In other words: If cost of the tools weren’t a factor, and the only goal was to automate as much human labour as possible, how much work could technology take over?
When the Oxford researchers, Carl Benedikt Frey and Michael A. Osborne, were conducting their study, IBM Watson, a question-answering system powered by artificial intelligence, had just shocked the world by winning Jeopardy! Test versions of autonomous vehicles were circling roads for the first time. Now, a new wave of studies follows the rise of tools that use generative AI.
In March, Goldman Sachs estimated that the technology behind popular AI tools such as DALL-E and ChatGPT could automate the equivalent of 300 million full-time jobs. Researchers at Open AI, the maker of those tools, and the University of Pennsylvania found that 80 per cent of the US workforce could see an effect on at least 10 per cent of their tasks.
“There’s tremendous uncertainty,” said David Autor, a professor of economics at the Massachusetts Institute of Technology, who has been studying technological change and the labour market for more than 20 years. “And people want to provide those answers.”
But what exactly does it mean to say that, for instance, the equivalent of 300 million full-time jobs could be affected by AI?
It depends, Autor said: “Affected could mean made better, made worse, disappeared, doubled.”
One complicating factor is that technology tends to automate tasks, not entire occupations. In 2016, for instance, AI pioneer Geoffrey Hinton considered new “deep learning” technology capable of reading medical images. He concluded that “if you work as a radiologist, you are like the coyote that’s already over the edge of the cliff but hasn’t yet looked down.”
He gave it five years, maybe 10, before algorithms would “do better” than humans. What he probably overlooked was that reading the images is just one of many tasks (30 of them, according to the US government) that radiologists do. They also do things like “confer with medical professionals” and “provide counselling.” Today, some in the field worry about an impending shortage of radiologists. And Hinton has since become a vocal public critic of the same technology he helped create.
Frey and Osborne calculated their 47 per cent number in part by asking technology experts to rate how likely entire occupations like “telemarketer” or “accountant” were to be automated. But three years after their paper was published, a group of researchers at the ZEW Center for European Economic Research, based in Mannheim, Germany, published a similar study that assessed tasks — like “explain products or services” — and found that just 9 per cent of occupations across 21 countries could be automated.
“People like numbers,” said Melanie Arntz, the lead author of the ZEW paper. “People always think that the number must be somehow solid, you know, because it’s a number. But numbers can really be very misleading.”
In some scenarios, AI has essentially created a tool, not a full job replacement. You’re now a digger who can use an excavator instead of a shovel. Or a nurse practitioner with access to better information for diagnosing a patient. It’s possible that you should charge more per hour, because you’re going to get a lot more done.
In other scenarios, the technology is replacing your labour rather complementing it. Or turning your job from one that requires special skills to one that doesn’t. That is not likely to go well for you.
In either case, said Autor, technological developments throughout history have tended to mostly affect wages and wealth distribution — not how many jobs are available. “This kind of exercise risks missing the forest by focusing on one very prominent tree,” he said of studies that look at how much human work could be replaced by AI.
What he considers to be another key focus — how artificial intelligence will change the value of skills — is difficult to predict, because the answer depends partly on how new tools are designed, regulated and used.
Take customer service. Many companies have handed the task of answering phones to an automated decision tree, bringing in the human operator only to troubleshoot. But one Fortune 500 enterprise software company has approached the problem differently. It created a generative AI tool to provide the agents with suggestions for what to say — keeping humans, and their ability to read social cues, in the loop. When researchers at Stanford and MIT compared the performance of groups who were given the tool with those who weren’t, they found the tool significantly improved the performance of lower-skilled agents.
Even if a job becomes completely automated, how displaced workers fare will depend on how companies decide to use technology in new kinds of work, especially work we can’t yet imagine, said Daron Acemoglu, a professor at MIT and an author of Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity. These choices will include whether to automate work entirely or use technology to augment human expertise.
He said that the seemingly scary numbers predicting how many jobs AI could eliminate, even if it’s not clear how, were a “wake-up call.”
He believes that people could “steer in a better direction,” he said, but he is not optimistic. He does not think we are on a “pro-human” path.
All estimates for how much work AI could take over are very dependent on humans: the researchers making the assumptions about what technology can do. Frey and Osborne invited experts to a workshop to score the likelihood of occupations becoming automated. More recent studies rely on information such as a database tracking AI capabilities, created by the Electronic Frontier Foundation, a nonprofit digital rights group. Or they rely on workers using platforms like CrowdFlower, where people complete small tasks for money. The workers score tasks on factors that make them prone to automation. For instance, if it’s something with a high tolerance for error, it’s a better candidate for a technology like ChatGPT to automate.
The exact numbers are not the point, say many researchers involved in this type of analysis.
“I would describe our methodology as almost certainly precisely wrong, but hopefully directionally correct,” said Michael Chui, an AI expert at McKinsey who was an author of a 2017 white paper suggesting that about half of work, and 5 per cent of occupations, could be automated.
What the data describes is, in some ways, more mundane than often assumed: Big changes are coming, and it’s worth paying attention.
This article originally appeared in The New York Times.
Written by: Sarah Kessler
Photographs by: Glenn Harvey
©2023 THE NEW YORK TIMES