NRPLUS MEMBER ARTICLE E arly in the 2010s, academics and entrepreneurs began to raise concerns about the economic consequences of artificial intelligence. Technological advances, the thinking went, would soon render vast swathes of the labor force obsolete, deepening income inequality and destabilizing society. Proponents of the “automation revolution” thesis called on policymakers to cushion workers from the effects of technological displacement through fiscal transfers and increased job training for technical fields.
MIT economist Erik Brynjolfsson, a pioneer in the economics of AI, said in 2014 that job loss due to automation would be “the biggest challenge of our society for the next decade.” Six years into the decade in question, it is time to take stock of his prediction. Is it true, as former Secretary of the Treasury Lawrence Summers said, that rapid automation “isn’t some hypothetical future possibility” but “something that’s emerging before us right now”?
Not quite, say economists Keller Scholl and Robin Hanson. In a paper published last month, they found that over the past 20 years, both the level and growth rate of job automation have been more or less flat. According to their analysis of 1,505 expert reports published by the Occupational Information Network (O*NET), while many workers are losing their jobs to machines, they are doing so at roughly the same rate as in the past.
Among the 261 occupational characteristics reported by O*NET — such as the degrees to which jobs require creativity, physical strength, or numeracy — two stand out in predicting automation: the importance of machinery and the importance of routine tasks. Unsurprisingly, assembly-line workers and data-entry clerks are particularly vulnerable to automation.
But factory work has seen a trend of automation going back several decades. Those sounding the alarms on AI have warned that not only factory workers but also skilled “knowledge” workers would face competition from machines. Indeed, algorithms are said to be capable of customer service, medical diagnostics, and news writing, among numerous other tasks. Yet the analysis of Scholl and Hanson indicates that workers are far more likely to be displaced by relatively dated technologies: manufacturing machinery, word processors, and spreadsheets. In other words, the types of jobs being automated haven’t changed much, despite technological advances.
The study also considers the vulnerability of jobs to automation by computers and machine-learning algorithms in light of two metrics devised by academics, called “computerizability” and “machine-learning suitability.” While the potential of digital technologies and AI to replace a given occupation appears to be a strong predictor of automation, its significance disappears when other factors, such as routineness, are taken into consideration. Which is to say that the “threat” posed by artificial intelligence is more or less the same as that posed by older technologies.
“The fundamental nature of automation hasn’t changed over the past 20 years,” Hanson tells National Review. “There’s this AI media story that’s been played over and over again for the last decade, and people are so familiar with it that they don’t bother to research it.”
These finding calls into question the need for policies to address automation, such as former Democratic presidential candidate Andrew Yang’s flagship universal-basic-income (UBI) proposal. Endorsed by a growing number of politicians and technologists, UBI is premised on the belief that automation will eliminate millions of jobs. The evidence doesn’t suggest the need for such large-scale structural changes to the U.S. economy, but the threat of automation serves as an easy talking point for politicians. “People pitch what they want to pitch, and frame it in terms of automation,” Hanson argues. “You can be pretty confident that those recommending a certain policy response to AI won’t change their minds” in light of his study’s findings.
Because it relies on subjective reporting, the paper does not definitively disprove the automation-revolution hypothesis. In general, it is hard to get an objective picture of the magnitude of automation, and it is possible that labor experts have underestimated the rate at which it is happening. Scholl and Hanson do find that the average job is significantly more susceptible to automation today than 20 years ago, even as the level of automation remains somewhat constant. And the paper’s central finding — that jobs dependent on technology are more likely to be automated — raises the possibility of a feedback loop in which automation begets automation, potentially spurring exponential growth in the number of jobs replaced by machines.
But that possibility remains remote, and the burden of proof lies with those arguing that technology is fundamentally transforming American life, and that we must fundamentally transform public policy in response. Until they can marshal convincing evidence, we should be skeptical of proposals that would remake the economy to fit their vision.