Stanford’s Institute for Economic Policy Research published a policy brief on July 24 that asks the question everyone in AI policy has been circling: if AI is destroying white-collar work, where is the evidence? The answer, after crunching U.S. labour market data through 2026, is that there isn’t any — not yet, and not at the scale the loudest voices predicted.
What is SIEPR? Stanford’s Stanford Institute for Economic Policy Research is one of the most cited economic policy shops in the United States. Its digital economy lab, led by Erik Brynjolfsson, has been tracking AI’s labour-market impact since before ChatGPT existed. When SIEPR publishes a jobs brief, policymakers and fund managers read it.
🔍 THE BOTTOM LINE
Unemployment for the most AI-exposed workers has risen 0.77 percentage points since 2022. For the least-exposed workers, it rose 0.85. If AI were uniquely destroying exposed jobs, those numbers would diverge. They don’t. The labour market is softening broadly, not selectively.
The Numbers That Kill the Headline
The brief’s central chart tracks unemployment by AI-exposure quintile — how likely a job’s tasks can be automated by AI. If the “AI jobs apocalypse” thesis were correct, the most-exposed workers (quintile 5) would show accelerating unemployment while the least-exposed (quintile 1) stayed stable.
The data shows the opposite of a divergence. Both groups rose, and the least-exposed rose slightly more. This is consistent with a general labour market cooling — the kind that follows interest rate hikes — not a targeted AI disruption.
The brief also notes that employment in highly AI-exposed occupations has been fairly stable. Software developer job postings have been growing faster than other occupations over the last year. Among firms that adopted enterprise AI, employment grew 10 percent in the two years following adoption, with the strongest growth at firms with the highest per-capita AI spending.
The Canary in the Coal Mine
There is one genuinely concerning signal. A widely discussed paper by Brynjolfsson, Chandar, and Chen found a notable decline in employment among early-career workers in AI-exposed occupations — particularly software developers and customer service reps — since ChatGPT’s launch in late 2022. Employment among older workers in the same occupations stayed stable or grew. The authors called these young workers “canaries in the coal mine.”
Recent graduates face the toughest job market in years: 5.6 percent unemployment for new grads, up 1.6 points from three years ago. Junior roles involve the routine research, analysis, and writing tasks that AI tools handle best.
But the brief is careful about causation. The timing is awkward: AI model capabilities were very limited in 2022, yet the decline in entry-level hiring begins around then. Two new papers found that hiring in AI-exposed occupations began declining after the Federal Reserve’s March 2022 rate hikes — months before ChatGPT’s November 2022 release. The shift to remote work during the pandemic, which slows on-the-job learning, also eroded the value of hiring junior workers. When Brynjolfsson added controls for these confounders, the employment decline for entry-level workers wasn’t notable until 2024 — when both AI adoption and model capabilities had genuinely advanced.
What AI Actually Does to Productivity
The brief catalogues what controlled experiments show AI tools doing to worker speed: GitHub Copilot let developers complete tasks 56 percent faster, with gains concentrated among less-experienced programmers. A generative AI assistant in a call centre boosted productivity 15 percent, with novices seeing 30 percent improvements. ChatGPT reduced writing time across all ability levels and improved quality for low-ability writers.
But these experimental gains haven’t shown up in aggregate productivity statistics. The brief offers several reasons: the share of tasks AI can profitably speed up may be small relative to total economic activity; process bottlenecks limit how much individual task speed translates to organisational output; and firms adopting new technologies often see a temporary productivity dip as they reorganise.
There’s also the jagged-capability problem. AI performance can be strongly positive or negative depending on the specific task. A study of Kenyan entrepreneurs found that less-skilled business owners posted lower revenues from using an AI tool because they acted on generic advice that was wrong for their situation. The tool wasn’t broken — the judgment to deploy it well was missing.
The AI Adoption Curve Is Still Early
U.S. Census Bureau data shows AI adoption at work rising steadily through 2026, but the absolute share of firms meaningfully using AI remains modest. If only 10-20 percent of firms are deploying AI at scale, you wouldn’t expect to see aggregate labour-market effects yet — regardless of how powerful the technology is.
This is the steelman of both sides. The AI-is-disrupting-jobs camp can say “the effects are concentrated in entry-level workers, and the lag between adoption and aggregate impact means we’re still early.” The AI-is-overhyped camp can say “three years after ChatGPT, the aggregate data shows nothing, and the confounders are large.” Both are partially right.
NZ Angle
New Zealand’s labour market tracks the U.S. with a lag, and our policy debate mirrors the American one. The same Dario Amodei claim — AI could wipe out half of white-collar jobs — that drove the Guardian analysis we covered this morning is the same claim Stanford’s data now pushes back on. For NZ policymakers, the Stanford brief’s most useful insight is the distinction between aggregate impact (none yet) and distributional impact (entry-level workers are bearing whatever burden exists). If AI is going to hurt anyone first, it’s the 22-year-old graduate, not the 45-year-old senior developer.
❓ FAQ
If AI isn’t causing job losses, why are companies announcing AI-driven layoffs? Stanford notes that industry leaders and labour economists are skeptical of layoff claims that cite AI. Many appear driven by a desire to free up cash for AI investment or to correct pandemic-era over-hiring. HR executives say AI’s impact is more visible in role consolidation and hiring avoidance than in mass redundancies.
What about the 56% Copilot speed gain? It’s real and measured. But task-level speed doesn’t automatically translate to organisational productivity, and the brief notes that the share of economically relevant tasks AI can speed up may be small. A developer who writes code 56% faster still spends time in meetings, code review, debugging, and deployment.
Does this mean AI won’t disrupt jobs? No. It means the disruption hasn’t arrived at aggregate scale yet, and what has arrived is confounded by other factors (interest rates, remote work, pandemic over-hiring). The entry-level signal is real and growing. The question is timing, not direction.
Should NZ graduates be worried? The data says entry-level workers in AI-exposed roles are the most vulnerable group. But the same studies show AI tools disproportionately help less-experienced workers — when they have access to them. The risk isn’t AI replacing juniors; it’s firms deciding not to hire juniors because AI can do the entry-level tasks. That’s a hiring-avoidance story, not a job-destruction story.
🔍 THE BOTTOM LINE
Stanford’s data doesn’t say AI won’t transform the labour market. It says we can’t see it in the aggregate data yet, and the signals we do see are confounded. The honest position is: the technology is powerful, the adoption is early, and the one group clearly affected — new graduates — faces a hiring market shaped by at least three forces beyond AI. Anyone claiming the apocalypse is here, or that it’s cancelled, is ahead of the evidence.
