This piece first appeared in The Algorithm, our weekly newsletter on AI. To receive stories like this in your inbox before anyone else, enroll here.
In the realm of Silicon Valley, a job apocalypse spurred by AI is regarded as inevitable. The atmosphere is so bleak that a societal impacts researcher at Anthropic, in response to a request for more hopeful outlooks regarding AI’s future, mentioned a possible recession in the near future and a “collapse of the early-career ladder.” Her more outspoken peer Dario Amodei, the company’s CEO, has described AI as “a general labor substitute for humans” capable of performing all jobs within five years. And those sentiments are not exclusive to Anthropic, obviously.
Consequently, these discussions have understandably left many employees feeling anxious (and are likely fueling support for initiatives to entirely halt the development of data centers, some of which gained momentum last week). This anxiety is not alleviated by politicians, none of whom have presented a clear strategy for what follows.
Even economists who have warned that AI has not yet resulted in job losses and may not lead to a crash are starting to acknowledge that it could uniquely and unprecedentedly affect our work patterns.
Alex Imas, from the University of Chicago, is one such economist. He conveyed two key points during our conversation on Friday morning: a frank evaluation that our predictive tools are quite deficient, and a “call to action” for economists to begin gathering the one type of data that could facilitate a plan for addressing AI in the labor market.
Regarding our inadequate tools: consider that any occupation consists of various individual tasks. For instance, a part of a real estate agent’s role is to inquire about what type of property clients wish to purchase. The US government has documented thousands of these tasks in a vast catalogue initiated in 1998 and continually updated. This was the data that OpenAI researchers utilized in December to assess how “vulnerable” a job is to AI (for example, they determined that a real estate agent is 28% vulnerable). Then in February, Anthropic leveraged this data in analyzing millions of Claude conversations to identify the tasks people are actively employing its AI for and the points of overlap.
However, understanding AI vulnerability in tasks leads to a deceptive comprehension of how at risk a specific job may be, according to Imas. “Vulnerability alone is a completely ineffective metric for forecasting displacement,” he asserted.
Indeed, it is revealing in the most dire scenario—for an occupation where virtually every task could be performed by AI without any human supervision. If an AI model can accomplish all those tasks for less than what you’re being compensated—which is not assured, as reasoning models and agentic AI can accumulate significant costs—and perform them effectively, that job is likely to vanish, asserts Imas. This reflects the frequently cited case of the elevator operator from years past; perhaps today’s equivalent is a customer service agent solely tasked with handling phone call triage.
Yet, for the majority of occupations, the situation is not as straightforward. The specifics are crucial as well: while some jobs may face ominous prospects, discerning how and when these changes will occur is challenging when solely considering vulnerability.
Take coding as an example. A person creating high-quality dating applications, for instance, may utilize AI coding tools to produce in one day what previously required three days. This boosts productivity for the worker. The employer, maintaining the same financial outlay, can now yield more output. Consequently, will the employer seek to hire more employees or fewer?
This is the question that Imas argues should keep any policymaker awake at night, as the answer will vary across industries. And we are navigating in uncertainty.
In the coder’s instance, such efficiencies allow dating applications to reduce prices. (A skeptic might anticipate companies merely pocketing the profits, yet in a competitive marketplace, they risk being undercut if they do.) These lower prices will invariably stimulate some surge in demand for the applications. But to what extent? If millions more individuals desire it, the company might expand and ultimately recruit more developers to fulfill this demand. However, if demand hardly increases—perhaps those who don’t currently use premium dating apps will still not want them even at a lower price—fewer coders will be necessary, resulting in layoffs.
Regularly applying this hypothetical across every job with tasks that AI can perform yields the most critical economic inquiry of our era: the particulars of price elasticity, or the extent to which demand for something shifts when its price shifts. This is the second aspect of what Imas highlighted last week: we currently lack this data throughout the economy. However, we could.
We possess figures for grocery products like cereal and milk, Imas notes, because the University of Chicago collaborates with supermarkets to gather data from their price scanners. Yet we lack such figures for tutors, web developers, or dietitians (all professions identified as having “exposure” to AI, by the way). Or at least not in a manner that’s been extensively compiled or made accessible to researchers; sometimes it’s distributed among private companies or consultancies.
“We require, akin to a Manhattan Project, to gather this information,” Imas states. Moreover, it’s not solely necessary for occupations that could evidently be impacted by AI now: “Fields not yet exposed will inevitably become exposed in the future, warranting a tracking of these statistics across the entire economy.”
Acquiring all this data would demand time and resources, but Imas argues that it’s valuable; it would provide economists with the first genuine insight into how our AI-driven future could evolve and offer policymakers an opportunity to devise a strategy for it.
