AI, Workers, and the 2 Myths | Roundup #14
Discerning the different varieties of AI that help and harm workers
Thanks for your patience on this one — spent a little more time on research in between travel stints! But wanted to get one last post out before taking a bit of a break (I’ll return in mid-January).
We have some great things planned for the year ahead, including an interview series with people I think you’ll be really excited to hear from…
Thanks for reading! As always, I’d love to hear your thoughts here or on Twitter.
ChatGPT and Generative AI
At the end of November, artificial intelligence (AI) company OpenAI released ChatGPT, a conversation-based AI. Within days, enthusiastic users were posting to Twitter how ChatGPT was able to do everything from writing their essay for a history class to describing programming concepts like a 1940s gangster would:
I even used ChatGPT to help me get started with this post!
If you haven’t already, I’d encourage you check out ChatGPT yourself here (it’s free).
ChatGPT is just the latest example of generative AI — which allows people to generate content with a simple text prompt. And generative AI is not just limited to generating text. You’ve probably seen images from Midjourney, which can generate images from a simple description like “Bob Ross as a muppet” (because why not). Runway is working on a product that can generate entire videos from just a line of text. Dozens of other companies are creating new products around this emerging field of generative AI.
And the field is accelerating quickly, in recent years surpassing human capabilities in speech, reading, and image recognition:
Things are evolving extremely quickly. Many who try out GPT-3, for example, see it potentially replacing tasks we currently perform at work, like writing outreach emails, drafting blog posts, or even writing code to fix a software bug.
But is generative AI, and automation in general, good or bad for workers?
The Two Myths
When people talk about AI and workers, you tend to hear loud voices from one of two groups, each with its own strongly-held belief about technology and its influence on society. Either:
AI will usher in expansive productivity growth, with workers taking more interesting and safer jobs, and consumers benefiting from lower-priced goods and services.
AI is going to cause mass unemployment, with wealth primarily accruing to corporations that create and deploy AI.
We are either heading towards a techno-utopia or techno-dystopia, depending on who you ask. I’d argue that both situations are their own myths. We know from economic research that some variants of AI, and technology more broadly, do in fact help workers, increasing wages and employment. Meanwhile, there are ways that AI can be deployed that would result in stagnant wages and worker displacement.
Good Automation vs Bad Automation
Economists Daron Acemoglu (MIT) and Pascual Restrepo (Boston University) point out that when technology is used to automate work tasks that otherwise require human labor, there are two countervailing effects on labor demand. Depending on which effect predominates, the automation may cause higher labor demand (e.g. higher wages, employment) or lower labor demand (e.g. lower wages, displacement). These two effects are:
Displacement Effect: by having technology perform tasks that used to be done by human labor, it takes less labor to get to a certain amount of output; displacement naturally leads to lower labor demand
Price-Productivity Effect: automations reduce the cost of production for a firm, which in turn causes that firm to expand, and thereby increases that firm’s demand for labor to perform the remaining non-automated tasks
Not all automations are created equal. It seems that for automations where price-productivity offsets displacement, wages and employment can remain steady or even increase. Acemoglu and Restrepo point out that, when comparing countries that invested in robotic automation, Germany, Japan, and South Korea probably had better labor outcomes than the United States because the productivity effect was stronger in the former countries.
But of course the opposite is also possible. A study of robotic automation in midwestern cities indicated that it decreased wage growth and employment, especially for workers at the bottom half of the earnings distribution.
This begs the question: what causes some automations to generate higher productivity? Acemoglu and Restrepo look at one potential reason: productivity boosts seem to be higher when the labor market is tight. Because workers in a tight labor market are typically earning higher wages, the cost savings from automation are much higher. And with those cost savings, firms can afford to recruit and retain workers with higher wages, gaining a competitive edge in a hot labor market.
This seems to bode well for the tight (albeit loosening) labor market we have right now. We’ve already been seeing signs that firms are increasingly interested in automation. But when we automate work tasks using things like generative AI, we should be mindful of the potentially significant displacement effects on workers.
New AI Tasks for Workers
While it’s helpful to have a framework to think about automation, technologies like generative AI won’t just be used for automation. Acemoglu and Restrepo identify that when technologies create new tasks for workers to perform, this universally increases labor demand, through two economic mechanisms:
Price-productivity effect: similar to automation, the new tasks created due to technology are productive, generating profits for firms that can then hire more workers
Reinstatement effect: the opposite of the ‘displacement effect’ for automation (which removes tasks for workers to do), in this situation we are adding new tasks for labor, which effectively increases labor demand
Just like the industrial revolution automated away some work but also created jobs for engineers, machinists, conductors, and financiers, AI has the potential to create new tasks and even entirely new jobs for people. I just checked Linkedin and saw 160+ job listings mentioning GPT-3. But even products built on top of models like GPT-3 or Stable Diffusion will probably wind up getting used in the jobs of tomorrow.and roon, in a recent post about the benefits of generative AI for workers, discuss how workers will have a big, new task to perform: “managing” the AIs that are helping them with work. They envision something they call the “sandwich workflow”:
This is a three-step process. First, a human has a creative impulse, and gives the AI a prompt. The AI then generates a menu of options. The human then chooses an option, edits it, and adds any touches they like.
You could imagine there may be other new tasks that we can’t even fully think of yet. The developers of the internet couldn’t have predicted that people would soon be earning income by renting out their apartments on Airbnb, creating Airtable templates, or even making memes. Just the same, it is early days for AI, and there are many new needs that will come up as the technology matures and becomes more widespread.
Generative AI, as well as AI and automation more broadly, has the potential to change the way we work in profound ways. Unlike what many like to think, this is not universally good or bad.
Automation where the displacement effect predominates is likely to lower wages and employment for workers in that industry. We need to upgrade our public safety net policies so even if workers lose their job or become underemployed due to automation, they can get back on their feet, retrain, and get back into the workforce.
But other types of automation, as well as the prospect of new task creation for workers, will serve to increase both employment and wages. We need to incentivize these further. Acemoglu and Restrepo warn readers that over the past few decades, technology development has focused more on the labor-displacing type of automation.
Not only is this exacerbating inequality, it also restricts economic growth — which harms all Americans.
Of course, we shouldn’t be under any illusions that new task creation will be a piece of cake -- we'll have to work for it. As(writer of the great newsletter) said in a recent post:
[E]ven if innovation creates new tasks—and thus new jobs—displaced people will have to pay a high cost to relocate (e.g. accepting a lower income, a temporal sacrifice to go back to study, and enduring psychological damage) which entails a drastic reduction in life quality.
We need to put the right policies and infrastructure in place to smooth the transition for workers to AI-mediated tasks and jobs.
In 2023, I hope to dive more into this general theme of new technological developments, like AI, and their implications for workers. As we’ve explored in the past roundups, it is an exciting time both for labor and technology — but also a pivotal one. If we are thoughtful about these issues and place our support behind the right policies, we can build an dynamic, equitable economy that works for everyone.
Cheers, and hope you have a happy holiday!
Thanks for reading Workonomics! Subscribe for free to receive new posts and support my work.
Other News and Perspectives
More Perfect Union: Video Overview of the Union-Busting Industry with LaborLab’s Bob Funk
Washington Post: New Penalties for Companies that Illegally Fire Workers Who Unionize
Derek Thompson (The Atlantic): Your Creativity Won’t Save Your Job From AI
- : The State of America's Labor Market
Aaron Sojourner (Upjohn Institute): Wealth grew almost 20% for the bottom 50% of Americans Since October 2021
Dean Baker (CEPR): The November Jobs Report Was Not as Strong as Advertised
The Baffler: Train in Vain
Fast Company: Can your boss make you sleep at work?
Virtual Uncle: Giant VR Robots are Building Railways in Japan