New technologies typically boost the economy in one of two ways: by increasing quality or by increasing efficiency. Higher quality means producing better goods or providing services than before, while higher efficiency means that, all else being equal, a good or service can be produced or provided with fewer inputs, such as fewer hours of work or less energy and materials.

The invention of the elevator in 1852 improved quality of life by enabling people to work and live in high-rise buildings rather than climb dozens of flights of stairs. Nearly a century later, the invention of the self-service elevator boosted efficiency and, in the process, eliminated tens of thousands of elevator operator jobs, freeing those workers to produce other goods and services.

AI, like past technologies, will have both effects on today’s economy. But the magnitude remains to be seen and in the meantime is a subject of heated debate. The prevailing narrative is that AI will be revolutionary, if not downright dystopian. Bill Gates reflected this elite consensus opinion when he said earlier this year, “We won’t need humans for most things in 10 years with AI.”

But the likely reality will be more modest. Over the next decade, AI’s effects will probably not be revolutionary. But if we are lucky, the technology will help address a persistent challenge: slow productivity growth.

The Productivity Problem

The United States is stuck in a productivity rut. From 1947 to 2007, labor productivity (or economic output per hour of work) grew by 2.3 percent per year. Since 2007, however, growth has slumped to an average of 1.6 percent a year. It’s not entirely clear why, but Schumpeterian long-wave theory offers perhaps the best explanation.

That theory holds that the emergence, then the full use, and later the exhaustion of general-purpose technologies lead to long-term cycles of growth. Examples include electricity, internal-combustion engines, and semiconductors—powerful technologies that enable new products or services, or otherwise enhance productivity across all industries. Once these technologies become good enough, cheap enough, and broadly applicable, the economy typically experiences 20 to 30 years of strong growth as they become more commonly used. Then, after they become ubiquitous and related improvements slow, growth opportunities also slow, and we wait again for the next ChatGPT to arrive.

This has been the pattern in the West since the development of the steam engine. So, one theory of the slowdown we have experienced in recent decades is that the last wave of productivity-enhancing technologies, including personal computers and the Internet, reached the stage of full adoption and ceased to produce significant marginal improvements, while powerful new technologies, like AI and robotics, have not yet been fully developed. If AI is in fact a general-purpose technology rather than a very effective narrow technology, then it is likely to bring us closer to that benchmark of 2.3 percent productivity growth.

But it’s still early days. We don’t yet know how much better AI will become, especially now that current large language models (LLMs) have already absorbed most web data for training. We also don’t know if LLMs will keep getting more expensive to train and operate. For now, limitations like these will slow AI adoption and fully effective use. And we don’t know if the emergence of widespread opposition to AI in the West will significantly slow its development and adoption.

Still, it appears to be a safer bet that AI will boost productivity than peter out halfway through the Gartner hype cycle.

Three Possible Pathways to Productivity Growth

1. Automating and Augmenting Knowledge Work

AI should either boost workers’ efficiency or automate occupations for which the job requirements and information involved are relatively simple and rules-based.

In the first category, AI will allow workers to produce more, but not completely automate their tasks. Early evidence suggests this is already occurring. For example, the St. Louis Federal Reserve found that users of generative AI tools reported saving an average of roughly 2.2 hours per 40-hour workweek.

Meanwhile, other occupations could be automated entirely. For example, a Microsoft study concluded that the top 20 jobs at greatest risk of being eliminated by AI included telephone operators, translators, ticket agents, telemarketers, brokerage clerks, and customer service reps. As the technology improves, this list could extend to graphic artists, legal clerks, accountants, and other occupations. And a recent MIT report found that 11.7 percent of the labor market could be automated by AI, which is not a particularly large number, given that it would mean about a 1.2 percentage point increase in annual productivity if this occurred over a decade.

This does not mean AI will soon do everything humans can. I used AI to help research this blog, saving perhaps 5 to 10 percent of my usual time. But AI couldn’t have written this article, because I wanted to say something unique, and what current AI largely does is regurgitate existing content. When I asked an LLM to write an article on AI and productivity, it said:

The most promising pathway forward involves neither wholesale replacement nor suspicious rejection of AI, but rather thoughtful integration that leverages the complementary strengths of human and machine intelligence. This partnership model recognizes that AI and humans excel at different aspects of work.

That sounds reasonable, and reflects the consensus view online, but it is wrong: The most promising pathway, from an economics standpoint, is not necessarily the “partnership” model. Replacing workers with technology boosts productivity more than providing them with AI helpers. In the former case, all the freed-up workers are now available to work at other jobs that raise GDP. In the latter case, only some of them are.

Moreover, current AI tools also have real limitations. One study found that AI increased errors by 19 percent when used in consulting. We’ve all experienced AI hallucinations that provided nonexistent references or fabricated information. So, productivity results can be mixed. A telemarketing study found that AI boosted productivity for skilled workers, but reduced it for less-skilled workers, while another study found the opposite, with less-skilled workers’ productivity increasing 30 percent versus just 15 percent for more-skilled workers. The MIT report referenced above estimated that around 100,000 jobs were lost due to AI in 2025, which represents just 0.5 percent of all layoffs.

The reality is that AI boosts productivity for many knowledge workers, freeing up time for them to do other things, usually not leading to job loss, but to output and quality improvement.

Of course, the legions of AI pessimists like Andrew Yang predict we’ll automate everything, which seems scary. But despite the hype, most functions remain too complex for even more effective AI to fully replace human workers. Many jobs that theoretically could be automated likely won’t be, at least in the near future. Consider school bus drivers: Autonomous vehicles are, or will soon be, safer than human-driven ones, but few parents are likely to let their eight-year-old get on a school bus without adult supervision.

So, regardless of whether a driver is driving the bus, someone will be paid to keep an eye on the kids. Planes will likely retain pilots and flight attendants. Higher-end service establishments—like hotels, stores, and restaurants—will keep human workers as status markers if for no other reason, just as tailored clothing signals status today. Only the most extreme AI doomsayers imagine AI embalming corpses and running companies.

2. Enabling More Capable Physical Systems

AI can boost productivity by powering machines that interact with the physical world, where most jobs (defined broadly) still function. Think: cooking food, moving patients in nursing homes, repairing plumbing, and the like.

Autonomous trucks and trains will increase freight productivity, reducing labor needs. AI-powered robots will deliver food to tables in restaurants, reducing the need for wait staff. AI-driven “dark factories” will produce assembly-based products such as smartphones with minimal human labor.

In general, recent progress in automating physical jobs has been much slower than progress in automating knowledge work. AI-powered robots are still early in their development, but there is reason to believe that they will improve, given the significant amount of research on integrating AI into robotics and other mechanical items, and the steady progress of robotics over the last decade. At some point, this “physical AI” will be a key to boosting productivity in the physical world, from factories to farming to restaurants. But it is highly unlikely to be able to do all the physical work.

3. Reducing Waste

AI is also likely to boost productivity by reducing society’s need to devote scarce labor to specific activities. Here, it’s not that the productivity of the task is improved, as in the case of knowledge workers; it’s that society doesn’t need to spend scarce resources on that specific task.

For example, as automobiles become more autonomous—even without reaching full self-driving capability—accidents, including minor fender benders, will fall dramatically. As such, people will need far fewer autobody repair services, and so most of the 176,000 workers in autobody repair shops will then be able to perform other work that AI currently cannot do, such as road repair or house construction. This will lead to greater economic growth, because it won’t affect car quality, but these workers will now be employed producing additional goods and services.

Additionally, agentic AI—meaning autonomous AI systems that can plan, reason, and act to complete complex tasks with minimal oversight—can unburden individuals from daily tasks like paying bills, scheduling appointments, and finding information. While this will not boost measured GDP, because personal chores are not included in these measurements, it will free up considerable time.

Quality Improvements

AI will also improve the quality of products and services. Drug discovery exemplifies this potential. In 2021, Google DeepMind’s AlphaFold2 system solved a pivotal aspect of the long-standing “protein-folding problem,” a 50-year-old challenge in biology. By predicting the 3D structures of nearly all known proteins from their amino acid sequences, AlphaFold2 has transformed biological research, reducing prediction time from months to minutes. Hopefully, this opens the door to the development of more and better drugs.

AI will also reduce errors that cost the economy hundreds of billions of dollars annually. Though AI is prone to hallucinations, humans are imperfect as well, and AI is likely to improve faster than humans can. In many cases, AI-powered systems can inspect products with greater accuracy than the human eye—for example, identifying defects on assembly lines to improve product quality and compliance. Automating tasks like data entry or quality checks minimizes the potential for human mistakes in production and service delivery. It also reduces errors in medical clinics. Indeed, AI is more accurate than humans for many tasks, especially those with bounded data sets.

We’ll likely see AI quality improvements across many domains. AI already enables personalized learning in K-12 education, where the teacher remains central but AI can customize the curriculum to each student’s level and learning style, leading to more effective learning. AI improves weather forecasts, saving lives and money. It enhances product design across industries, enabling better manufacturing and greater customer satisfaction.

Since GDP growth isn’t just about producing more but also takes into account quality improvements—for example, a car that resists rust is worth more than a car that does not—these upgrades will lead to higher overall economic output.

What About Jobs?

Two concerns dominate the public debate: overall job numbers and the pace of change.

Regarding total employment: Don’t worry. Historically, automation—whether from tractors or computer-aided machine tools—hasn’t caused net job losses, and it likely never will.

The reason is straightforward: Automation boosts productivity, and increased productivity lowers prices or raises wages, or sometimes both. Both lower prices and higher wages increase spending and investment, which in turn creates jobs.

If we automate half of legal jobs, for instance, legal services will become cheaper, leaving consumers with more money for personal trainers, restaurant meals, or new fishing boats. This spending will create labor demand that the laid-off legal workers can fill.

But won’t companies just pocket all the gains and leave workers to survive on the crumbs of universal basic income, as many AI job doomers assert? They would, if we could somehow repeal the laws of economics and competition. Productivity has increased around 9 times since 1900, yet corporate profits as a share of the economy haven’t materially changed. This is because companies, even the largest companies of our day, face competition. If they didn’t, then we’d all be paying $5,000 for our televisions, $50 for our fast-food burgers, and $100 a month for Google and Facebook.

Competition forces companies to reduce prices and/or raise wages. Nobody will buy your $5,000 TV if a $500 model is available. And nobody will work for you if every other company in your industry pays more.

But won’t we run out of things to buy? No. If AI turns out to be vastly more powerful than any past technology and somehow automates half the jobs, then U.S. per-capita income would double from $73,400 to $146,800 through higher productivity. In which case, people would turn around and spend that new income on things like smaller class sizes in schools, first-class airline tickets, bigger houses, personal trainers, and more meals out at restaurants. Americans will never run out of things to spend money on, and their spending will continue creating jobs, because there will always be things that still require humans to produce or provide.

Even if we still have robust labor demand, won’t AI continue to boost productivity faster than workers’ wages, as many assert? The answer is that when measured correctly, it turns out that the growth of workers’ compensation has generally tracked economic productivity. Indeed, one definitive study found that when using appropriate measures, wages increased 63 percent and productivity increased around 76 percent from 1972 to 2013, not a big difference. There is no reason to believe that AI changes that relationship.

What if the pace of automation is so fast that millions of workers can’t find new work? That could pose problems, such as large numbers of workers becoming so discouraged that they stop even trying to find work.

But history suggests otherwise. No GPT has ever fully spread within 5 or 10 years. There are always lead adopters and laggards. Technology never emerges fully formed; each generation improves and falls in price. The first integrated circuit for semiconductors was developed in 1957, but it was not until the mid-1990s that chips became powerful enough to propel the Internet-connected PC revolution. Thomas Edison powered the first home with electricity in 1882, but it was not until the late 1940s that 90 percent of U.S. homes had electricity.

AI may spread faster than that, but it will still take time. However, that doesn’t mean Congress shouldn’t improve legislation that helps workers navigate job transitions, but fears of mass unemployment are fever dreams.

The Bottom Line

It is still early. We don’t yet know how much better AI will become. It still has numerous limitations that constrain widespread adoption and fully effective use. But if we are fortunate, AI will restore the productivity growth that has eluded us for 15 years—not through dystopian transformation, but through steady, incremental improvements across the economy.