Review of "Power and Progress"
The book by Acemoglu and Johnson has nothing of interesting to say about the future.
I’ve been on a quest to better understand the impact of AI on jobs. This week, I read Power and Progress, by MIT economists Daren Acemoglu and Simon Johnson. I was disappointed by how little they had to say about the future.
. . .
In tech circles, the attitude towards AI-powered automation has been bimodal. There are the doomers on one side, the techno-optimists on the other, and not many voices in between. But what does academia have to say? I spoke to a friend who suggested the works of Acemoglu. He’s an MIT professor who shared the 2024 Nobel Prize in Economics for his work on institutions and how they determine the prosperity of a country. In 2012, he co-wrote the best-seller Why Nations Fail about this.
A decade ago, Acemoglu turned his attention to automation and its effects on the economy. He applied a “task-based framework” to better reason about how technology displaces and later reinstates workers.1 (I recommend this paper from 2019 as a good entry point.)
The task-based framework got me excited. I saw it allows us to make concrete predictions about the rollout of AI. That’s what I expected going into Power and Progress.
. . .
The book, published in 2023, is mostly a history of automation in the West, starting with the late Middle Ages. It tells the following story:
Historically, automation has been short-term negative for workers, and long-term positive. In the short-term, it compresses wages and worsens working conditions. This is especially true when workers lack the freedom to leave their employers, as under serfdom, or to fight for better working conditions.
Over time, automation turns into a positive. It makes goods affordable, which improves quality of life for workers. Lower prices increase demand, which increases economic output, and causes workers to be rehired. New tasks emerge that machines can’t yet automate. This is happened throughout the industrial revolutions in Europe, and in the US during the postwar boom, from the 1940s to the early 1970s.
In the 1970s, the situation flips. Manufacturing jobs start getting shipped overseas; the computer revolution begins to automate routine back-office work. These and other factors end up squeezing unskilled labor. Wage growth in the US stalls for all but the highly educated, never to pick up again.
The book blames this reversal of fortune on an ideological shift in America. Deregulation; neoliberalism; the rise of business schools preaching cost-cutting… This ideological shift is used to prop up the thesis of the book: that the impact of any technology can be positive or negative — it depends on what the public and private sectors choose to do.
Technology does not have a preordained direction, and nothing about it is inevitable. Technology has increased inequality largely because of choices that companies and other actors have made. (p. 263)
This becomes their core take on the future of AI and automation. AI can be negative or positive for workers; it’s up to civil society to guarantee the latter.
. . .
I don’t have much to say on the first half of the book. I know little about the industrial revolutions, or the impact of unions in 20th century America. The authors’ account of automation prior to the 1970s sounded plausible. After that, I think the book starts to fall apart.
What bothers me most is the central thesis: that technology has no “preordained direction”. I believe it’s wrong for two reasons. First, because certain automations are much more hostile to workers than others. Second, because market forces and political realities often do let us predict the direction of technology adoption. So the idea that “nothing about [technology] is inevitable” falls somewhere between a misleading tautology and a lazy conclusion.
In Chapter 9, the authors identify two conditions required for automation to benefit workers: it must create new tasks, and the gains must be shared with workers.
There were two pillars of shared prosperity in the postwar period: alongside automation, new opportunities were created for all kinds of workers, and robust rent sharing (meaning the splitting of productivity and profit gains between capital and labor) kept wages buoyant. (p. 258)
This passage alone seems to contradict to the book’s central thesis. If technology is neutral, we’d expect no correlation with how it splits rent2 between shareholders vs. workers, or how many new tasks it creates. But that’s obviously not the case. Some technologies are naturally more disruptive than others.
One of the most interesting contributions from Acemoglu’s 2019 paper is the concept of “so-so automations”. These are automations that replace workers with machines that are only slightly more productive. One such example is self-checkout kiosks. Kiosks are cheaper than cashiers but not cheap enough to lower prices for the end consumer. So-so automations displace workers without generating savings that increase demand and with it economic output. This is displacement without reinstatement.
So why do the authors insist on framing technology itself as neutral? Can they not spot the obvious contradiction?
I think they can. The authors want to make the point that, for any technology, government and civil society still have a say on how it’s applied. We can choose whether to adopt it, how quickly to do it, how much support to direct towards workers being automated, and so on.
It’s an inspiring thesis for a book about automation; it’s also ultimately useless. By the same lens, enriched plutonium is not dangerous because nuclear weapons don’t build themselves: humans build nuclear weapons. It’s the same logic behind “guns don’t kill people, people kill people”. These ideas are true but they entirely miss the point.
. . .
In the direction of AI really that undecided? I don’t think so. In 2026, white collar automation feels inevitable. The open question is how much of it will be automated. Trillions of dollars worldwide have poured into infrastructure and businesses built on this assumption. We’re moving fast in a very particular direction.
More fundamentally, I don’t believe in the project of the book. Knowing what I know about LLMs, I’m skeptical that a history of the industrial revolutions has much to teach us about what’s coming. The parallels are thin. When you try to draw a line from the spinning jenny to GPT-4, all you can do is wave your hands and say “who knows” — which is exactly what the authors do. The historical framing is weak. It gives the book a pass on doing harder, more specific work about the technology actually in front of us.
Ultimately, Power and Progress has nothing to contribute on the future of jobs after AGI, or how automation may play out. To the readers interested in this question, I suggest reading Acemoglu’s papers and skipping the book. 3
Traditional frameworks of productivity look something like this:
They model technology as a multiplier on the effectiveness of labor (employees) and machines (capital). This can be a useful way to reason about displacement but empirically leads to all sorts of issues. For example, it can’t explain the observed decline in real wages of low-skilled workers in the US since the 1970s.
Since I originally posted this, a friend pointed me at Bill Gates’ review of Acemoglu’s best-seller Why Nations Fail, from 2014: https://www.gatesnotes.com/why-nations-fail. I found some parallels between Gates’ critique of that book and my issues with Power and Progress.