One version of the AI jobs narrative is already essentially self-written. When a new report arrives, researchers calculate the number of jobs at risk, sometimes in the millions, by counting the tasks that large language models are capable of performing and mapping those tasks onto current job titles. The narrative is published. People nervously share it. After a few months, the unemployment rate returns to its previous level, and the cycle repeats itself.
Recently, columnist John Burn-Murdoch of The Financial Times identified the problem with this analysis. The issue is not that scientists are misrepresenting the capabilities of AI. The issue is that in an argument that actually calls for much more, “can AI perform this task?” is doing nearly all the heavy lifting. It’s similar to determining whether a new highway will destroy a neighborhood by measuring the width of the road; while technically relevant, this method is incredibly inadequate.
The difference is more important than it may seem. To use Burn-Murdoch’s own example, spreadsheets could actually do the majority of the tasks that bookkeepers performed. Jobs in bookkeeping did decline. However, because the spreadsheet took care of the mechanical work and freed up human judgment for everything else, financial analyst jobs grew, expanded in scope, and became more lucrative. The majority of the fascinating economic history actually exists in the space between capability and displacement.
Although coming from a different angle, Sam Altman appears to have reached a similar conclusion. For years, he talked confidently about how AI would replace whole categories of white-collar jobs. He proposed that entry-level positions would be most affected. Then, in May, he declared at a Commonwealth Bank of Australia conference that he was “delighted to be wrong”—that the human aspect of work, where coworkers genuinely want to communicate with one another rather than a chatbot, had proven to be more resilient than he had anticipated. It’s difficult to say with certainty whether this is a true intellectual update or a calculated softening ahead of OpenAI’s anticipated IPO. Most likely a combination of the two.
The jobpocalypse narrative actively misleads people and serves the interests of AI companies trying to justify premium pricing, according to Andrew Ng, whose credibility in this field is hard to challenge. His reasoning is straightforward: charging $10,000 for your product suddenly makes sense if it can replace a $100,000 employee. It turns out that, even when the scarcity isn’t quite real yet, fear and scarcity are great marketing tools.
However, dismissing the anxiety would be too easy. It is truly comforting that the Yale Budget Lab found no significant change in unemployment among workers in high-AI-exposure jobs through March 2026. However, according to a different Stanford analysis cited by researcher Paul Fergus, entry-level hiring in AI-exposed roles has already decreased by about 13%, with employment in the most exposed positions falling by nearly 20% from its peak in 2022 for workers between the ages of 22 and 25. The aggregate appears good. Booking Holdings CEO Glenn Fogel bluntly stated at the TIME100 Summit that the bottom rungs of the ladder are being pulled away.
In the binary argument between proponents and opponents of the jobpocalypse, that particular detail is overlooked. In the past, technological disruption has been absorbed by the economy through reallocation, where workers who were displaced in one sector eventually found employment in another. However, reallocation makes assumptions about mobility, retraining, and geographic flexibility that aren’t always true in real life. Coal mines were shut down in Thatcher’s Britain; economists noted the increase in national productivity, but the communities that were left empty as a result of those closures remain empty.
The FT’s framing accurately conveys that the true question is not whether AI is capable of doing anything. It concerns whether companies will reorganize around that capability, when they will do so, and whether the impacted employees will be able to find other employment. No task-mapping exercise can address those slower, messier, and more human questions. Additionally, the majority of the concerning reports subtly ignore these questions. Altman’s updated optimism might turn out to be accurate in the coming years. It’s also possible that they won’t. At the very least, it seems reasonable to start with the appropriate question.
