Has the A.I. Job Apocalypse Been Postponed?

How can these findings be explained? One possibility is that many professions may be more A.I.-resistant than first thought. In a new book, “Messy Jobs,” the economists Luis Garicano, of the London School of Economics, and Jin Li and Yanhui Wu, both of the University of Hong Kong, discuss jobs that are hard for A.I. models to displace because they involve a broader range of tasks than might appear at first glance. They cite radiologists as an example. When A.I. first came in, many observers thought they would be displaced en masse. That hasn’t happened, and that may be because, in addition to reading scans, radiologists also do things like talk to patients, coördinate with surgeons, and reach judgments on tricky cases—things that A.I. models may struggle with. Another factor to consider is that, for now, anyway, many professions that involve manual labor seem happily oblivious to A.I. Data from Anthropic shows that cleaners, carpenters, electricians, and plumbers account for virtually none of the usage of Claude, its A.I. system.

From an employment perspective, that’s good news. But A.I. and robotics are progressing by the month. Perhaps the hammer is still hanging over America’s knowledge workers; it just hasn’t fallen yet. Take the legal profession. In theory, generative-A.I. models can carry out some of the tasks that lawyers have traditionally undertaken, such as drawing up and checking documents, researching past cases, and even making legal arguments. (One recent blind test found that law professors preferred A.I.-generated answers to legal questions over those written by their fellow-professors.) And yet, according to the Bureau of Labor Statistics, the number of lawyers is stable, or even rising slightly. Last month, there were 1,245,9srcsrc people providing legal services in the U.S. That figure is up about half a per cent from a year ago, and it’s at roughly the same level as it was a decade ago.

But how long can this situation persist? Ultimately, client fees pay the salaries of all these lawyers, and, thanks to A.I., the fee structure is being challenged on at least two fronts. The legal world is awash with A.I. startups promising to transform the industry’s economics. And now Wall Street banks, which spend enormous sums on legal services, are demanding fee reductions from their lawyers. “If the number of hours they’re working on a matter has come down because of A.I. . . . our expectation is for costs to come down significantly per transaction,” Adam Meshel, Citigroup’s global head of legal, told the Financial Times. It seems likely that some law firms, to preserve their profit margins, will further automate their operations and reduce their head count, particularly at junior levels.

Law is merely one area where competition and the profit motive could make A.I. bite workers. Given the rapid development of cheaper open-weight models from China and elsewhere, it’s easy to imagine similar scenarios playing out in many other industries where cognitive tasks are central, such as consulting, finance, and publishing, not to mention coding, accounting, and data processing. Ever since the rollout of ChatGPT, there have been two scenarios for the impact of A.I. on workers: the technology could complement people’s skills and boost their productivity and wages, or it could render lots of skills obsolete and end up substituting for many workers entirely. Recently, concerns about the negative possibilities of A.I. prompted more than two hundred economists and researchers to sign a public letter calling on policymakers and tech leaders to “build the incentives, guardrails, and institutions needed to steer AI in a direction that complements humans and benefits society.”

One upside of many people in Silicon Valley overstating the speed of the A.I. revolution is that we may have a bit more time to introduce reforms that elevate the interests of workers, encourage firms to deploy A.I. in a positive manner, and make sure that the profits A.I. generates are widely shared. (On the existential question of safety, OpenAI’s Hugging Face saga suggests that time may be in considerably shorter supply.) Although there is no magic solution to the economic challenges that A.I. pose, economists have suggested a number of policies, including changing the corporate tax laws to favor investments in people over machines; exploiting the government’s procurement powers in areas where it’s a big spender on technology, such as health care and education; establishing federal jobs subsidies; and even a federal jobs guarantee. (In the political realm, of course, there is a growing movement to block or delay the development of data centers.)