The Jobs AI Will Change, Not Replace – by Robert VerBruggen – City Journal Substack

Everyone is worried about AI taking over jobs. Each week I’m deluged with papers trying to measure how “exposed” to AI competition each job is—and whether workers in these fields are suffering for it.

A new working paper from the Burning Glass Institute’s Soumitra Shukla flips this concept on its head—or at least on its side. Instead of focusing on which jobs might be displaced by artificial intelligence entirely, it asks which jobs contain an awkward mix of tasks—some of which can, but many of which cannot, be massively sped up with AI.

These are jobs that could change profoundly rather than disappear. In these jobs, workers’ efforts will likely narrow to focus on the tasks AI can’t do. Such tasks will become a “bottleneck” to productivity, and improving human performance in those areas will become especially valuable.

Using data from job postings on the tasks that different types of work involve, Shukla constructs an index of this “bottleneck pressure,” which reaches its maximum of 25 when a job is a perfectly even mix of AI and non-AI tasks, or 50 percent AI exposure.1

He also provides a handy list of how different jobs fare. For example, accountants face extremely high AI exposure, 90 percent, and thus have a low bottleneck score (8.7): AI is a big deal in this field, but there are few tasks that can’t be done that way to serve as bottlenecks. Fast-food cooks, with very low exposure (about 15 percent), similarly rate just a 13: humans aren’t bottlenecks when AI can’t do much work either way.

Occupations facing approximately the maximum pressure, thanks to roughly 50 percent AI exposure, include orthodontists and various types of technicians. This is where AI can do a lot, but only if humans keep pace with the other tasks.

Shukla also looks at the issue through the lens of the individual tasks rather than overall jobs. He highlights that the same tasks might serve as bottlenecks to different degrees in different fields.

Imagine that two different jobs involve the same task—a task that can’t be done with AI. However, one job is staying the same over time, while the other involves a lot of other tasks newly being done by AI. The task will be more of a bottleneck in the second job, and employers in that field might make special efforts to bring in employees good at the task.

For example, both fast-food workers and customer-service representatives need to greet customers and make them feel welcome. But the latter job’s other tasks are more exposed to AI, which could, over time, make human interaction a more important part of what’s left.

The paper is necessarily on the speculative side. You can empirically measure the tasks mentioned in job listings, but you can’t say for sure how AI will play out in each field and how jobs will evolve to accommodate it. Not all of the method’s job ratings made intuitive sense to me, either. Can AI really affect about 40 percent of the tasks required of a retail salesperson, which puts that occupation toward the high end of the bottleneck index?

Nonetheless, Shukla starts an important conversation here. It’s critical to understand what tasks AI can do and the jobs it might destroy—but maybe we should pay attention to the tasks it can’t do, and the jobs it can change, too.

From Manhattan Institute

• From yours truly, a look at the future of antidiscrimination law’s “disparate impact” doctrine, which applies, for example, to hiring tests with uneven results across demographic groups. I focus on two threats to the doctrine in particular: the Trump administration’s efforts to address its constitutional issues, and the difficulty of applying its principles to new AI- and algorithm-based hiring technology.

Other Work of Note

• California’s billionaire tax, on the ballot this November, is likely to collect less money than previously expected, and has already driven some to leave the state, according to this study. MI’s Shawn Regan discussed a previous version of this research here.

• A different way of looking at the tradeoffs of wealth taxes.

• “[I]mprisoning 431,632 felony contact sex offenders for an average of 4.3 years each between the years 2000 and 2019 may have prevented between 120,744 and 551,708 rapes in the United States.”

• Are recessions good for health and mortality?

• How will medical costs affect the federal budget going forward? And how much do Medicaid dental benefits cost?

• More Medicaid, more credit-card borrowing.

• Some complications of forcing states to take on more of the costs of the food-stamp program if they have high payment error rates.

• Can you nag someone into finishing high school via text message?

• Across many different countries, boys are more likely to have a bad attitude toward academic effort.

• “Active” learning seems more effective than lectures.

• In Chicago’s school-choice system, students’ demand for top-rated schools depends on how far they’d have to travel to reach them.

• How culture affects fertility.

• New data from Pew on how Americans are using social media, cell phones, and broadband internet.

• Do people vote primarily based on policy—not demographics, or even partisanship—after all?

• Some helpful points about the tax code should treat business investments.

• How big of a deal are fossil fuel subsidies?

• Some agencies and organizations intentionally add noise to the data they release, to protect privacy. How much does that affect analyses of the data?

• A new federal report on “debanking” incidents.

• “Fee waivers” granted by Fannie Mae and Freddie Mac are mostly captured by mortgage intermediaries, not passed through to borrowers.

• How to improve the National Science Foundation.

• What (and how much) are kids vaping these days?