Two Cheers for Automation

When an economist blames technology for the challenges facing workers, run, do not walk, the other way. If you stick around too long, he is likely to pivot from his lecture on “skills-biased technological change” straight into another one about how “labor shortages” are holding the economy back—apparently, the reduced demand for workers makes it difficult to find workers to fill all the demand. The whiplash can be quite hazardous.

Misunderstandings about the relationship between technology and work are as old as capitalism itself, dating back at least to the Luddite revolt against textile machinery in the early 1800s. Intuitively, technological progress undermines workers and their interests. The machine encountered in an office or factory or store tends to be doing something a worker might once have done, and the competition seems fundamentally unfair. The machine needs neither wages, nor benefits, nor breaks. It has no family to feed, no aspirations. For the person who remains on site, the machine may have eliminated the human element that made a job meaningful, replacing personal relationships and artisanal craft with a panel of buttons to push.

Yet at the same time, technological progress is responsible for most of the positive developments in the experience of work over the past two centuries. Few good jobs would exist without machines. Most jobs that have always existed in some form now offer better pay, are less physically strenuous or dangerous, and achieve more thanks to machines. For all the hesitation about whatever new technology might be around the corner, few workers can identify a major existing technology, once new itself, that they would wish away if they could.

Reconciling these narratives in a coherent economic framework begins with a simple statement: Technology allows a given amount of labor to produce greater output. Put even more succinctly, it boosts productivity. From there, all effects, good and ill, follow. Which effects follow are almost never inherent to the technology itself, but result from the economic, social, and policy context in which it is deployed.

Rising productivity is necessary for the sustainable economic growth that delivers improved outcomes for workers, but it is not sufficient. Opposing progress is self-defeating and prevents any possibility of desired gains for workers materializing. Blind embrace of progress, in the style of techno-optimism and market fundamentalism, can lead to worse outcomes—making stubborn opposition rational by comparison. The American economy desperately needs faster productivity growth, but it also needs better incentives for investment to leverage that growth into higher output and more worker power to ensure that the growth translates into better and higher-paying jobs. Policymakers, innovators, investors, business leaders, and unions all have roles to play in creating the conditions for technological progress that can earn wide political support and generate widely shared prosperity.

Technology as Productivity Improvement

Generally speaking, firms introduce new technology into a production process when the result will be to increase the value of outputs relative to the cost of inputs. Consider a hypothetical Worker Augmentation Module, or “WAM,” which cuts the time required to make a widget in half. Whereas a factory once required two full shifts to produce 10,000 widgets per day, it now only needs one shift. How would we describe the WAM’s effect, and should workers celebrate it?

If the WAM is a robotic arm that rapidly attaches various components, we would call it “automation” and perhaps worry that it has rendered half the workers superfluous, likely leading to a substantial reduction in employment. But if the WAM is an app on the worker’s phone that optimizes the sequencing of necessary customizations for each individual widget as it progresses through production, is this still “automation” or is it instead “process improvement”?

What if, instead of optimizing a process, the app delivers real-time coaching to the worker, allowing him to execute the same steps with the same equipment much more efficiently? Or perhaps the module is a training exercise that the team works through together for 30 minutes before hitting the shop floor. Now it is “skill enhancement.” Does training eliminate jobs, to the detriment of workers? Good luck finding someone who talks about it that way.

For any hypothetical WAM, the narrative might also change in another way. A factory that automated half the steps in the production of 10,000 widgets sounds like one poised to eliminate half its jobs. But a factory that trained its workers to produce widgets twice as fast sounds like one poised to double its output to 20,000 widgets daily.

Finally, for any WAM and its results, the vital question remains: cui bono? In some combination, units of widget output are up and units of labor input are down. But how much are costs down—is the reduction in labor cost per unit of output offset entirely by the increased cost of the new technology? Neither robotic arms, nor training staff, nor apps are free. If the money once paid to workers is instead invested in machinery, who manufactures the machinery? Perhaps workers are earning better wages doing that, or perhaps the equipment is simply imported from China. Most importantly, how much of the remaining gain is captured by workers in higher wages versus the firm in higher profits?

The Industrial Revolution provides a quintessential illustration of how the very same technology can benefit or harm workers. In its initial years, despite extraordinary productivity gains, the introduction of steam power and more advanced machinery to factories proved a disaster for the working class. “The average person was not reaping the benefits of economic change,” The Economist concluded, a fact best established by declining height and life expectancy. That is not because industrial automation is bad for workers, as subsequent periods of extraordinary gain make clear. Rather, the problem was that technology was introduced within a political, social, and economic framework that treated labor as an expendable input and minimized its power.

In The Technology Trap, Carl Benedikt Frey provides an especially useful case study. Introducing machines that reduced the need for physical strength led to the rapid expansion of child labor, often in conditions of extreme exploitation. Children came to account for half of textile workers, and “many were pauper apprentices who worked in factories far from their families and friends. When they made up the bulk of the workforce, as they often did, they lacked the protection others were given by the mere presence of adult coworkers. They were often consigned to work without wages or rewards.” Obviously, this arrangement was detrimental to both children and any adult workers who might attempt to compete with them. Only after the introduction of labor regulations did industry pursue the adoption of steam, the installation of the larger and more complex equipment it could power, and the creation of good jobs for adults operating it all more productively.

Gains for workers also tend to initiate a virtuous cycle in which higher incomes lead to rising demand, which ensures that productivity gains lead to rising output rather than reduced employment, which leads to higher incomes and yet more demand. In 1914, Henry Ford famously doubled wages to $5 per day, in part to ensure that his workers could afford the cars they were helping to build. Throughout the middle of the twentieth century, with union density and worker power at their peak, strong productivity growth in the manufacturing sector was matched by even stronger output growth, ensuring that wages rose rapidly and employment expanded. The middle class emerged and prospered.

Contrast that experience with the first quarter of the twenty-first century, an era when wages stagnated and manufacturing employment collapsed. Economists have been quick to blame automation for these problems, but productivity growth is slower now than it was in the earlier era—for all the talk of robotics and now artificial intelligence, generating more output with as many or fewer workers has been getting harder.

The struggles for labor are the consequence not of improved technology, but of stalled output. The American consumer’s appetite remains nearly insatiable, but floods of cheap imports sate it. The domestic industrial base has eroded, investment has slowed, and productivity growth has been negative for the past decade—an absence of investment in new technology has meant more workers are required than in the past to produce the same output and global competitiveness has declined as U.S. producers fall far behind counterparts around the world in automation. Advanced economies like Germany, Sweden, Japan, and Korea deploy robots at far higher rates. So does China, which is now installing half of all industrial robots globally, allowing it to pass German density and approach the Korean level. As that diverse list of countries underscores, aggressive automation hardly guarantees good treatment of workers. But the United States cannot hope to rebuild its own industrial base or improve the quality of American manufacturing jobs without closing the gap. Quick adoption of technology is not the problem facing American workers— and slow adoption is doing them no favors.

The Tech Stall

The same story is playing out across a range of sectors, where productivity growth has been slow and the major complaint is a shortage of workers available for the labor-intensive roles that firms have failed to supplement with technology. In the construction industry, productivity has been falling for 50 years. The healthcare sector is notorious for low productivity growth and reports labor shortages at every level, from doctors to home health aides. The agricultural sector has expanded its use of workers with H-2A seasonal visas from fewer than 50,000 in 2005 to more than 300,000 in 2023.

In those sectors where the technology of the digital revolution has been deployed most effectively in recent decades—white-collar professions that rely heavily on symbol manipulation, data processing, and connectivity—employment has grown rapidly. From the legal profession to fields such as business and financial services to armies of computer programmers, productivity gains have made workers more valuable and led to the creation of many more positions. (It is an open question whether AI will change this.)

As a result, the best evidence suggests that unequal productivity gains have been a major driving force behind the rise in income inequality. Pay at the top has not diverged from pay at the middle because those at the top are unfairly expropriating more than their fair share; pay for those at the top has diverged from pay for everyone else because opportunities for those at the top to deliver more output with less labor have diverged, and those opportunities are precisely what deliver gains to workers. The American Enterprise Institute’s Scott Winship explains the dynamic well:

“Industries with a higher level of education in 1989 saw stronger productivity growth through 2017. … A recent paper finds that firms with higher productivity have a larger wage gap between their highest- and lowest-paid workers. Even more strikingly, increases in firm productivity raise the pay of all the firm’s employees, but not equally. The highest-earning workers in a firm with productivity growth receive a bigger earnings boost than do the lowest-earning workers. … Rising productivity in the economy may be driven by greater inequality in productivity and productivity growth and, hence, lead to relatively sluggish growth in median pay. …

There is another way in which productivity growth became less equal, which relates to the shift to a lower-productivity service economy and the way that shift differentially affected men and women. … Part of why men have done worse comes down to their earlier overrepresentation in high-productivity sectors of the economy, such as manufacturing, which employed a smaller share of the workforce over time. … Productivity is generally lower in the service sector, since there are fewer opportunities to improve efficiency using machines, computers, and equipment. Think of haircutting or performing in a play. Moreover, over the long run, productivity grows at a slower rate in the service sector for the same reason. … Over time, men increasingly ended up in lower-productivity, lower-paying service jobs.”

The end result is a labor market dominated by low-quality, insecure jobs. In the modest definition developed by American Compass, a “secure” job is one with annual pay above $40,000, health insurance and paid time off, and predictability in scheduling and future earnings. As of 2023, only 40% of the economy’s jobs met that definition; among workers without college degrees, only 30% had secure jobs. This is not a market disrupted by extraordinary progress that allowed a much smaller number of workers to do the jobs of many, but rather one where the disruption seems never to have taken hold.

Is AI Different?

The new new thing, of course, is artificial intelligence, which advocates insist will transform the labor market and render entire classes of workers unemployable in a matter of years. But as with the last wave of innovation, which never quite reached shore, the AI revolution is thus far delivering froth without impact. As the Economic Innovation Group demonstrated recently, the unemployment rate is rising most slowly for those jobs that should be most susceptible to AI. The aggressively reported stories of new college graduates facing sudden challenges landing jobs describe a trend that has been underway for at least five years and say more about the higher education system’s failings than the competence of Claude.

Indeed, assuming that AI is having no effect on anything gets one closest to an accurate prediction. Economy-wide productivity in mid-2025 is precisely where the Congressional Budget Office forecasted it would be at the start of 2020, long before the release of ChatGPT. Most amusingly, Federal Reserve economists Mark Wynne and Lillian Derr have compared the predictions made about automatable jobs ten years ago to those forecasters are making today, and found no correlation between the two sets of predictions or between predictions and actual results.

Microsoft CEO Satya Nadella has made the link between technological progress and productivity growth explicit, arguing that abstract definitions of “artificial general intelligence” are beside the point, because proof that computers have surpassed human capabilities should be readily apparent in economic data suddenly showing 10% annual growth. And indeed, if AI achieves this effect, it will be operating far beyond all prior human experience. Perhaps, in that world, technology would be a genuine threat to workers.

But while Nadella has proposed 10% growth as a litmus test, almost no one expects it to be met. For instance, Tyler Cowen, an economist generally quite bullish on AI’s prospects, has suggested that AI might add a quarter to a half of a percentage point to annual growth. That would be meaningful, but hardly transformative. An economy achieving 1.5% annual productivity growth doubles the output from a given set of workers in 47 years; increasing the growth rate to 1.8%, the doubling occurs in 39 years. That is a meaningful increase in prosperity—in year 40, the faster-growing society has input more than 10% higher. But it is not the difference between progress that could benefit workers and progress that would pass them by. Given the choice, 1.8% productivity growth is unquestionably better, 2.1% better than that, 2.4% better still.

How can technologies that seem so transformative—whether robotics and other automation, computers and the Internet, or AI—consistently deliver labor market impacts so gradual? Even in Korea, which has automated most aggressively, the latest evidence suggests that initial disruption gave way quickly to sustainable gains for workers. At least four factors likely come into play.

First, in our imagination, we grant the new breakthrough far greater power than the old one we already take for granted. The great economic historian Robert Gordon has shown that, despite all these technological innovations, growth in U.S. real GDP per capita has never exceeded 2.5% for an extended period. Yes, advanced robotics and AI are impressive. But are they more impressive than the personal computer, or the car, or electricity? In some respects, those earlier breakthroughs had the benefit of picking the lower-hanging fruit, whereas each subsequent step brings greater cost and complexity with less marginal benefit.

Second, deploying new technology is hard and goes slowly. Fifty years after Thomas Edison first demonstrated a lightbulb lit by a power generator, fewer than one in ten American farms had electricity. The infrastructure to support deployment can take decades to build. Transforming processes within firms can take years longer. If AI could “destroy” half of all jobs in the contemporary economy over the next 25 years (i.e., allow half of today’s workers to produce the output that currently requires all of them), that would be an extraordinary and unprecedented disruption. This would also imply an annual productivity growth rate of just 2.8%.

Third, while many individual tasks have the potential to be automated, almost any job consists of a wide range of tasks, only some of which meet that criterion. Thus, rather than entirely replacing the need for a human worker, technology invariably augments what the worker can do, enabling more or better output. Ultimately, no technology can be deployed faster than the people who will need to work with it can learn to use it effectively. Early efforts at using AI for computer programming offer a good example. This area is the one in which, at this point, AI appears to offer the greatest potential for replacing substantial amounts of human labor. One recent paper from Stanford University suggests that AI is affecting employment in the sector, mostly by reducing new hiring into entry-level jobs. But across the economy, employment in software publishing and computer system design has declined less than 2% from its 2022 high and remains 15% above its level of five years ago. In a detailed study of programmers using AI, Model Evaluation & Threat Research found that while everyone expected AI-driven increases in productivity across a range of tasks, in fact using AI tended to reduce productivity.

Fourth, as technology changes the contours of a job, it also creates new opportunities. The self-driving truck can move itself from point A to point B, but it still needs someone to load and unload and, as is often the responsibility of a delivery driver, stock the shelves. If the driver job transforms from one of traveling long distances on the road for most of the year to one that spends part of each day overseeing a fleet of trucks from a control center and part of the day unloading all the trucks that make deliveries in a certain area, the worker in the role can earn far more than the one watching the miles fly by, and he can be home for dinner every evening.

Blaming technological progress and productivity growth for bad labor-market outcomes is not only inaccurate, but also deeply counterproductive, leading workers to resist precisely the investments that would most benefit them and giving employers a convenient excuse that insecure jobs and low pay are somehow “inevitable” results of desirable change. The message instead should be that the United States is going to accelerate productivity growth and ensure that everyone shares in the gains: with better and higher-paying jobs for workers, more and higher-quality output for consumers, stronger firms better able to compete globally, and higher returns on investment. That narrative is entirely coherent; that outcome is entirely plausible. Indeed, it is the mechanism by which American capitalism has delivered rising prosperity in those eras when it has been working well.

What Needs to Change?

Of course, one badly needed change is higher productivity growth. Much could be said about the policies related to innovation, investment, workforce, and so on that would foster the development and deployment of more productivity-enhancing technologies. But for those policies to gain widespread support, and for the resulting productivity gains to actually deliver benefits for workers, thereby validating and further building support, several key features must be present.

First, business leaders and investors must return to a mindset that prioritizes long-term growth over short-term value extraction. Deploying technology to boost productivity will always leave a choice between using that higher output per worker to increase output or using it to eliminate jobs. An unfortunate side effect of financialization generally, which encourages managers and shareholders to treat firms as financial assets for generating cash rather than as organizations of people that exist to create value, is that reducing headcount has become an end in itself and a sign of effectiveness.

The Wall Street Journal captured the dynamic well in a July report titled, “CEOs Are Shrinking Their Workforces—and They Couldn’t Be Prouder.” Announcing layoffs or hiring freezes used to be a sign of distress or an admission of mismanagement. Now, at Wells Fargo, 20 straight quarters of declining employment is cause for celebration. Verizon’s CEO defines being “very, very good” on headcount as “it’s going down all the time.”

That approach is antithetical to capitalism, which, going all the way back to Adam Smith’s The Wealth of Nations, presumes that “upon equal, or only nearly equal profits, therefore, every individual naturally inclines to employ his capital in the manner in which it is likely to afford the greatest support to domestic industry, and to give revenue and employment to the greatest number of people of his own country.” Under those conditions, the “invisible hand” ensures that the individual pursuing only his own self-interest serves the public interest too. But if the profit-maximizing strategy instead calls for employing the fewest number of people, no one should expect capitalism to deliver prosperity. This was understood by the classical economists and subsequently forgotten or ignored by the market fundamentalists. As Chris Griswold observes, David Ricardo raised exactly this fear himself: that new technology can indeed hurt workers when employers use it to reduce their demand for labor.

Policy can bring the profit incentive back into alignment with the interests of workers, both by actively supporting productive investment and by discouraging unproductive activity. For instance, a trade policy that discourages offshoring and makes domestic production relatively more attractive can help to place the productive employment of American workers back at the center of business strategy. Immigration restrictions and enforcement that curtail access to cheap and easily exploitable foreign labor also bring American workers back into focus and create an imperative to address labor shortages by automating processes and training employees. Workforce development should directly subsidize employers who take on inexperienced workers and place them on promising career paths. Industrial policy can support investment in the development and deployment of new technologies.

Strengthening worker power is fundamental to success on all these fronts, simultaneously creating incentives to boost productivity and positioning workers to share fully in the resulting gains. The “worker shortages” caused when trade and immigration policy foreclose the option of looking abroad to meet domestic demand are powerful market signals that making a profit will depend upon creating better jobs for American workers. Employment regulation can simply foreclose “low road” strategies that aim to hold costs down by relying upon easily exploitable labor. Worker organizations that provide leverage in the labor market and voice in the workplace can advance these same goals and also help to ensure that new tools and processes are deployed with workers’ interests in mind. Where workers have power, the companies developing new technologies will also have a much stronger incentive to consider how a product will serve the people using it. If technology were for improving worker productivity and job quality, rather than merely for increasing profit, developers would make very different decisions about the tools and applications they created and marketed.

The Luddites, it turns out, were not protesting technology. To the contrary, the machines at issue had been in use for years, and in many cases the protestors were the most adept users. Their demands were fair wages and better working conditions, not an end to automation. They smashed machines because that was the only form of power they had. Whether workers benefit from technology, and whether our politics supports our deployment, is not a question about the technology. It is a question for all of us.