Large language models such as ChatGPT are emerging as powerful tools that not only make workers more productive but also increase the rate of innovation, laying the foundation for a significant acceleration in economic growth. As a general purpose technology, AI will impact a wide array of industries, prompting investments in new skills, transforming business processes, and altering the nature of work. However, official statistics will only partially capture the boost in productivity because the output of knowledge workers is difficult to measure. The rapid advances can have great benefits but may also lead to significant risks, so it is crucial to ensure that we steer progress in a direction that benefits all of society.
On a recent Friday morning, one of us sat down in his favorite coffee shop to work on a new research paper regarding how AI will affect the labor market. To begin, he pulled up ChatGPT, a generative AI tool. After entering a few plain-English prompts, the system was able to provide a suitable economic model, draft code to run the model, and produce potential titles for the work. By the end of the morning, he had achieved a week’s worth of progress on his research.
We expect millions of knowledge workers, ranging from doctors and lawyers to managers and salespeople to experience similar ground-breaking shifts in their productivity within a few years, if not sooner.
The potential of the most recent generation of AI systems is illustrated vividly by the viral uptake of ChatGPT, a large language model (LLM) that captured public attention by its ability to generate coherent and contextually appropriate text. This is not an innovation that is languishing in the basement. Its capabilities have already captivated hundreds of millions of users.
Other LLMs that were recently rolled out publicly include Google’s Bard and Anthropic’s Claude. But generative AI is not limited to text: in recent years, we have also seen generative AI systems that can create images, such as Midjourney, Stable Diffusion or DALL-E, and more recently multi-modal systems that combine text, images, video, audio and even robotic functions. These technologies are foundation models, which are vast systems based on deep neural networks that have been trained on massive amounts of data and can then be adapted to perform a wide range of different tasks. Because information and knowledge work dominates the US economy, these machines of the mind will dramatically boost overall productivity.
The power of productivity growth
The primary determinant of our long-term prosperity and welfare is the rate of productivity growth: the amount of output created per hour worked. This holds even though changes in productivity are not immediately felt by everyone and, in the short run, workers’ perceptions of the economy are dominated by the business cycle. From World War II until the early 1970s, labor productivity grew at over 3% a year, more than doubling over the period, ushering in an era of prosperity for most Americans. In the early 1970s productivity growth slowed dramatically, rebounding in the 1990s, only to slow again since the early 2000s.
Figure 1: The Components of labor productivity growth 1948 to 2022. Contribution of total factor productivity. Contributions of capital intensity and labor force composition.
Figure 1 illustrates the story. It decomposes the overall growth in labor productivity into two components: total factor productivity (which is a measure of the impact of technology) and the contribution of the labor composition and capital intensity. The figure illustrates that the key driver of changes in labor productivity is changes total factor productivity (TFP). There are many reasons for America’s recent economic struggles, but slow TFP growth is a key cause, slowly eating away at the country’s prosperity, making it harder to fight inflation, eroding workers’ wages and worsening budget deficits.
The generally slow pace of economic growth, together with the outsized profits of tech companies, has resulted in skepticism about the benefits of digital technologies for the broad economy. However, for about 10 years starting in the 1990s there was a surge in productivity growth, as shown in Figure 1, driven primarily by a huge wave of investment in computers and communications, which in turn drove business transformations. Even though there was a stock market bubble as well as significant reallocation of labor and resources, workers were generally better off. Furthermore, the federal budget was balanced from 1998 to 2001—a double win. Digital technology can drive broad economic growth, and it happened less than thirty years ago.
Early estimates of AI’s productivity effects
The recent advances in generative AI have been driven by progress in software, hardware, data collection, and growing amounts of investment in cutting-edge models. Sevilla et al. (2022) observe that the amount of compute (computing power) used to train cutting-edge AI systems has been doubling every six months over the past decade. The capabilities of generative AI systems have grown in tandem, allowing them to perform many tasks that used to be reserved for cognitive workers, such as writing well-crafted sentences, creating computer code, summarizing articles, brainstorming ideas, organizing plans, translating other languages, writing complex emails, and much more.
Generative AI has broad applications that will impact a wide range of workers, occupations, and activities. Unlike most advances in automation in the past, it is a machine of the mind affecting cognitive work. As noted in a recent research paper (Eloundou et al., 2023), LLMs could affect 80% of the US workforce in some form.
There is an emerging literature that estimates the productivity effects of AI on specific occupations or tasks. Kalliamvakou (2022) finds that software engineers can code up to twice as fast using a tool called Codex, based on the previous version of the large language model GPT-3. That’s a transformative effect. Noy and Zhang (2023) find that many writing tasks can also be completed twice as fast and Korinek (2023) estimates, based on 25 use cases for language models, that economists can be 10-20% more productive using large language models.
But can these gains in specific tasks translate into significant gains in a real-world setting? The answer appears to be yes. Brynjolfsson, Li, and Raymond (2023) show that call center operators became 14% more productive when they used the technology, with the gains of over 30% for the least experienced workers. What’s more, customer sentiment was higher when interacting with operators using generative AI as an aid, and perhaps as a result, employee attrition was lower. The system appears to create value by capturing and conveying some of the tacit organizational knowledge about how to solve problems and please customers that previously was learned only via on-the-job experience.
Criticism of large language models as merely “stochastic parrots” is misplaced. Most cognitive work involves drawing on past knowledge and experience and applying it to the problem at hand. It is true that generative AI programs are prone to certain types of mistakes, but the form of these mistakes is predictable. For example, language models tend to engage in “hallucinations,” i.e., to make up facts and references. As a result, they clearly require human oversight. However, their economic value depends not on whether they are flawless, but on whether they can be used productively. By that criterion, they are already poised to have a massive impact. Moreover, the accuracy of generative AI models continues to improve rapidly.
Quantifying the productivity effects
A recent report by Goldman Sachs suggests that generative AI could raise global GDP by 7%, a truly significant effect for any single technology. Based on our analysis of a variety of use cases and the share of the workforce doing mainly cognitive work, this estimate strikes us as being reasonable, though there remains great uncertainty about the ultimate productivity and growth effects of AI.
It is useful to rigorously break down the channels through which we expect generative AI to produce growth in productivity, output, and ultimately in social welfare in a model.
The first channel is the increased efficiency of output production. By making cognitive workers engaged in production more efficient, the level of output increases. Economic theory tells us that, in competitive markets, the effect of a productivity boost in a given sector on aggregate productivity and output is equal to the size of the productivity boost multiplied by the size of the sector (Hulten’s theorem). For instance, if generative AI makes cognitive workers on average 30% more productive over a decade or two and cognitive work makes up about 60% of all value added in the economy (as measured by the wage bill attributable to cognitive tasks), this amounts to a 18% increase in aggregate productivity and output, spread out over those years.
The second, and ultimately more important, channel is the acceleration of innovation and thus future productivity growth. Cognitive workers not only produce current output but also invent new things, engage in discoveries, and generate the technological progress that boosts future productivity. This includes R&D—what scientists do—and perhaps more importantly, the process of rolling out new innovations into production activities throughout the economy—what managers do. If cognitive workers are more efficient, they will accelerate technological progress and thereby boost the rate of productivity growth—in perpetuity. For example, if productivity growth was 2% and the cognitive labor that underpins productivity growth is 20% more productive, this would raise the growth rate of productivity by 20% to 2.4%. In a given year, such a change is barely noticeable and is usually swamped by cyclical fluctuations.
But productivity growth compounds. After a decade, the described tiny increase in productivity growth would leave the economy 5% larger, and the growth would compound further every year thereafter. What’s more, if the acceleration applied to the growth rate of the growth rate (for instance if one of the applications of AI was to improving AI itself), then of course, growth would accelerate even more over time.
Figure 2: Possible growth trajectories. Source: Authors’ simulations.
Figure 2 schematically illustrates the effects of the two channels of productivity growth over a twenty year horizon. The baseline follows the current projection of the Congressional Budget Office (CBO) of 1.5% productivity growth, giving rise to a total of 33% productivity growth over 20 years. The projection labeled “Level” assumes that generative AI raises the level of productivity and output by an additional 18% over ten years, as suggested by the illustrative numbers we discussed for the first channel. After ten years, growth reverts to the baseline rate. The third projection labeled “Level+Growth” additionally includes a one percentage point boost in the rate of growth over the baseline rate, resulting from the additional innovation triggered by generative AI. At first, the resulting growth trajectory is barely distinguishable from the “Level” projection, but through the power of compounding, the effects grow bigger over time, leading to a near doubling of output after 20 years, far greater than the baseline projection.
Barriers and drivers of adoption
For the productivity gains to materialize, advances in AI have to disseminate throughout the economy. Traditionally, this has always taken time, so we would not expect potential productivity gains to show up immediately. The advances need to be taken up and rolled out by businesses and organizations that employ cognitive labor throughout the economy, including small and medium-sized businesses, some of which may be slow to realize the potential of adapting advanced new technologies or may lack the required skills to use them well. For example, the Goldman report assumes it takes 10 years for the gains to fully materialize.
The “productivity J-curve” (Brynjolfsson et al., 2021) describes how new technologies, especially general purpose technologies, deliver productivity gains only after a period of investment in complementary intangible goods, such as business processes and new skills. In fact, this can temporarily even drag down measured productivity. As a result, earlier general purpose technologies like electricity and the first wave of computers took decades to have a significant effect on productivity. Additional barriers to adoption and rollout include concerns about job losses and institutional inertia and regulation, in areas from the medicine to finance and law.
However, in the case of generative AI there are also factors that can mitigate these barriers, or even accelerate adoption. First, in contrast to physical automation, one benefit of cognitive automation is that it can often be rolled out quickly via software. This is particularly true now that a ubiquitous digital infrastructure is available: the Internet. ChatGPT famously was the most rapid product launch in history—it gained 100 million users in just two months—because it was accessible to anyone with an internet connection and did not require any hardware investment on the users’ side.
Both Microsoft and Google are in the process of rolling out Generative AI tools as part of their search engines and office suites, offering access to generative AI to a large fraction of the cognitive workforce in advanced countries who regularly use these tools. Furthermore, application programming interfaces (APIs) are increasingly available to enable seamless modularization and connectivity between systems, and a marketplace for plug-ins and extensions is rapidly growing, making it much easier to add functionality. Finally, in contrast to other technologies, users of generative AI can interact with the technology in natural language rather than special codes or commands, making it easier to learn and adopt these tools.
These reasons for optimism suggest that the rollout of these new technologies may be faster than in the past. Still, the importance of training to make optimal use of these tools cannot be overstated.
Problems of measurement – silent productivity growth
The most common measure of productivity, non-farm business productivity, is quite adept at capturing increases in productivity in the industrial sector where inputs and outputs are tangible and easy to account for. However, productivity of cognitive labor is harder to measure. Statisticians who compile GDP and productivity statistics sometimes resort to valuing the output of cognitive activity simply by assuming it is proportional to the quantity of labor input being used to produce it, which of course eliminates any scope for productivity growth.
For example, generative AI enables economists to write more thought pieces and provide deeper analyses of the economy than before, yet this output would not directly show up in GDP statistics. Readers may feel that they have access to better and deeper economic analyses (contributing to channel 1 above). Moreover, the analyses may also play a part in enabling business leaders and policymakers to better harness the positive productivity effects of generative AI (contributing to channel 2 above). Neither of these positive productivity effects of such work would be directly captured in official GDP or productivity statistics, yet the benefits of economists’ productivity gains would still lead to greater social welfare.
The same holds true for many other cognitive workers throughout the economy. This may give rise to significant under-measurement or “silent productivity growth.”
Productivity growth, labor markets, and income distribution
A bigger pie does not automatically mean everyone benefits evenly, or at all. The productivity effects of generative AI are likely to go hand in hand with significant disruption in the job market as many workers may see downward wage pressures. For example, the Eloundou et al. paper cited earlier predicts that up to 49% of the workforce could eventually have half or more of their job tasks performed by AI. Will the demand for these tasks increase enough to compensate for such efficiency gains? Will the workers find other tasks to do? The answers are far from certain. In past technological transformations, workers who lost their jobs could transition to new jobs, and on average pay increased. However, given the scale of the impending disruption and the labor-saving nature of it, it remains to be seen whether this will be the case in the age of generative AI.
Moreover, the current wave of cognitive automation marks a change from most earlier waves of automation, which focused on physical jobs or routine
GPT’s reaction to this article:
tasks. This means that the impact on the labor market and income distribution may be different this time around. It is important to ensure that the benefits of productivity growth are shared widely and that workers have access to the skills and training needed to adapt to the changing nature of work.
In conclusion, the emergence of large language models and generative AI has the potential to significantly boost productivity and accelerate economic growth. However, it is important to ensure that the benefits are shared widely and that progress is steered in a direction that benefits all of society. The impact on the labor market and income distribution is uncertain, and it is crucial to invest in training and skills development to support workers in adapting to the changing nature of work.