Brian Doherty, Oxford, 2019
- In 2013, Oxford economist Carl Benedikt Frey estimated that nearly half of all US jobs could be replaced by AI.
- New technologies like ChatGPT have taken the business world by storm in recent months.
- We spoke with Frey about whether ChatGPT is coming for Americans’ jobs.
In a paper he co-authored titled “The Future of Employment,” Frey estimated that 47% of all US jobs were at risk of being replaced by automation as soon as the “next decade or two.” The paper has been cited over 5,000 times and covered by countless media outlets, including a 2019 episode of Last Week Tonight with John Oliver.
AI hasn’t replaced all of us just yet, and the paper isn’t without its critics, but Frey, who’s served as a consultant for the United Nations, the OECD, and several Fortune 500 companies, maintains that AI will disrupt plenty of industries in the future.
Now a decade since the paper’s publication, Insider spoke with Frey about what he got right — and wrong, whether ChatGPT is coming for all of our jobs, and why even if it doesn’t replace you, it might help someone else do so. It’s not all bad, however. Frey says the AI takeover could be a good thing for all of us in the long run.
Artificial intelligence might not take your job, but it will have impacts
In 2013, Frey concluded that “low-skill, low-income jobs” were most at risk of being replaced by automation, and he says he still believes this to be the case.
“When autonomous vehicles arrive, they’re not going to augment a taxi driver, they’re going to replace those drivers,” he told Insider. He also listed truck drivers, telemarketers, receptionists, cashiers, waiters, and fashion models as jobs particularly at risk of replacement.
When it comes to ChatGPT and related AI, however, Frey says job replacement is “probably not the right way of thinking about it,” but workers still have some reason to fear for their job security. The reason: “more competition.”
Frey points to “creative” industries like writing, music, and art, as well as graphic design, advertising, and computer programming roles, where he says widespread access to tools like ChatGPT will allow far more people to produce high quality work.
So rather than having your job “outright replaced,” Frey says, ChatGPT’s adoption may ultimately result in a different outcome — being replaced by another human. And even if workers manage to hang on to their jobs, he says the influx of new competition for their roles could result in lower wages.
He compares it to the introduction of ridesharing companies like Uber, which resulted in more drivers, increased competition, and ultimately led to incumbent drivers experiencing “wage cuts by around 10% or so,” per his research.
“Uber didn’t reduce the demand for taxi drivers,” he said. “It, if anything, increased the number of people driving cars for a living, but it reduced the amount the earnings capacity of incumbent drivers.”
He says a similar development could play out in industries where ChatGPT is most effectively deployed.
“In my view, it’s less about automation,” he said. “It’s more about democratization and competition, potentially leading to lower wages for people in some of these professions.”
That said, Frey doesn’t think AI will only replace blue-collar roles, citing fashion models and credit analysts as two jobs that could potentially be at risk.
AI has been slower to replace drivers than expected, but has also shown it has potential to do more than Frey thought
Looking back at the 2013 study, Frey acknowledges there are some things he’d have done differently if he could go back in time. That said, he believes the study’s estimates regarding potential AI job replacement are still “broadly on target.”
Frey points to one category of job in particular that’s been automated at a much slower pace than he expected a decade ago: drivers.
“I think most people we spoke to back then said it’s 10 to 15 years from now,” Frey said regarding the widespread adoption of autonomous vehicles. “And 10 years later, most people are still saying that it’s 10 to 15 years from now.”
On the flip side, he was skeptical whether waiters and fashion models, for instance, were jobs that could potentially be automated, but says new technologies have since convinced him this is possible.
In the distant future, Frey says it’s likely “everybody will be replaced” eventually by AI. As for his 47% prediction, the paper said this could happen in “perhaps a decade or two,” something many people latched on to.
But Frey says he and fellow researcher Michael A. Osborne did not intend to offer a precise forecast on the “pace of automation,” which he says there is not “any way to credibly forecast.” Instead, he says they were focused more on forecasting the scope of jobs that could be at high risk of replacement — whenever widespread AI adoption does materialize.
AI could make the jobs of the future better
While emerging technologies have impacted many jobs in decades past, it hasn’t always led to replacement.
Secretaries and bank tellers, for instance, are doing “completely different jobs” than they were a few decades ago, Frey says, but they haven’t been formally replaced — their roles have simply evolved.
“I think there’s this somewhat misguided distinction between replacement and jobs being changed and transformed by technology,” he said.
In general, Frey says technological progress like ChatGPT is good for society. “It’s the reason that we’re a lot more prosperous today than a couple of hundred years ago.”
And for many workers, AI tools may eventually be able to make their jobs easier.
“Some of the more boring parts of the job may disappear,” he said. “We may be more focusing on generating the right ideas, asking the right questions, things that are more interesting.”
At least in the near-term, Frey says humans should maintain an edge over AI in many areas. That’s because there are arguably some things ChatGPT can’t learn from scouring data from millions of websites.
“Humans don’t just interact on the internet,” he said. “We’re out interacting in the real world and draw on experiences there as well. And often, we come up with things that are from a completely different domain that are unlikely to be in the training dataset.”