Artificial Intelligence has not Gutted White-collar Work — not yet, anyway

The data show a labour market bruised by many things. AI, so far, is a bit player

BAD NEWS travels fast, and few stories travel faster these days than the one about robots coming for the cubicle. Layoffs at Coinbase, Meta and Cisco are cited as proof that a white-collar reckoning has begun. Yet a look at the actual figures tells a duller, more reassuring story: America's aggregate labour market shows scant sign of an AI-driven collapse. Unemployment among occupations most exposed to artificial intelligence is, if anything, lower than among those least exposed. Nor is there any sign of a mass migration from AI-threatened desk jobs to supposedly safer manual ones — the kind of shift one would expect if the robots really were coming.

None of this means AI is harmless to workers. It means the disruption, if it is coming, has not yet shown up in the statistics that economists watch most closely.

Slow-moving technology, fast-moving fears

Erika McEntarfer, who ran the Bureau of Labor Statistics until she was dismissed last autumn, argues that the muted impact should not surprise anyone who has studied earlier technological revolutions. New tools take years to reshape whole industries; AI cannot transform jobs until it first transforms the businesses that employ people. Census Bureau data back this up: only around one in five American firms currently use AI in any part of their operations. That is hardly the profile of a technology already devouring payrolls.

Adoption vs. impact, at a glance

Indicator	Finding
Firms using AI in any business function	~20%
Workers using generative AI regularly	~40%+ (varies by sector)
Unemployment, recent college graduates	~5.6%
Decline in entry-level jobs, most AI-exposed occupations (since 2024)	~16%
Slowdown in annual coder employment growth since ChatGPT	~3 percentage points

 

Where the pain is real: the young

The exception to this sanguine picture is entry-level workers, particularly those aged 22 to 25 in occupations such as software development and customer service. Using payroll data from ADP, researchers at Stanford's Digital Economy Lab sorted 950 occupations by exposure to AI and tracked employment by age group. The pattern, in the words of the lab's director, Erik Brynjolfsson, was striking: head counts for young workers in the most exposed jobs began falling, with the drop accelerating sharply after 2024. Older workers in the very same occupations kept getting hired.

The mechanism appears to matter as much as the exposure. Jobs in which AI merely assists human workers grew faster than average. Jobs in which AI could substitute for a worker with little oversight are the ones that shrank. One reading of this: entry-level roles lean on "codified" knowledge — the kind absorbed in a classroom, and precisely the kind large language models are good at mimicking. Seasoned workers carry "tacit" knowledge built from experience, which machines still struggle to replicate.

Coding: transformed, not terminated

Software developers have become the poster children for AI anxiety, and the data give that story partial support. Economists at the Federal Reserve Board found that annual employment growth among coders has slowed by roughly three percentage points since ChatGPT's debut. But growth has slowed, not reversed — coding employment is still rising overall, just less briskly. Wages in heavily AI-exposed sectors have also climbed relatively quickly, suggesting employers still prize the experience AI cannot yet supply. The old model — hire a graduate, have AI (or a senior colleague) train them slowly into that expertise — looks shakier than the jobs themselves.

History's rerun, or something new?

Warnings of machines eating jobs are a recurring genre. A 2016 White House report fretted that driverless trucks could erase millions of positions; none of that has happened. Geoffrey Hinton, a pioneer of the technology now being deployed in radiology, once suggested hospitals should simply stop training radiologists. Radiologist numbers have grown since, because the job includes plenty that AI cannot do — consulting with patients, interpreting ambiguous scans, making judgment calls.

Economists caution against reading too much comfort into that history, though. As David Deming of Harvard puts it, researchers are largely "flying blind": the government's monthly household survey and even novel real-time projects, such as Deming's own quarterly survey of AI use since 2024, capture usage and productivity, not the eventual fate of specific jobs. Jed Kolko of the Peterson Institute notes that an economy without mass unemployment could still deliver a rough transition — jobs redefined, pay compressed, some workers simply unable to adapt.

The bottom line

For now, the safest reading of the evidence is this: AI is reshaping certain corners of the labour market, especially entry-level technical work, while leaving the broader picture largely intact. Whether that stays true depends on a variable nobody can yet measure — the speed of change. If disruption arrives at the "normal pace" of past technological shifts, argues Ms McEntarfer, labour markets and policymakers will have time to adjust. If it arrives suddenly, they will not. Getting better, more granular data now, economists agree, is the best insurance against being caught flat-footed the way America was by the "China shock" two decades ago — a disruption whose scale only became clear years after the damage was done.