slot demand, while JOLTS hiring rates capture whether matches are actually clearing.

A divergence between the two is consistent with a market where positions exist but AI-enabled workflow changes mean each position absorbs more output per hire, suppressing throughput without eliminating postings.

Counter-view: Dario Acemoglu (MIT) and Pascual Restrepo's earlier automation-displacement work (2022) identified that the share of occupations in structural decline tracks closely with the period of technology adoption, not the period of product launch; the NY Fed's finding that exposure-weighted posting declines predate ChatGPT is consistent with Acemoglu's model, not contrary to it.

Pravin Krishna at Johns Hopkins, drawing on comparative offshoring literature, notes that similar pre-launch patterns appeared in routine-task automation in the 2010s and were later confirmed as structurally causal once multi-year panel data became available.

Key tension: Whether the pre-ChatGPT timing of the NY Fed decline exonerates current AI from causing displacement, or whether it documents an earlier adoption wave that the 2022 ChatGPT moment then accelerated rather than originated.

Sources:Federal Reserve Bank of New York

Mentions:Federal Reserve Bank of Dallas →Stanford Digital Economy Lab →Bureau of Labor Statistics →Federal Reserve →ChatGPT →Liberty Street Economics →Job Openings and Labor Turnover Survey →

First Reported In

Update #10 · Rival studies split on AI's hit to jobs

Federal Reserve Bank of New York· 24 May 2026

Read original →

Causes and effects

Caused by

Stanford: AI costs 34 hires per layoff

NY Fed's occupation-level study was published in direct response to the Stanford JOLTS finding that AI prevents approximately one million US hires annually, with the two studies now split on whether AI is costing jobs.

Occurred 10 Apr 2026

Read story →

BLS skips; NY Fed fills the vacuum

The BLS skipping its GenAI workplace paper left the measurement vacuum that the NY Fed's Liberty Street publication is now attempting to fill.

Occurred 14 Apr 2026

Read story →

Fed admits: no single AI adoption figure

The Fed Board's three-survey reconciliation showing 18%, 41% and 78% adoption for the same period compounds the NY Fed's methodological split into a federal-measurement crisis.

Occurred 3 Apr 2026

Read story →

This Event

Two federal studies split on AI jobs

The case for AI-jobs intervention lost its agreed evidence base the same fortnight it lost its strongest law.

Different Perspectives

TSMC and Taiwan chip supply chain

Nvidia's 17% headcount growth to 42,000 on $81.6 billion in quarterly revenue depends on TSMC's CoWoS advanced packaging capacity constraining H100 and B200 supply, sustaining margins above 70%. The AI build-out's sole headcount-growth story runs through a Taiwan supply chain that has no parallel in downstream software.

Displaced tech workers globally

CrowdStrike's SEC disclosure puts AI attribution on a material regulatory record for the first time, but Oracle's Massachusetts WARN clock expired unfiled after up to 14 workers were logged as remote despite office proximity. The legal apparatus cannot enforce what it cannot see: hybrid reclassification, GCC transfers, and hires never made.

UK workforce and policymakers

ONS recorded UK vacancies at 705,000, below the pre-pandemic baseline for the first time, as payrolled employment fell 210,000 year on year with real wage growth at 0.1%. The Bank of England's AI worst case assumed 500,000 additional unemployed from a baseline above 730,000; the UK is already below that floor, and ONS still publishes no AI-exposure breakdown.

India IT workforce and graduates

NASSCOM's FY2026 data shows net sector growth of 140,000, but entry-level hiring fell 20-25% as the growth concentrated in in-house GCC offices requiring mid-career specialists. Indian graduates who previously entered through TCS, Infosys and Wipro fresher programmes find that channel closing at both ends: outsourcers cutting and GCCs not hiring at the junior level.

IG Metall and European trade unions

European labour bodies see the market reward pattern, cuts on record revenue, as investor preference for short-term margin extraction over validated AI productivity. They note the EU Digital Omnibus provisional deal has dropped binding employer AI-literacy obligations at the precise moment the ILO-NASK index has quantified that 3.3% of global workers are in the highest AI exposure category.

Federal Reserve Board

Governor Cook told Stanford's SIEPR on 27 May that speculative-grade software bond spreads have widened on AI-disruption concern, moving AI displacement from a labour observation into the Fed's financial-stability mandate. The Fed cannot resolve structural labour transformation through rate policy, so Cook routed the concern through the one channel the Fed does control.