A big AI-driven productivity boom is unlikely
Like the information and communications technology (ICT) revolution of the 1990s, the impact will be limited
Slow productivity growth and stagnant real wages
Most economists associate economic growth, rising wages, and improved living standards with increases in productivity, both labor productivity and Total Factor Productivity (TFP), which includes capital input and the effects of technological innovation and efficiency. This was certainly the case in the 30 years after World War II and had been the case since the mid-19th Century due to the impact of major innovations like electricity, the internal combustion engine, indoor plumbing, antibiotics, mass production techniques, air travel, and the mechanization of agriculture.
But productivity, however defined and measured, has grown only slowly in advanced democracies since the 1970s, and, despite an upturn in the 1990s, the slow-growth trend has continued since 2004 in the U.S., even before the Great Recession, and has worsened in recent years. Productivity growth in Europe has followed a slightly different pattern, but the entire advanced sector has been similarly affected. Economist Robert Gordon calls this “supply-side secular stagnation,” a long-term decline in the potential for economic growth (see our previous post on Gordon here).
The U.S. Bureau of Labor Statistics (BLS) had this to say about the U.S. labor productivity situation in April 2021 (Monthly Labor Review):
Labor productivity—defined as output per labor hour—has grown at a below-average rate since 2005, representing a dramatic reversal of the above-average growth of the late 1990s and early 2000s….$10.9 trillion represents the cumulative loss in output in the U.S. nonfarm business sector due to the labor productivity slowdown since 2005, also corresponding to a loss of $95,000 in output per worker.
The BLS goes on to comment on the expectation that ICT would kick off a new industrial revolution and sustainable growth in productivity and prosperity:
….One of the more perplexing aspects of the current slowdown is its genesis: that it came immediately following a historic productivity boom in the United States, and represented a swift rebuke of the popular idea of that time that we had entered a new era of heightened technological progress. The suddenness and size of the reversal were difficult to comprehend. For some background, in the late 1990s, when that much-cited productivity boom had begun, U.S. labor productivity growth had accelerated to rates of change that had not been seen since the late 1960s and early 1970s.
….A debate ensued among economists: Was the tremendous productivity growth of the late 1990s here to stay—a fundamental change generated by the computing and internet-related innovations that were all around us—or was it a temporary phenomenon that would pass? …. Many economic observers were yet again surprised, in this case at just how drastically growth rates slowed, given the recently observed high rates of growth and the continued technological innovations that were proliferating throughout the economy.
The BLS also notes that labor productivity has declined even further since 2010:
In the years since 2005, labor productivity has grown at an average annual rate of just 1.3 percent, which is lower than the 2.1-percent long-term average rate from 1947 to 2018. The slow growth observed since 2010 has been even more striking: labor productivity grew just 0.8 percent from 2010 to 2018.
One of the consequences of this slow-growth trend has been the flatlining of real wages since the 1970s, and while demand and pay for qualified tech workers and professionals has risen, the average wage over this period disguises the fact that real wages for many non-tech low and middle-income workers have actually fallen.
The trend toward weak productivity and stagnant or declining real wages is the same across the advanced sector. Here’s a McKinsey commentary on the G7 nations:
Slowdowns in productivity growth can be seen across the G7 since 2005, with most countries experiencing declines of 0.6 to 1.6 percentage points—equivalent to a halving of productivity growth. Other countries confront more considerable challenges: Italy has barely sustained any productivity growth at all, averaging only 0.1 percent per year since 2005. The United Kingdom has experienced the most significant drop-off, falling from robust rates of 2.1 percent over 1995–2005 to 0.5 percent from 2005–19. These declines represent significant challenges to economic growth and individual prosperity in these economies.
The BLS report drew similar conclusions about profits, wages, and living standards in the U.S.:
As the slowdown in labor productivity growth has steadily held on throughout the past decade , economic observers have been trying to understand this phenomenon, which has the effect of placing downward pressure on economic growth, worker compensation gains, profits growth, and gains in living standards.
This has created a serious problem for social-democratic politics which has historically relied on continued capitalist dynamism and surpluses in order to protect and improve wages and pensions, and expand welfare state benefits (see our previous post on this here). Wage earners will continue to abandon social-democratic parties if they don’t start directly addressing the stagnant real wages and rising economic insecurity that have accompanied the productivity slowdown.

Note: The trend is calculated using a Hodrick-Prescott filter with a smoothing parameter lambda of 10.
Productivity change in the U.S. nonfarm business sector, 1947-2022
Why did the productivity boost driven by the Information and Communications Technology (ICT) revolution run out of steam?
As the BLS notes, experts have been baffled by the productivity stagnation since 2004 and the weak recovery since the Great Recession. Two main theories have been offered: 1- The full impact of the 1990s ICT revolution is still to come; and 2- The scope of the ICT impact has been too narrow.
Is the ICT impact still to come?
Despite the fact that the slowdown in the U.S. commenced before the Great Recession and has continued since, some argue that the full economic impact of the ICT revolution of the 1990s is still to come. They say that most innovation in the past took decades to be reflected in the general economy and that this is likely going to be true for ICT. They also point to barriers created by the Great Recession, regulatory restrictions, trade conflicts, and, more recently, inflation.
The theory is correct up to a point. While most basic software applications used in business, government, academia, and elsewhere have been around for a long time, access to high-speed internet is an area where we can expect a continuing impact. There are still companies and communities without access to high-speed connectivity and office work will continue to be more efficient at higher download speeds.
In addition, working from home and video-conferencing has been growing and will continue to grow as internet speed increases. Rural areas will benefit in many ways and many small businesses are not yet connected. As today’s young people grow into tomorrow’s mature workforce, their superior facility with IT tools will surely influence labor productivity.
While skeptics recognize the continuing value of the ICT revolution, they disagree as to the scale of what’s left to be impacted. They point out that ICT has been around for over 30 years, plenty of time for most of the labor productivity value to be realized, and that most advanced societies have already picked the low-hanging fruit. Secretaries, bank tellers, toll collectors, inventory clerks, telephone operators, and many other occupations have already been fully or partially replaced by digital technology.
In addition, business and office software is more or less the same today as it was 20 years ago and email works the same way. Smart phones have changed lives across the planet, but recent upgrades offer mostly boutique new features. Most companies and organizations already have a web presence and even physicians and lawyers, late-comers to the revolution, have migrated online over the past decade, increased their use of internet tools, and reduced the need for legal secretaries and assistants.
Since the ICT transformation occurred just once, however, most productivity gains have already been baked into the numbers.
There are also certain factors that can blunt productivity growth. ICT systems are relentlessly upgraded and legacy systems are often replaced by new products. While this will improve efficiency, it often requires a new layer of high-priced staffing. Companies without a large in-house IT staff are forced to hire managers and support staff, or engage consultants, to implement and maintain the new systems.
In sum, skeptics argue that weak recent productivity data reflect diminishing returns from the ICT revolution. The gains of the 1990s were transformative but their economic impact has already been baked into the numbers. While advances, like greater internet speed, will help some businesses become even more efficient, it will not be enough to extend the boom.
The scope of the 1990s ICT revolution impact was relatively narrow; the effect on the broader economy has been limited
Despite the transformative nature of the ICT revolution, its greatest productivity impact has been in the arena of knowledge work, i.e., business back office, research, and other white collar work. Digital tools and early pattern-recognition software have significantly improved revenue management, for example, in travel, hospitality, and many other industries.
There have been important impacts on manufacturing, supply-chain efficiency, inventory control and many industries in the material economy, but they don’t compare with effects in the knowledge economy. Industries like construction, energy, mining, metallurgy, agriculture, lumber, water supply, sewer systems, infrastructure maintenance and repair, mass transit, trucking, freight transport, military hardware, furniture, cosmetics, fashion, home repair, auto repair, and many others, have surely been impacted by ICT, the effects have been more modest, below what most would consider “transformative.” Today’s cars, trucks, trains, barges, and ships, perform the same basic tasks they did in the 1960s despite being loaded up with digital technology. Other than the addition of microwave ovens, today’s household and restaurant kitchens work much like the kitchens of the 60s, despite the introduction of “smart” technology.
In addition, as key elements of the manufacturing and extractive industries have moved overseas in search of cheap labor, service sector growth has accelerated becoming the dominant source of job growth in the advanced economies. Much of the growth has been in labor-intensive work, however, largely resistant to innovation and improved efficiency. Care workers are in very high demand, for example, including nurses and nurse aides in institutions, home-care workers and child-care workers. Aging populations will guarantee continued growth in demand for home-care workers for older adults, and the increasing need for young women to be in the workforce to maintain a middle-class lifestyle for their families will ensure high demand for child-care and pre-K workers.
Precarious (gig) work driven by online platforms is one arena in which labor productivity gains are real and growing and are contributing to economic growth in the broader economy. Some see this as the future of work and consider it a positive, even liberating, phenomenon that will keep the advanced sector growth strong.
Others see the gig economy as oppressive, retrograde, and contributing to the stagnation or decline in real wages, economic security, and living standards of working and middle-class families. While some have thrived in the environment, those who do gig work full-time are often underpaid and lack basic benefits. Those with other full-time employment who do gig work part-time mostly say they are doing it because they want to improve their retirement reserves, pay for college tuition, or get out from under serious debt. In other words, they are trying to maintain a middle-class lifestyle in the face of stagnant wages.
It should come as no surprise that gig workers are starting to unionize and fight against laws that categorize them as independent contract workers who are not entitled to overtime, health insurance, and other benefits.
While the ICT-driven gig economy may help reverse the decline in labor productivity, it is having no, or negative impact on real wages, which is, or should be, the primary concern of social democrats.
Will artificial intelligence generate a new productivity boom and boost real wages?
Increasing numbers of tech entrepreneurs, techno-optimists, and libertarians have expressed great hope in the power of ICT. They are typically proponents of free markets and believers in technological disruption and creative destruction. They are also typically skeptical of big government solutions. They are somewhat baffled and concerned that the ICT revolution has not shown up in productivity statistics, but believe that we are on the cusp of a major acceleration of the revolution, driven largely by advances in artificial intelligence and intelligent autonomous robots, including driverless cars.
Generative AI is a powerful tool, but its impact will be limited
Here’s how McKinsey defines “generative AI”:
Generative artificial intelligence (AI) describes algorithms (such as ChatGPT) that can be used to create new content, including audio, code, images, text, simulations, and videos. Recent breakthroughs in the field have the potential to drastically change the way we approach content creation.
This is how, according to McKinsey, ChatGPT, defines itself:
Ready to take your creativity to the next level? Look no further than generative AI! This nifty form of machine learning allows computers to generate all sorts of new and exciting content, from music and art to entire virtual worlds. And it’s not just for fun—generative AI has plenty of practical uses too, like creating new product designs and optimizing business processes. So why wait? Unleash the power of generative AI and see what amazing creations you can come up with!
Common to both these descriptions is the word “content”, reminding us that the primary impact of the ICT revolution was among knowledge workers, including scientific research, medical diagnosis and similar skilled professionals. With 95% of all knowledge available in digital form, generative AI algorithms have the ability to quickly scan content libraries and create summaries and, in response to queries, fresh content. In many cases, they outperform humans in terms of time to perform a task, accuracy, and completeness.
With the dominant value being in the domain of knowledge work, however, generative AI will continue to affect an important but narrow part of the broader economy. We have Google to search for information, AI will serve as a super-Google. We have Wikipedia to provide a consensus summary and description of numerous content items, AI will be a super-Wiki. The effect on the material economy, however, will continue the pattern we’ve already seen, impactful but not transformative.
This will certainly have a productivity impact in the world of knowledge work, but even many passionate proponents do not feel that it will have much of an effect on employment. The tool will serve as a super-Alexa, a powerful assistant, but still an assistant.
Part of the reason for suggesting that job loss and overall impact will be modest are a number of important, looming concerns that will be amplified by AI’s more powerful capabilities, and will keep human workers actively involved:
Fact-checking: Generative AI has difficulty distinguishing between fact and fiction, especially if the fiction sounds right and is widely repeated, and facts themselves are often open to different interpretations. The internet is also a notorious repository of inaccuracies, odd interpretations of facts, outlandish conspiracy theories, lies, falsehoods, and half-truths. Early tests have produced some reasonable-sounding pure nonsense in response to queries, and bad actors have already begun to use AI tools to produce realistic deep fake photos, misinformation, and propaganda.
Accuracy: Beyond fact-checking, there are knowledge areas where accuracy is at a premium. We don’t yet know how accurate AI algorithms will be in making medical diagnoses, engineering fault analyses, or military assessments, among other things. Inaccuracy in many domains can be a matter of life or death.
Privacy and security: Given its powerful capabilities, Generative AI may need additional layers of regulation and security to keep it from accessing and divulging sensitive or proprietary information. Care will also be needed to assure that chatbots are not engaged in plagiarism or violating intellectual property laws.
Bias: The internet is a place where racial, ethnic, religious, gender and social biases proliferate widely. Historical texts and images that were previously common, may be unacceptable in today’s culture. Early experiments with chatbots found a surprising number of racist and sexist comments blended into knowledge products.
Can Generative AI be efficient and effective if it is accurate only 80-90% of the time, or if it occasionally violates legal constraints and social convention? It seems that humans will be needed to “proof” the work of the algorithms, further limiting its productivity value. Generative AI will be a powerful tool, but still a tool, and one which will affect only a segment of the economy.
Intelligent autonomous robots: A technology looking for a deep economic purpose
Robotic technology has been around for a long time. It was instrumental in improving mass production techniques and now dominates the world of advanced manufacturing in industrial economies. Labor-intensive manufacturing involving products like apparel have long since moved to less-developed economies.
The last generation of robotics, however, involved largely single-purpose machines, both fixed and mobile, that did not make “decisions” about their next moves in real time.
The advent of artificial intelligence has created expectations, however, that powerful machine learning techniques will allow robots to become fully autonomous and capable of making their own decisions while interacting independently with the world of humans. Since humanoid robots still have significant difficulties carrying out many simple physical maneuvers, like turning a doorknob, picking up a dime, cracking eggs, handling a spatula, or using a stepstool to reach a high shelf, attention has turned to physical robots that can carry out complex, open-ended tasks.
A promising domain involves the potential for military use of AI-driven robots. Robot “soldiers” capable of assessing a battle situation and making real-time decisions about attack and retreat in a complex environment may soon be a reality. They will not need to look like humans or have human capacities beyond those needed to carry out battle functions. Such soldiers would be fed millions of possible scenarios and would “learn” how to develop optimal situational responses even under uncertain and changing conditions. There will be many variables at work, but without having to worry about battlefield casualties, the military will likely be able to produce robots that can eventually match or even outperform human soldiers. Many expect AI-driven robotics to be a true game-changer in the domain of military conflict.
Driverless vehicles
The prospective autonomous robot technology attracting the greatest public interest and enormous levels of investment is driverless vehicles, but it is unclear as to just how such vehicles would impact productivity. Having a previous driver use the time available to do productive work on a computer is still in the domain of knowledge work, and since ICT has already extended the work day for many, it’s just as likely that the driver will take a nap or browse YouTube.
Some justify going to driverless cars as a way to prevent injuries and fatalities in vehicle accidents. Yet studies in this area usually assume that all cars would need to be driverless to realize large gains, and no one sees the transition as being so abrupt, especially since vehicles equipped with lidar, radar, and powerful computers would likely be far more expensive than current vehicles. As long as driverless cars share the road with human drivers, pedestrians, bicycles, and motorcycles, significant reductions in accidents may not ensue. This has already been borne out even in the simplified conditions of early testing.
In addition, if driverless cars must interact with humans for any significant length of time, there will be some likely insurmountable problems:
Combinatorial explosion: While battlefield simulations will involve variables defined by the circumstances and goals of a tactical combat situation, driverless cars may face situations with so many variables that change so quickly that the ensuing combinatorial explosion will make it impossible to make decisions in real time, decisions that humans make readily using intuitive, heuristic thinking. Sharing a busy rush-hour highway with human-driven vehicles going an average of 60 mph in a 50 mph zone, might require a driverless car to assess and predict the behavior of a dozen drivers who may be entering the highway, exiting the highway, or changing lanes for a variety of reasons.
Moreover, moves made by one driver may cause another driver to change behavior, as when a faster driver comes up quickly behind a slower driver and then attempts to change lanes to get around the slower driver as the slower driver simultaneously tries to move to the same lane in an effort to accommodate the faster driver. The possible combination of moves by multiple vehicles would be large enough itself, but the circumstances would be changing constantly in very busy conditions, often in a matter of seconds. The slow driver may opt to interrupt the move to a new lane and the faster driver may do the same, while other nearby drivers and the driverless vehicle are forced to orient to the same situation.
Conservatism: To avoid liability, manufacturers will have no choice but to build conservative driving features into driverless vehicles. They will need to obey traffic regulations and allow extra time and space to manage borderline situations. Traffic regulations are usually established by local authorities, so vehicles would have to have the regulations built into their systems which might then change with every new jurisdiction. Will vehicles be allowed to cross a double line to pass a cyclist? Will a vehicle be allowed to enter a crosswalk or intersection controlled by a traffic signal before all vehicles are clear of the intersection? Would a driverless car be allowed to exceed the speed limit, especially if they are creating a slow-driver hazard as other cars travel 10-15 mph faster?
The behavior of pedestrians can flummox a driverless vehicle, including at low speeds. Moving through a small town main street on a busy Saturday afternoon can lead to a very uncomfortable ride as the driverless vehicle slows down or stops every time it “sees” a pedestrian stepping off a curb to, say, put a package in a parked car, or in anticipation of a traffic signal change, or in preparation to jay walk. The vehicle will be programmed to assume the worst while the passenger suffers through the jerky starts and stops. In the extreme case, driverless cars will simply stop and not proceed at all if faced with an edge situation it can’t assess as affirmatively safe including in severe weather conditions or on unmapped unpaved roads at night.
Ethical dilemmas: There have been numerous essays debating whether and how a driverless vehicle would “decide” on an action that might endanger its passengers or others. Would it be programmed to save its passengers in an edge situation even if it meant potentially killing others? Would it act to endanger an older passenger if it meant not endangering many children in a bus? Humans, of course, could face the same decisions, and need to make an instantaneous decision. It’s part of life. But how should programmers guide the autonomous robot? Would they set hard rules? Would it even be possible to allow for conditional flexibility for obeying regulations?
There is also the question of liability should an accident, injury, or death occur. Will insurers provide liability insurance, and how costly would it be to the vehicle owner? With no driver involved, how would fault be established? Could the manufacturer face charges?
Vandalism and sabotage: Driverless vehicles will be particularly vulnerable to all sorts of vandalism, especially while traveling empty. Vehicles could easily be stopped by a pedestrian or another vehicle and subjected to physical assaults. Vehicles with or without passengers would be vulnerable to interference with wireless communications.
These considerations threaten to derail the driverless vehicle industry even before it gets going. Consumers will surely resist if they feel the technology is dangerous, too costly, or inconvenient. States will be reluctant to implement costly infrastructure changes to accommodate driverless vehicles, like separate lanes and road sensors, if they are not sure the technology will be realized. Investors and manufacturers have already started withdrawing from the space as wildly exaggerated predictions have not been borne out.
There may still be a number of ways that driverless vehicles will play a role in productivity growth, however, though it will have limits. While driving in dense cities will run into the problems listed above, long-distance trucking on major highways leading to stations where a driver would take over appears promising and could decrease the need for truckers.
Taxi services limited to well-mapped local routes are already operating in a number of cities across the globe which would reduce the need for drivers over the long run. Driverless buses could operate on fixed bus routes under controlled conditions using designated bus lanes. Driverless taxis and buses could be monitored at a control center with the capacity to redirect the service and respond to operational obstacles. These would be important but not transformative innovations.
Conclusion
Artificial intelligence and robots are not the only future technologies being touted as able to advance economic growth and increase living standards. Genomics, synthetic biology, and 3-D printing are among the others, but the potential impact has not yet been reliably calculated. Extending lifespans may seem like a good thing but octogenarians will not be in the workforce and are the most likely to need labor-intensive long-term care. Fusion energy has great potential, however, since it would provide limitless cheap energy, but the timeline for its commercial development has not yet been established.
This is a newsletter that wants to help revive the political prospects of social democrats. Social democrats should fight the “old” battles for direct public investment into quality affordable housing, universal health care, long-term care, and child care, improved pensions and social security, increased public services, as well as the “new” battle to mitigate the effects of climate change.
Technological innovation will be an essential element, but we can no longer expect “capitalist dynamism” to bring about the future that working people deserve.