AI Is Not Eating Your Job. At Least Not Yet.

Everyone, it seems, is convinced that white-collar jobs are being hollowed out by AI. Tech layoffs at the likes of Coinbase, Meta and Cisco keep getting cited as portents of a broader collapse. The narrative has momentum. It also, currently, lacks much supporting evidence.

The actual labour market data tells a more boring story: AI has not yet caused any measurable large-scale disruption to US employment. Bureau of Labor Statistics figures show that unemployment in occupations most exposed to AI is actually lower than in occupations with minimal AI exposure. There is also no sign of workers flooding out of at-risk roles and into manual trades, which is exactly what you would expect to see if displacement were happening at scale.

Erika McEntarfer, who ran the BLS until the Trump administration fired her last autumn after an inconvenient jobs report, puts it plainly: the impact AI is having on current labour market conditions is likely small. She is now at the Stanford Institute for Economic Policy Research and points to Census data showing only one in five companies are using AI in any business function whatsoever. Her view is that disruption of labour markets cannot happen at scale until AI first transforms how businesses actually operate, and that process takes time. History bears this out. Every major technological shift has worked its way through economies slowly and unevenly.

None of which is to say the job market is fine. It is not, particularly for younger workers. Unemployment among recent graduates is sitting around 5.6 percent, a level not seen since the immediate aftermath of the 2008 recession. Hiring rates have been poor throughout the post-pandemic period. If you are 23 and looking for a tech job right now, the market feels brutal.

The question is how much of that pain is AI's fault versus other macroeconomic factors. Researchers at the Stanford Digital Economy Lab dug into payroll data from ADP, which covers a far larger sample than BLS surveys, and found something striking: head counts for 22-to-25-year-olds in high-exposure occupations like software development and customer service started dropping around late 2022, roughly when ChatGPT launched. By 2025, entry-level jobs in the most AI-exposed roles had fallen around 16 percent. Meanwhile, employment for older workers in the same occupations grew, as did jobs in less exposed fields.

There is a plausible explanation for this pattern. Entry-level work tends to rely on codified knowledge, the kind learned in classrooms and easily replicated by a capable language model. Junior developers write code that AI can now produce competently. Experienced workers carry tacit knowledge built through years of context, judgement and client interaction, which is considerably harder to automate. The earn-while-you-learn model, where graduates did the grunt work and slowly accumulated experience, may be breaking down in certain fields.

Interestingly, wages in high-exposure sectors have actually risen since ChatGPT arrived. If AI were simply replacing workers, you would expect the opposite. Instead, it suggests employers are still paying a premium for the experience and judgement that AI cannot yet replicate. The problem is not that experienced developers are being pushed out. It is that fewer people are being hired to become experienced in the first place.

Coding is the clearest case study. A Federal Reserve paper found annual employment growth for programmers has slowed by around 3 percent since 2022. But overall employment in coding is still growing, just less quickly than before. The jobs are not disappearing. The occupation is being reshaped.

This should sound familiar. In 2013, Geoffrey Hinton declared that training new radiologists was pointless because AI would obviously replace them. Obama's economic advisors warned that autonomous vehicles would eliminate up to 3.1 million jobs. There are now more radiologists than ever. Driverless trucks remain conspicuously absent from motorways. AI became a useful tool for screening images, but radiologists turned out to do a lot of things beyond staring at scans.

The 'this time is different' argument is always available, and it might eventually prove correct. The capabilities of current AI systems are genuinely without precedent. But the same argument has been trotted out repeatedly over the past century and has repeatedly been wrong about the scale and pace of displacement.

What most economists are pointing to instead is a transition problem rather than an elimination problem. Jobs will be redefined. Some will pay worse. Some workers will struggle to adapt. Sectors could shift faster than training and reskilling programmes can keep pace. The comparison being drawn is to the so-called China shock of the early 2000s, when trade liberalisation gutted manufacturing communities and it took years for researchers to even understand what had happened, let alone for policy to respond. An AI transition done badly could be considerably larger in scope.

Erik Brynjolfsson at Stanford is unusually bullish on AI's economic potential and thinks we may be approaching some of the strongest productivity growth in living memory. He also notes, with some frustration, that while hundreds of billions are being spent rolling out AI, a fraction of a percent of that sum is going toward understanding what it is actually doing to the economy.

That is the core problem. The data we have is patchy and slow. Harvard economist David Deming, who has been surveying thousands of workers quarterly on AI usage since 2024, describes the current situation as flying blind. The BLS household survey gives a broad picture. A handful of academic projects are trying to get more granular. But nobody has a clear, real-time view of how AI adoption is reshaping employment at the level of specific tasks, firms and demographics.

McEntarfer's read is that speed is the critical variable. If disruption unfolds at the typical pace of technological change, labour markets will adapt. If it hits suddenly, policymakers will be caught flat-footed with no adequate response ready. Given that we were not adequately prepared for trade shocks we saw coming years in advance, that is not a particularly comforting baseline.

The apocalypse is not here yet. But that is not the same as saying there is nothing to worry about.