Some More Thoughts on AI, Jobs, Wages, Skills, the Future, &etc.
Why it's important to go past the headlines and read the research cited in the article
My previous post on the headline topics drew some interest, so let’s keep the conversation going.
A good place to start is with a recent article by Christopher Mims in the WSJ. It’s a good place to start because of its headline: “AI Doesn’t Kill Jobs? Tell That to Freelancers.” One of the first things I learned when I started writing for newspapers and magazines is that I had little input into what the piece’s headline would be, and no veto power over what the editors came up with.1
The Gig Is Up?
Headlines are the original clickbait. They’re created to entice, not inform, and they’re often like the cover of a comic book in that they oversell the underlying content. So the WSJ article doesn’t actually document the death of freelancing as AI rises. It instead quotes a couple freelancers who have seen their work shrink in the era of Generative AI. It then cites three recent studies that look at what happened to postings on platforms like Fiverr and Upwork for different kinds of work after ChatGPT became widely available. In each case (and speaking a bit loosely), the authors separate the work into two categories: stuff that ChatGPT is really good at, like generating computer code, market content, and images, and stuff that ChatGPT and other LLMs aren’t as good at.
All three studies find what you’re probably expecting, especially if you read the article’s headline: a post-ChatGPT decline in postings for stuff that ChatGPT is good at. This seems like evidence that AI is in fact killing jobs. But spend a bit of time with these studies and two important things become clear.
First, in two of the three studies the employment declines are relative, not absolute . In other words the research doesn’t find that the total number of posts for AI-substitutable work (AISW) declines post-ChatGPT. It instead finds that compared to non-AISW work, the total number of posts for AISW declines. There’s no information on the total number of AISW posts over time. So as far as I can tell, the two studies are silent on whether the total number of AISW postings declined post-ChatGPT (And I can’t tell if the same is true for the third study; its full text was not available for download, and I couldn’t find a pdf of it.)
Second, the (relative) decline in AISW postings was not at all the only interesting thing observed by the researchers. Some of the other things they found show me that Jan Tinbergen really deserved to be the first Nobel-prize winning economist.23
Tinbergen pioneered the study of income distribution: why do some folk earn more than others? And why do these differences often get bigger over time? In answering these questions, Tinbergen highlighted two of the most powerful forces at work: education (skill acquisition, in other words) and technological progress.
When powerful new technologies come along they generally automate less-skilled labor: routine work like doing the same thing over and over on an assembly line, or adding up columns of numbers. The people who used to make a living doing this work either have to acquire new skills — via education, for example — or face declining prospects and wages. The new technology also has another effect: it tends to make advanced skills (again, often acquired via education) more valuable. As automobile factories get more mechanized and productive, for example, car designers can command higher pay, The more spreadsheets can do, the more a valuable a good financial analyst is. Etc., etc.
Skills Pay the Bills
The Tinbergian upshot of all this is that inequality is “a race between technology and education” as Larry Katz and Claudia Goldin (herself a Nobel prize winner) put it. Tech progress tends to increase inequality (by reducing willingness to pay for low-skilled labor) and education tends to decrease it (by raising skills, and hence raising the wage floor). There are important caveats here, but the broad pattern has held true for a long time. “Skill-biased technological change” is, as one study put it, “a canonical framework that does an excellent job of explaining US wage structure changes”
Do we see evidence of SBTC in online freelancing platforms’ recent “wage structure changes?” We absolutely do. One study found “a significant increase in demand for idea planning services” and that “demand for text reviewing services and services provided by human writers with a PhD degree remains unaffected by the launch of ChatGPT” (PhD-level skills, in short, remained in demand). Another study found that even though there was reduced (relative) demand for AISW, “the remaining automation-prone jobs are of greater complexity and offer higher pay” And Mims’ WSJ article contains fascinating data from Upwork:
Of the 12 task types in this chart, 8 of them have seen pay for their high-value versions — complex and non-routine versions of the tasks — go up more than pay for low-value versions has gone down. So far, GenAI seems to increasing demand for many types of skilled freelance labor.
Still Workin’ 9 to 5
About those freelancers. The WSJ article claims that they “represent an increasing proportion of the workforce” but this does NOT mean that folk with full-time jobs represent a decreasing proportion of said workforce. Matt Darling has been making this point for a while now: America is not becoming a nation of freelancers. Or if we are, it’s in the same sense that we’re becoming a nation of hoteliers because of Airbnb.
The main trend is that most of us continue to have job-y kinds of jobs. We’re near an all-time high of full-time workers as a % of the prime-age labor force:
And as Darling has pointed out, the percentage of people holding multiple jobs is significantly lower than it was in the 1990s:
Also, freelancers themselves don’t seem too worried: An Upwork survey conducted in late 2023 — well into the era of GenAI, in other words — found that “over 85% of freelancers say the best days are ahead for freelancing.”
All of the above reinforces two points for me. The first is, as I wrote in my last post here, “Both the creative and destructive sides of creative destruction are complicated phenomena — far more complicated than “powerful new technology means fast and big job losses.”” Or fast and big pay declines.
The second point is that LLMs are powerful but weird entities — ones that do by far their best work when they’re combined with a skillful human being. “For best results, combine machines and people” is one of the oldest mantras of the computer era. It dates back at least to tech guru Douglas Engelbart’s 1962 insights, and continues to be repeated by modern gurus like Jeremy Howard. Companies that get overenthusiastic about automation, like the car dealership that launched a customer-facing chatbot and wound up having to give someone a car for $1, keep relearning it. So, apparently, do headline writers at major business publications.
For books, your editor and publisher are usually willing to work with you to find a mutually agreeable title. However, they’re weirdly inflexible on the design of the cover. My guess is that this is because most authors have lousy ideas here. Whatever the reason, interactions between authors and publishers about the cover are bizarre. I get the strong impression that the designer and the publisher’s art department in general want absolutely nothing to do with the actual people who actually wrote the book. It got comical on one of my books; my editor set up a back channel to send me images of what the art department was thinking and get my feedback. Such back and forth was a deep norm violation on his part.
Custom demands a footnote here clarifying that the full name of the prize is “The Sveriges Riksbank Prize in Economic Sciences in Memory of Alfred Nobel.” I don’t make the rules.
Jan’s younger brother Nikolas also won a Nobel, in 1973 for Physiology and Medicine. Something tells me that you had to come correct to dinner in the Tinbergen household.