Book review: "Power and Progress"

In which Daron Acemoglu and Simon Johnson fail to convince me that innovation needs to be steered away from automation.

“Do not be fooled by the monumental technological achievements of humankind.” — Acemoglu and Johnson

It’s hardly surprising that Power and Progress made it onto practically every list of the most important business books of 2023. First, there’s the unrivaled pedigree of the authors themselves. To call Daron Acemoglu a powerhouse in the world of economics would be a ludicrous understatement:

Acemoglu is also the main proponent of the institutional explanation for national development, through his famous book Why Nations Fail and its sequel, The Narrow Corridor (both with James Robinson). If you hear me talk about “inclusive institutions” and “extractive institutions”, I’m channeling Acemoglu.

Simon Johnson, meanwhile, is the author of some of my favorite popular books about economic policy, especially Jump-Starting America (with Jonathan Gruber) and 13 Bankers (with James Kwak). When I write more about the need to spend more on science and to restrain the excesses of the finance industry, I’m channeling Johnson.

The second reason this book was destined to garner attention is that it brings together two extremely timely strains of thought: 1) the widespread distrust of tech companies that has grown in American society over the last few years, and 2) the wave of anxiety over AI-driven automation. Power and Progress weaves those two anxieties into a more-or-less coherent whole — a sum of all technological fears, if you will. And it seems to have been spectacularly well-timed, since its release coincided closely with the coming of ChatGPT and other generative AI.

But given all of those powerful tailwinds, I have to say I’m kind of surprised at how little of a splash Power and Progress seems to have made. This is anecdotal of course, but in the 9 months since it came out, I’m not sure I’ve once heard someone reference the book or any idea in it. The authors clearly intended it to be a handbook for people who are scared about AI putting humans out of a job, the way Thomas Piketty’s Capital in the Twenty-First Century became a handbook for people worried about inequality, or Robert Gordon’s The Rise and Fall of American Growth became a handbook for people concerned about technological stagnation. But I don’t think it did.

Why not? One reason might be that the timing wasn’t as favorable as it might appear. Contrary to Acemoglu and Johnson’s assertion (on p.24 of the hardcover edition) that we live in an age of “blind techno-optimism”, the internet is absolutely chock-full of arguments and warnings about the downsides of AI. Concerns over the risk of rogue Artificial General Intelligence resulted in a boardroom coup attempt that almost drove Sam Altman out of OpenAI. Worries that AI wouldn’t uphold diversity led Google to implement some pretty hilarious countermeasures. Fears of mass surveillance, deepfakes, etc. are widespread. And of course the idea that AI is going to lead to mass unemployment is absolutely ubiquitous — so much so that practically every San Francisco tech event I go to features discussions about exactly this subject. Yes, even dance parties.

In other words, Power and Progress may have come out a little too late to make a big splash, and instead ended up just being one more voice shouting in the chorus.

On top of that, though, I have to say that this book…well, I just don’t think it’s very good. I winced while I wrote that sentence, because Simon Johnson is a personal friend, and Acemoglu is a celebrated genius, and because both of them have written such good books in the past. This is the first broadly negative book review I’ve written since 2014, and I’m a lot less combative of a blogger than I was a decade ago. I did not want to pan this book, especially because I think the topic is a good and important one, and I think the authors are brilliant people whose hearts are in the right place.

But I just don’t think the way this book was written ends up supporting the conclusions it draws. The historical examples it cites simply don’t support a narrative of out-of-touch technologists inventing the wrong sorts of technologies and hurting workers in the process. The book embraces a highly questionable definition of “power” in which persuasion in an open democratic society is painted as a threat. It often seems to assume its conclusions about the impacts of specific technologies, and it tells a jumbled and confusing story about the role of productivity growth. And its central claim — that society can push entrepreneurs to steer innovation in a direction that augments humans instead of replacing them — is not well-supported.

All in all, Power and Progress just fails to convince.

The basic idea

Power and Progress is of the “magisterial sweeping tome” class of econ book, like Capital in the Twenty-First Century, The Rise and Fall of American Growth, or Brad DeLong’s Slouching Toward Utopia. Much of the book is a history of technological innovation in general. As such, it tends to ramble; the authors often seem to get so caught up in the telling of this history that they neglect to tie each event to their central theses. In fact, those are often the most fun and fascinating parts of the book. But if I were to boil down Power and Progress to a set of core ideas, it would be:

1. Technological innovation’s impact on human welfare depends crucially on social choices about how those innovations are used.

2. Those choices are determined by the relations of power in a society, and in recent decades our choices have been steered in a negative direction by the power of tech company founders and venture capitalists.

3. The type of technologies that society invents can be chosen so as to distribute benefits more widely, by avoiding technologies that replace workers and inventing technologies that complement workers.

It’s the last of these that the book is most known for, because it’s the boldest, the most original, and the most controversial. But first let’s talk a bit about the other two.

Questionable historiography

The idea that technology’s impact on society is not determined solely by the nature of the tech itself, but depends on how we choose to use it, is obvious enough to be a truism. Everyone knows how the industrial technologies that have created so much wealth are also put to destructive uses in wars. Everyone knows that the same camera technology that lets you talk to your friend in a different city can allow governments to spy on their citizens. Everyone knows that there is a vast system of laws, international agreements, and social norms whose purpose, at least in theory, is to ensure that technology is used for good and not for ill.

But even though “technology can be used for bad purposes” should be a simple truism, Acemoglu and Johnson pick some very odd examples to illustrate the principle. For example, in the prologue, they have a list of what they claim are “new inventions that brought nothing like shared prosperity”. Here’s the fifth item on their list:

At the end of the nineteenth century, German chemist Fritz Haber developed artificial fertilizers that boosted agricultural yields. Subsequently, Haber and other scientists used the same ideas to design chemical weapons that killed and maimed hundreds of thousands on World War I battlefields.

The idea that the Haber-Bosch process has “brought nothing like shared prosperity” is an absolutely wild claim. Nitrogen fertilizers are so important to human existence that by the most common estimates, about half of the entire population of Earth — 3.5 billion people — is only sustained thanks to this technology. But because that same chemical reaction was used to create one particular type of chemical weapon that was responsible for a tiny fraction of the deaths in one particular war, Acemoglu and Johnson feel comfortable saying that a technology that literally gives life to half of humanity “brought nothing like shared prosperity”. It is the kind of claim that is so obviously wrong as to leave the reader slack-jawed — and yet it is deployed in support of an overall thesis for which countless better examples exist.

Unfortunately, this kind of questionable selection of historical examples is a hallmark of Power and Progress all the way through. For example, in Chapter 6, the authors write:

[Belief in the power of productivity] suggests that as technology advanced rapidly during the early phases of the Industrial Revolution, wages should have risen. Instead, real incomes of the majority stagnated.

Acemoglu and Johnson conclude that because textile manufacturing technologies were biased toward automating workers, they immiserated the working class of 1700s Britain. But those same textile manufacturing technologies have been at the center of the early stage of every other country’s industrialization as well. China went through a period where it made most of the world’s clothes, with its share peaking in the late 2000s. In 1995, apparel was China’s biggest export category.

But during this time, when Chinese garment workers were getting the descendants of the original British industrial technologies of power looms, their wages were skyrocketing — as were wages in the economy as a whole. The same is now true of Bangladesh — the country focuses relentlessly on the garment industry, and has access to all of the old automation technologies, and yet incomes in the country have tripled since 1990.

(As a side note, it’s kind of funny that after we’ve used “Luddite” as a slur for technophobes for all these years, Acemoglu and Johnson explicitly try to rehabilitate the original Luddites, writing that they “were right to worry about knitting frames decimating their livelihoods”. For this reason, I considered subtitling this review “the Bible of the Luddites”, but decided that the negative connotation of the word was too strong and it would be rude.)

A third questionable example in the book is the story of the Panama Canal. Acemoglu and Johnson describe the brutal exploitation of the workers who built the canal, and declare the project a “colossal failure”. That brutality was certainly real. But the authors cite it as a reason that the technology of the canal itself failed to bring broad-based prosperity. In fact, the opposite seems true; thanks to the canal, the people of Panama today enjoy a standard of living much, much higher than that of their Central American neighbors. This is not to say those economic benefits were worth the human cost. But the canal’s problems clearly seem associated with its construction, rather than unfair distribution of the benefits from the technology itself. Could the same canal have been built using more humane labor standards? The authors decline to speculate, simply declaring the whole project a failure and not even mentioning Panama’s prosperity.

A fourth dodgy example is the story they tell about Japan. In Chapter 8, Acemoglu and Johnson praise Japanese companies for “combin[ing] automation with the creation of new tasks”, noting that Japanese automakers didn’t reduce their workforces like American automakers did. But Japan’s manufacturing sector wages, like wages throughout the rest of the country, have been falling since the early 1990s, while American wages have stagnated but not fallen. So this story doesn’t fit the data.

A fifth example is in Chapter 7, when Acemoglu and Johnson write that “Henry Ford was a pioneer” in developing “a more cooperative relationship” with his workforce. I’m just wondering how this “more cooperative relationship” involved hiring thugs to gun down union organizers. Ford did pay higher wages to increase efficiency, but his actual dealings with representatives of labor was brutal and intolerant.

I could go on citing these questionable examples — my copy of Power and Progress is stained blue with all the notes I made in the margins — but this review would run into the dozens of pages, and you would quit long before you finished. But because there are so many questionable examples, Power and Progress is the kind of book that must be read closely and with a critical eye.

Source?

Another issue with the book’s examples is the lack of footnotes or endnotes. Instead of citing specific works in support of each specific claim — as most books do — Power and Progress has a bibliographic essay at the end. Many sources are mentioned in this essay, but it’s often difficult, and sometimes impossible, to match the sources to specific claims. As a result, you often end up having to choose between exhaustively searching multiple sources to figure out where the authors got a particular point, or simply giving up and trusting that the authors are accurately representing the data.

For example, in Chapter 1 the authors ask “What if…AI also impoverishes billions in the developing world?”, and asserts that “evidence is mounting” that this concern is “valid”. But where is the evidence that AI threatens to impoverish billions? That’s an astonishingly strong claim about a technology about which little is known, and I can’t find any source in the bibliographic essay. An empirical study I do know is Acemoglu, Autor, Hazell and Restrepo’s 2022 paper “AI and Jobs: Evidence from Online Vacancies”, whose abstract concludes:

We find no discernible relationship between AI exposure and employment or wage growth at the occupation or industry level, however, implying that AI is currently substituting for humans in a subset of tasks but it is not yet having detectable aggregate labor market consequences.

So that paper certainly doesn’t include mounting evidence that AI threatens to impoverish billions. But I can’t find which paper the authors relied on to make this claim.

In fact, because I’ve read many of the Acemoglu papers that undergird this book, I also know that there are instances where the data doesn’t quite say what the authors claim. For example, in Chapter 8, Acemoglu and Johnson argue that “digital technologies became the graveyard of shared prosperity” over the last few decades. In this chapter, they attribute some meaningful piece of the recent rise in inequality to the introduction of digital technologies to the workplace. But because I’ve read Acemoglu and Restrepo’s 2020 paper “Robots and Jobs: Evidence from U.S. Labor Markets”, as well as the working paper version from 2017, I know to be skeptical of this claim.

Acemoglu and Restrepo found that a narrow category of automation — industrial robots — was associated with decreased employment and wages. But as the Economic Policy Institute’s Larry Mishel and Josh Bivens noted, when Acemoglu and Restrepo measured the effect of workers’ “exposure to IT capital” in general — i.e., how much the employers invested in IT tech overall — they found either no effect or a positive effect on employment and wages. Here’s the relevant table from the 2020 version of the paper:

Source: Acemoglu & Restrepo (2020)

The estimates of a positive impact of IT capital on employment and wages are in the original working paper version, in table A9.

Now, this doesn’t mean that computers and the internet weren’t a driver of mass unemployment or stagnant wages. Maybe they were! Acemoglu and Restrepo (2020) could simply be wrong; in fact, since their working paper first came out in 2018, many other studies have ended up contradicting their findings about the negative impacts of robots. And all of these papers look within specific industries or companies — as does Acemoglu and Restrepo’s 2022 follow-up paper about automation and inequality between demographic groups. The overall effect of automation on economic growth, absolute wages, and the composition of industries in the economy simply isn’t known.

In other words, it could be very well be that automation has been impoverishing people, or it could be that it has been enriching people overall. I’d simply like to know where the authors get the data to back up their claim about information technology, especially when one of the authors’ most famous papers appears to contradict that claim.

To sum up, footnotes and endnotes are a technology that has unambiguously benefitted the world, and even though they can be a bit of a pain the butt, authors should include them.

A questionable definition of “power”

Anyway I digress; back to the book’s central theses.

Despite the questionable examples, it’s clearly true that technology can be used to benefit average people or to hurt them. But how does society choose how to use technologies? Acemoglu and Johnson’s answer is “power”, from which they get the title of their book. But what is power? Here, in Chapter 3, Acemoglu and Johnson deploy a definition that veers into the tautological:

Power is about the ability of an individual or group to achieve explicit or implicit objectives. If two people want the same loaf of bread, power determines who will get it.

Using this definition, how could we ever conclude that power wasn’t the reason for an observed outcome? Two people want a loaf of bread, and one of them gets it; we know this was due to “power”, because “power” is defined by who gets a loaf of bread. This kind of definition is semantically valid, but empirically useless; if you define “power” such that it simply means “whatever caused an outcome to happen”, you haven’t isolated causality, you have simply given it a new name.

Acemoglu and Johnson have a reason for employing a definition this infinitely broad; it allows them to include persuasion and compulsion in a single category of “power”.

The authors’ historical examples of when power determined the distribution of the benefits of technology include cases when laws and the threat of violence allowed some people to extract the benefits of technology for themselves — the cotton gin increasing slaveowners’ profits in the American South, or lords extracting agricultural surplus from peasants in medieval Britain. These are instances of compulsion, which certainly fit with our common, everyday, colloquial use of the word “power”.

But Acemoglu and Johnson also spend a lot of time arguing that persuasion is also a form of power. They cite instances in which techno-optimists and businesspeople in 18th century England and 21st century America persuaded the public to enact pro-business policies, through articles, speeches, conversations, and so on. Their explanation for why inequality has increased since the 1970s is, in effect, that silver-tongued technologists managed to persuade American society to weaken pro-worker institutions, and to allow the technologists to invent technologies that replaced human labor instead of complementing it.

The authors don’t venture to say exactly why these techno-optimists’ pro-business vision prevailed — they write that “an idea is more likely to spread if it is simple, is backed by a nice story, and has a ring of truth to it,” but they admit that “quite a bit of this process [of persuasion] is random,” and declare that “you are enormously lucky if you get the right idea, with just the right ring to it, at just the right time.”

I have to admit, this kind of surprised me. I expected to see some sort of pseudo-Gramscian theory of cultural hegemony (or at least some references to Gramsci or similar writers). Instead, the authors just sort of shrug and put it all down to luck. For some reason, the techbros just wrote really good posts, and by doing so they ruled the world — at least until their luck ran out and the world turned against them, I suppose.

In fact, I have to confess that the entire chapter on power and persuasion left me bewildered. I do not understand why we should put accidental success in a nonviolent marketplace of ideas in the same conceptual category as chattel slavery and feudalism. It seems to yield neither understanding nor solutions. Perhaps the historical example of the cotton gin might give us some insight about how to explain the spread of laissez-faire economics, but simply labeling both things as varieties of “power” does not yield that insight. And the idea that persuasion is power doesn’t seem to suggest any kind of systemic fix for the problem that sometimes society is persuaded to do things that increase inequality.

That doesn’t mean Acemoglu and Johnson have no solutions to recommend, though. They want to strengthen institutions like unions and labor laws, but their main idea is to redirect technological innovation toward technologies that complement workers instead of replacing them.

When all you have is a hammer, everything looks like a loom

“With such persuasiveness, you tend to convince yourself that you are correct.” — Acemoglu and Johnson

Over the last six years, Acemoglu and Restrepo wrote a series of theoretical papers in which they lay out a number of different ways that new technology can affect workers’ jobs and wages. Basically, this ends up being a fancier version of a very old idea — capital can either substitute for labor or complement it. If capital substitutes for labor, then capitalists win, because they can replace people with machines, and pay people accordingly less. But if capital complements labor, then workers win, because capitalists have no choice but to hire them to work the new machines, and pay them good wages. Acemoglu and Restrepo make this theoretical breakdown a bit more nuanced, but in the end it really boils down to whether machines replace people or augment their abilities (either by making them more productive or by giving them new things to do).

Power and Progress attempts to analyze the history of technology through the lens of this theory. In eras where technology seemed to progress rapidly but workers’ wages didn’t grow much, like 18th century Britain, Acemoglu and Johnson argue that the main cause was technologists inventing machines that replaced human labor; in other eras where wages grew rapidly, like the late 19th century, they argue that technologists were inventing machines that created new tasks for humans to do.

At times this can begin to feel like a just-so story. For example, Acemoglu and Johnson cite electricity as a technology that was good for workers, because it created so many new industries for people to work in. But didn’t electrification also replace human labor in quite a large variety of ways? Electric lights save us the labor of making candles, electric dishwashers and washing machines and dryers automate our housework, and so on. How do we know this task automation outweighed by the new tasks electricity creates?

They also argue that wages in industrial Britain began to increase in the late 19th century because steamships and telegraphs — as opposed to looms — “expand[ed] the set of tasks and opportunities for workers”. As far as I can tell, this is a claim without evidence. Why was the telegraph’s automation of message couriers less significant than its creation of new jobs for telegraph operators? In Chapter 8, Acemoglu and Johnson blame communication technology for increasing inequality by enabling the offshoring of jobs to China and other countries. Why would modern communication technologies have exactly the opposite effect of the telegraph?

And in Chapter 6, Acemoglu and Johnson praise the early United States for its direction of innovation, writing that American businesses compensated for a lack of skilled labor by, in the words of engineer Joseph Whitworth, “call[ing] in the aid of machinery in almost every department of industry.” In Chapter 7 the authors write that the American use of interchangeable parts “was first and foremost an effort to simplify the production process so that workers lacking in artisanal skills could produce high-quality products.” But how is that any different from the British use of power tools to help unskilled make textiles in the 1700s? The difference is never explained.

On the other side of the coin, Acemoglu and Johnson cite most modern information technology as something that automates more tasks than it creates. But what about all the new tasks that IT creates — mobile developers, web designers, digital media marketers, content moderators, and so on and so forth? Why are these less economically important than the tasks that the internet automates away (encyclopedia salespeople, etc.)?

One answer is that if we assume that Acemoglu’s theory of automation is the main thing that’s going on, we can just infer the effects of particular technologies from macroeconomic outcomes — if we see wages stagnate, it must be because task automation outweighed task creation. But for anyone who suspects that Acemoglu’s theory might not actually be the main thing going on in the economy, just saying that the proof is in the pudding is a bit unsatisfying. It feels like a just-so story.

Is productivity good or bad?

In fact, the historical examples in Power and Progress leave themselves open to alternate narratives. The main alternative narrative is about productivity.

Acemoglu and Johnson repeatedly argue that if productivity gains are produced by automation, workers don’t see the benefits. They cite the idea that productivity naturally uplifts workers — which they call the “productivity bandwagon” — as one of the main nefarious narratives that technologists use to persuade society to allow them to invent technologies that replace workers.

But what if the “productivity bandwagon” narrative is true?

There are two main historical periods the authors cite as examples of excessive automation leading to stagnant wages — the early Industrial Revolution in 18th century Britain, where textile machines like looms replaced human artisans, and America since the 1970s. (Note: they make an error when they say, in Chapter 8, that “declines in real wages…have been a major part of U.S. labor market trends.” In fact, when you include benefits, real hourly compensation has grown a bit more slowly since 1973, but has still consistently risen.)

But Acemoglu and Johnson also note that both of these eras had sluggish productivity growth. Perhaps wages were stagnant in those eras because productivity was also stagnant?

Regarding the early Industrial Revolution, some researchers argue that the labor share didn’t actually fall. For example, here’s Crafts (2020):

[R]eal wages grew more slowly than real GDP per worker during the industrial revolution. However, the discrepancy was much less than has been claimed such that in 1820 the former had risen by about 12 per cent since 1770 and the latter by about 16 per cent. Second, labour productivity grew quite slowly prior to 1830 averaging a little below 0.4 per cent per year in the 60 years after 1770. Nevertheless, in the context of demographic pressure this was a very good outcome by pre-industrial standards. Third, as relative prices changed and exportable manufactures became cheaper, over the long run real product wages grew somewhat faster than real consumption earnings. Fourth, the share of profits in GDP rose over time from 17.2 per cent in 1770 to 31.3 per cent in 1860 but this was associated with a decline in the share of land rents and the share of labour was little changed. Fifth, looked at through the lens of growth accounting the evidence is of total factor productivity (TFP) growth accelerating only gradually to 0.6-0.8 per cent per year during 1830 to 1860 with the steam age only materializing after 1830.

In sum, this looks more like a story of paradoxically slow productivity growth than of pro-rich growth. The story of the industrial revolution is definitely not one of a new general-purpose technology boosting productivity growth at the expense of a big shift in the distribution of income which is the current fear about AI.

As for the U.S. since the 1970s, inequality has definitely increased, but — contrary to what you may have heard — pay has largely kept pace with productivity. Automation may have made wages more unequal (this is the argument of Acemoglu and Restrepo’s 2022 paper), but the modest decline in labor’s share of the national pie was probably mostly about land values increasing. Which means that aggregate wage stagnation was largely due to slowing productivity growth in recent years as well.

In fact, Acemoglu and Johnson also blame automation for stagnating productivity! In Chapter 8, they write that “productivity gains from automation may always be somewhat limited.” They coin the term “so-so automation” to describe technologies that take humans out of the loop but fail to increase productivity much by doing so. They argue that technologies that make better use of human capabilities lead to faster productivity growth as well as higher wages and lower inequality.

OK so…why isn’t that the book’s central this? You could write a very interesting book about how technologies that complement humans are better for both productivity and broad-based prosperity than technologies that try to substitute for humans wholesale. I would definitely read that book! But Acemoglu and Johnson did not choose to write that book; instead, they warn against a focus on productivity, claiming that it's a seductive but dangerous narrative used by the greedy, fast-talking techbros. In my mind, this weakens their overall narrative.

Where’s the menu?

Throughout Power and Progress, the authors tell a story about a “menu of technologies” that entrepreneurs can choose from. On one hand, companies can choose to invest in automation, replacing workers, increasing inequality, causing slow wage growth, and maybe reducing productivity in the process. On the other hand, they can choose to invest in technologies that create new tasks for humans to do, thus increasing wages and decreasing inequality. Their story is that out of greed and/or elitism, entrepreneurs often choose the former, so it’s in the interests of society to push them toward choosing the latter.

But when the authors approvingly cite examples of new industries springing into being, they never give an explanation of why entrepreneurs and technologists chose to create these new industries, instead of trying to cut costs in existing industries. My default assumption would be that the people who invented and commercialized steamships, telegraphy, interchangeable parts, autos, electricity, and telephones were driven by the same sort of motivations that animated the people who invented and commercialized power looms, computers, and the internet. If not, why not? Was there ever a case when governments or unions pushed entrepreneurs to select Option A from the “menu of technologies” instead of Option B?

When Acemoglu and Johnson do discuss union power, it’s in the context of worker training. In Chapter 7, they write:

In fact, for unions [in the 1960s] the central issue was worker training. They insisted on training provisions to ensure that workers could be brought up to the necessary skill level to operate the new machinery and benefit from it.

This is very different from affecting the direction of innovation! This is a case of workers collectively pushing companies to invest in human capital, so that worker skills can catch up to the direction in which innovation was already going.

As far as I can tell, this book does not contain even one single example of when a union or government supposedly pushed an entrepreneur or company to choose a different path of technology in order to benefit workers. As far as I can tell, it does not even contain one single example of when an engineer, entrepreneur, company or investor chose to create a technology in order to benefit workers more.

In other words, there is no evidence here that the “menu of technologies” actually exists. It’s not clear that technologists and industrialists even know in advance whether the inventions they create and commercialize will create more new tasks than they automate. And this raises pointed and troubling questions for the authors’ preferred solution to the problems of inequality and wage stagnation.

Can the mandarins act as a check on the techbros?

In Chapter 11, Acemoglu and Johnson roll out their proposed solutions. Having concluded that inequality and wage stagnation are due to “tech billionaires and their agenda” choosing the wrong technologies from the “menu”, they call for democratic people-power to force the techbros back onto the labor-augmenting path.

What’s totally unclear is how to do this. Acemoglu and Johnson admit that “redirecting” the path of technological innovation is going to be an incredibly tall order:

Determining how different digital technologies are used and their impact on wages, inequality, and surveillance is much harder [than assessing their climate impacts]…Moreover, given the difficulty of distinguishing automation from other uses of digital technologies, automation taxes are currently not practical.

And yet the authors still claim that this can be done! Yet they’re maddeningly vague on the details:

There is a telltale sign of automation technologies: reducing the labor share of value added, meaning that once these technologies are introduced, how much of value added goes to capital increases and how much gets captured by labor decreases…On this basis, technologies that increase the labor share can be encouraged via subsidies for their use and their development.

But how do we know in advance, before a technology is invented, whether it will increase or decrease the labor share? This is just replacing one target of guesswork — new task creation vs. automation — with another target of guesswork.

Fundamentally, it still boils down to some sort of mandarins in a room somewhere — economists? government engineers? bloggers? — trying to assess the economic effects of a technology that doesn’t even exist yet.

As I wrote in a post last June, this is probably an impossible tasks. Some of the world’s top experts thought that AI would replace radiologists within a few years, but demand for radiologists boomed even as the new AI tools were coming online. The technologists got it wrong.

And the economists are just as likely to get it wrong. For example, industrial robots are the one technology that Acemoglu consistently rails against as an example of harmful automation. That’s based on his 2020 paper with Restrepo, where they found that companies that buy more robots employ fewer humans. But a whole bunch of other economists followed up on this research and found the exact opposite — robot adoption is correlated with more jobs at a company or in an industry. Here was a list I made back in 2022:

I’ll list a few of these studies:

1. Mann and Püttmann (2018) — Where Acemoglu and Restrepo…looked at correlation, this paper attempts to identify causation. They look at automation-related patents in an industry — a proxy for innovation in the automation space — and then look to see whether that industry gains or loses jobs. They find that “advances in national automation technology have a positive influence on employment in local labor markets”, though this isn’t true for every area.

2. Dixon, Hong and Wu (2021) — These authors looked at robot adoption in Canada, at the level of the individual company (or “firm”, as economists say). They found that companies that adopted more robots hired more people, while also improving the quality of their products and services.

3. Koch, Manuylov and Smolka (2019) — This paper looks at firm-level data for manufacturing companies in Spain. They find that robot adoption is associated with a substantial increase in employment as well as output.

4. Adachi, Kawaguchi and Saito (2020) — This paper finds the same thing as the previous one, but for Japanese companies over the course of a 40-year period.

5. Eggleston, Lee and Iizuka (2021) — These authors look at robot adoption by nursing homes in Japan, and find that it strongly increases employment, although it does result in existing nurses working fewer hours (and thus getting paid less).

6. Hirvonen, Stenhammar, and Tuhkuri (2022) — This paper looks at a technology subsidy program in Finland that increased adoption of a broad range of advanced technologies at Finnish firms. They find that this led to employment increases.

In fact, by this point the trend is clear. Essentially everyone is finding that, contra Acemoglu and Restrepo…robots are correlated with — and probably cause — higher employment in the companies and areas where they’re adopted.

What’s happening is that companies that use more robots hire more humans (and retain their existing humans) in jobs that complement the robots. That’s exactly what we saw with previous waves of automation — people find new roles, robots increase their productivity, and they get paid more. Looking at the countries that use the most robots in their manufacturing industry, it seems likely that this virtuous cycle is happening even at the level of whole nations.

Zooming out from just the manufacturing sector, Hötte, Somers, and Theodorakopoulos have a very interesting 2022 review paper in which they look at the literature on the entire range of automation technologies. Here’s an article in which they explain their results. Hötte et al. find that automation does replace jobs, but that this effect is outweighed by the “reinstatement” effect — in other words, people find new jobs to do. And their incomes generally rise as a result.

So no, there’s no possibility that a council of mandarins — engineers, economists, or whoever — can sit there evaluating every potential new technology that companies or inventors want to create, and deciding whether it will raise or lower the labor share. I mean, you could make a council of mandarins, and it could look at plans for new technologies, and it could issue decisions, but in practice it would be throwing darts at a dartboard. And it would be an incredibly costly tax on our companies, since it would introduce massive delays into their decision-making process. Their Chinese rivals, on the other hand, would suffer no such delays.

In other words, I see no hope for Acemoglu and Johnson’s preferred solution. The utter vagueness with which the idea is presented in Power and Progress doesn’t suggest that the authors have thought carefully about how this solution might work in practice.

In particular, these solutions seem inferior to something far simpler: policies to increase labor share ex post. Labor market institutions like co-determination and sectoral bargaining, and direct interventions like wage subsidies funded by taxes on capital income, can push up the labor share without requiring panels of experts to predict the unpredictable. And if entrepreneurs really do have any degree of foresight about whether their innovations will tend to push the labor share up or down, these policies will act like a Pigouvian tax on the kind of cost-cutting that Acemoglu and Johnson decry. With a wage subsidy, for example, the higher the market rate you can afford to pay your workers, the more of a rebate you can get from the government. So if there are technologies that augment your workers and let you hire new workers, a wage subsidy gives you an incentive to create them.

Anyway, I think simple policies like these should be economists’ first go-to solutions, rather than the creation of whole new social institutions.

The old narrative, and the new

At the start of this review, I talked about how Power and Progress may have missed its moment by getting lost in a flood of fears about AI. But there’s another way in which the book might be poorly timed. Wage inequality — the very thing that Acemoglu and Restrepo (2022) try to explain — has flatlined since the early 2010s.

Source: Realtime Inequality

Meanwhile, real wages have been rising strongly for years now, interrupted only by the post-pandemic inflation. And wages for production and nonsupervisory workers have risen more robustly than those for managerial workers:

Source: Arin Dube

Meanwhile, employment for prime-age Americans is near all-time highs, and unemployment is at record lows; everyone who wants a job in America has one.

All this has happened in exactly the time frame during which AI has exploded. Predictive AI burst onto the national scene in 2012 with the ImageNet paper, the basic technology for generative AI was created in 2017 with the transformer paper, and generative AI became really widespread in 2022-23 with LLMs and AI art programs. To reiterate: essentially all of the commercialization and implementation of artificial intelligence has happened during a time in which wages have been rising, inequality has been flat or falling, and employment has been high.

Maybe AI just isn’t big enough to kill all the jobs yet; maybe we just have to wait a few years and we’ll all be unemployed or working for pennies. Or maybe AI is actually the kind of technology that improves task productivity and creates new tasks instead of automating old tasks away. Or maybe Acemoglu has something wrong in his models, and further theoretical and empirical explorations will overturn his conclusions about the key role of automation in fostering inequality. I’m not sure.

Whatever is going on, though, I think it should give the AI worriers pause. This was not on the menu. If you went back to 2012, and asked people to predict the impact of a new machine that could recognize objects and imitate speech and create beautiful art, they probably would have assumed that the rising inequality that they had experienced for the last 30 years would now be turbocharged. And — at least so far — they’d have been completely wrong.

To me, that thought experiment illustrates the folly of trying to predict the economic effects of new technology. It also suggests another reason why Power and Progress didn’t make the same kind of splash it might have made in 2018. Obviously, our economic problems haven’t all been solved. But perhaps, underneath all of the anger and pessimism, Americans realize that something has shifted in their economy, for the better. And perhaps that’s making them a bit less interested in the kind of pessimistic economic narratives that flew off the shelves in the 2010s.

---